diff options
323 files changed, 16863 insertions, 4520 deletions
diff --git a/.gitignore b/.gitignore index 88d7b56181..ae39c52b11 100644 --- a/.gitignore +++ b/.gitignore @@ -34,3 +34,7 @@ log/ spark-tests.log streaming-tests.log dependency-reduced-pom.xml +.ensime +.ensime_lucene +checkpoint +derby.log @@ -12,11 +12,16 @@ This README file only contains basic setup instructions. ## Building -Spark requires Scala 2.9.2. The project is built using Simple Build Tool (SBT), -which is packaged with it. To build Spark and its example programs, run: +Spark requires Scala 2.9.2 (Scala 2.10 is not yet supported). The project is +built using Simple Build Tool (SBT), which is packaged with it. To build +Spark and its example programs, run: sbt/sbt package +Spark also supports building using Maven. If you would like to build using Maven, +see the [instructions for building Spark with Maven](http://spark-project.org/docs/latest/building-with-maven.html) +in the spark documentation.. + To run Spark, you will need to have Scala's bin directory in your `PATH`, or you will need to set the `SCALA_HOME` environment variable to point to where you've installed Scala. Scala must be accessible through one of these diff --git a/bagel/pom.xml b/bagel/pom.xml index a8256a6e8b..b83a0ef6c0 100644 --- a/bagel/pom.xml +++ b/bagel/pom.xml @@ -3,8 +3,8 @@ <modelVersion>4.0.0</modelVersion> <parent> <groupId>org.spark-project</groupId> - <artifactId>parent</artifactId> - <version>0.7.0-SNAPSHOT</version> + <artifactId>spark-parent</artifactId> + <version>0.8.0-SNAPSHOT</version> <relativePath>../pom.xml</relativePath> </parent> @@ -102,5 +102,42 @@ </plugins> </build> </profile> + <profile> + <id>hadoop2-yarn</id> + <dependencies> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-core</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-api</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> + <scope>provided</scope> + </dependency> + </dependencies> + <build> + <plugins> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-jar-plugin</artifactId> + <configuration> + <classifier>hadoop2-yarn</classifier> + </configuration> + </plugin> + </plugins> + </build> + </profile> </profiles> </project> diff --git a/bagel/src/main/scala/spark/bagel/Bagel.scala b/bagel/src/main/scala/spark/bagel/Bagel.scala index 094e57dacb..5ecdd7d004 100644 --- a/bagel/src/main/scala/spark/bagel/Bagel.scala +++ b/bagel/src/main/scala/spark/bagel/Bagel.scala @@ -4,8 +4,37 @@ import spark._ import spark.SparkContext._ import scala.collection.mutable.ArrayBuffer +import storage.StorageLevel object Bagel extends Logging { + val DEFAULT_STORAGE_LEVEL = StorageLevel.MEMORY_AND_DISK + + /** + * Runs a Bagel program. + * @param sc [[spark.SparkContext]] to use for the program. + * @param vertices vertices of the graph represented as an RDD of (Key, Vertex) pairs. Often the Key will be + * the vertex id. + * @param messages initial set of messages represented as an RDD of (Key, Message) pairs. Often this will be an + * empty array, i.e. sc.parallelize(Array[K, Message]()). + * @param combiner [[spark.bagel.Combiner]] combines multiple individual messages to a given vertex into one + * message before sending (which often involves network I/O). + * @param aggregator [[spark.bagel.Aggregator]] performs a reduce across all vertices after each superstep, + * and provides the result to each vertex in the next superstep. + * @param partitioner [[spark.Partitioner]] partitions values by key + * @param numPartitions number of partitions across which to split the graph. + * Default is the default parallelism of the SparkContext + * @param storageLevel [[spark.storage.StorageLevel]] to use for caching of intermediate RDDs in each superstep. + * Defaults to caching in memory. + * @param compute function that takes a Vertex, optional set of (possibly combined) messages to the Vertex, + * optional Aggregator and the current superstep, + * and returns a set of (Vertex, outgoing Messages) pairs + * @tparam K key + * @tparam V vertex type + * @tparam M message type + * @tparam C combiner + * @tparam A aggregator + * @return an RDD of (K, V) pairs representing the graph after completion of the program + */ def run[K: Manifest, V <: Vertex : Manifest, M <: Message[K] : Manifest, C: Manifest, A: Manifest]( sc: SparkContext, @@ -14,7 +43,8 @@ object Bagel extends Logging { combiner: Combiner[M, C], aggregator: Option[Aggregator[V, A]], partitioner: Partitioner, - numPartitions: Int + numPartitions: Int, + storageLevel: StorageLevel = DEFAULT_STORAGE_LEVEL )( compute: (V, Option[C], Option[A], Int) => (V, Array[M]) ): RDD[(K, V)] = { @@ -32,8 +62,9 @@ object Bagel extends Logging { val combinedMsgs = msgs.combineByKey( combiner.createCombiner _, combiner.mergeMsg _, combiner.mergeCombiners _, partitioner) val grouped = combinedMsgs.groupWith(verts) + val superstep_ = superstep // Create a read-only copy of superstep for capture in closure val (processed, numMsgs, numActiveVerts) = - comp[K, V, M, C](sc, grouped, compute(_, _, aggregated, superstep)) + comp[K, V, M, C](sc, grouped, compute(_, _, aggregated, superstep_), storageLevel) val timeTaken = System.currentTimeMillis - startTime logInfo("Superstep %d took %d s".format(superstep, timeTaken / 1000)) @@ -50,6 +81,7 @@ object Bagel extends Logging { verts } + /** Runs a Bagel program with no [[spark.bagel.Aggregator]] and the default storage level */ def run[K: Manifest, V <: Vertex : Manifest, M <: Message[K] : Manifest, C: Manifest]( sc: SparkContext, vertices: RDD[(K, V)], @@ -59,12 +91,29 @@ object Bagel extends Logging { numPartitions: Int )( compute: (V, Option[C], Int) => (V, Array[M]) + ): RDD[(K, V)] = run(sc, vertices, messages, combiner, numPartitions, DEFAULT_STORAGE_LEVEL)(compute) + + /** Runs a Bagel program with no [[spark.bagel.Aggregator]] */ + def run[K: Manifest, V <: Vertex : Manifest, M <: Message[K] : Manifest, C: Manifest]( + sc: SparkContext, + vertices: RDD[(K, V)], + messages: RDD[(K, M)], + combiner: Combiner[M, C], + partitioner: Partitioner, + numPartitions: Int, + storageLevel: StorageLevel + )( + compute: (V, Option[C], Int) => (V, Array[M]) ): RDD[(K, V)] = { run[K, V, M, C, Nothing]( - sc, vertices, messages, combiner, None, partitioner, numPartitions)( + sc, vertices, messages, combiner, None, partitioner, numPartitions, storageLevel)( addAggregatorArg[K, V, M, C](compute)) } + /** + * Runs a Bagel program with no [[spark.bagel.Aggregator]], default [[spark.HashPartitioner]] + * and default storage level + */ def run[K: Manifest, V <: Vertex : Manifest, M <: Message[K] : Manifest, C: Manifest]( sc: SparkContext, vertices: RDD[(K, V)], @@ -73,13 +122,29 @@ object Bagel extends Logging { numPartitions: Int )( compute: (V, Option[C], Int) => (V, Array[M]) + ): RDD[(K, V)] = run(sc, vertices, messages, combiner, numPartitions, DEFAULT_STORAGE_LEVEL)(compute) + + /** Runs a Bagel program with no [[spark.bagel.Aggregator]] and the default [[spark.HashPartitioner]]*/ + def run[K: Manifest, V <: Vertex : Manifest, M <: Message[K] : Manifest, C: Manifest]( + sc: SparkContext, + vertices: RDD[(K, V)], + messages: RDD[(K, M)], + combiner: Combiner[M, C], + numPartitions: Int, + storageLevel: StorageLevel + )( + compute: (V, Option[C], Int) => (V, Array[M]) ): RDD[(K, V)] = { val part = new HashPartitioner(numPartitions) run[K, V, M, C, Nothing]( - sc, vertices, messages, combiner, None, part, numPartitions)( + sc, vertices, messages, combiner, None, part, numPartitions, storageLevel)( addAggregatorArg[K, V, M, C](compute)) } + /** + * Runs a Bagel program with no [[spark.bagel.Aggregator]], default [[spark.HashPartitioner]], + * [[spark.bagel.DefaultCombiner]] and the default storage level + */ def run[K: Manifest, V <: Vertex : Manifest, M <: Message[K] : Manifest]( sc: SparkContext, vertices: RDD[(K, V)], @@ -87,10 +152,24 @@ object Bagel extends Logging { numPartitions: Int )( compute: (V, Option[Array[M]], Int) => (V, Array[M]) - ): RDD[(K, V)] = { + ): RDD[(K, V)] = run(sc, vertices, messages, numPartitions, DEFAULT_STORAGE_LEVEL)(compute) + + /** + * Runs a Bagel program with no [[spark.bagel.Aggregator]], the default [[spark.HashPartitioner]] + * and [[spark.bagel.DefaultCombiner]] + */ + def run[K: Manifest, V <: Vertex : Manifest, M <: Message[K] : Manifest]( + sc: SparkContext, + vertices: RDD[(K, V)], + messages: RDD[(K, M)], + numPartitions: Int, + storageLevel: StorageLevel + )( + compute: (V, Option[Array[M]], Int) => (V, Array[M]) + ): RDD[(K, V)] = { val part = new HashPartitioner(numPartitions) run[K, V, M, Array[M], Nothing]( - sc, vertices, messages, new DefaultCombiner(), None, part, numPartitions)( + sc, vertices, messages, new DefaultCombiner(), None, part, numPartitions, storageLevel)( addAggregatorArg[K, V, M, Array[M]](compute)) } @@ -117,7 +196,8 @@ object Bagel extends Logging { private def comp[K: Manifest, V <: Vertex, M <: Message[K], C]( sc: SparkContext, grouped: RDD[(K, (Seq[C], Seq[V]))], - compute: (V, Option[C]) => (V, Array[M]) + compute: (V, Option[C]) => (V, Array[M]), + storageLevel: StorageLevel ): (RDD[(K, (V, Array[M]))], Int, Int) = { var numMsgs = sc.accumulator(0) var numActiveVerts = sc.accumulator(0) @@ -135,7 +215,7 @@ object Bagel extends Logging { numActiveVerts += 1 Some((newVert, newMsgs)) - }.cache + }.persist(storageLevel) // Force evaluation of processed RDD for accurate performance measurements processed.foreach(x => {}) @@ -166,6 +246,7 @@ trait Aggregator[V, A] { def mergeAggregators(a: A, b: A): A } +/** Default combiner that simply appends messages together (i.e. performs no aggregation) */ class DefaultCombiner[M: Manifest] extends Combiner[M, Array[M]] with Serializable { def createCombiner(msg: M): Array[M] = Array(msg) diff --git a/bagel/src/test/scala/bagel/BagelSuite.scala b/bagel/src/test/scala/bagel/BagelSuite.scala index 47829a431e..a09c978068 100644 --- a/bagel/src/test/scala/bagel/BagelSuite.scala +++ b/bagel/src/test/scala/bagel/BagelSuite.scala @@ -7,6 +7,7 @@ import org.scalatest.time.SpanSugar._ import scala.collection.mutable.ArrayBuffer import spark._ +import storage.StorageLevel class TestVertex(val active: Boolean, val age: Int) extends Vertex with Serializable class TestMessage(val targetId: String) extends Message[String] with Serializable @@ -22,6 +23,7 @@ class BagelSuite extends FunSuite with Assertions with BeforeAndAfter with Timeo } // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") } test("halting by voting") { @@ -79,4 +81,21 @@ class BagelSuite extends FunSuite with Assertions with BeforeAndAfter with Timeo } } } + + test("using non-default persistence level") { + failAfter(10 seconds) { + sc = new SparkContext("local", "test") + val verts = sc.parallelize((1 to 4).map(id => (id.toString, new TestVertex(true, 0)))) + val msgs = sc.parallelize(Array[(String, TestMessage)]()) + val numSupersteps = 50 + val result = + Bagel.run(sc, verts, msgs, sc.defaultParallelism, StorageLevel.DISK_ONLY) { + (self: TestVertex, msgs: Option[Array[TestMessage]], superstep: Int) => + (new TestVertex(superstep < numSupersteps - 1, self.age + 1), Array[TestMessage]()) + } + for ((id, vert) <- result.collect) { + assert(vert.age === numSupersteps) + } + } + } } diff --git a/bin/spark-daemon.sh b/bin/spark-daemon.sh index 0c584055c7..8ee3ec481f 100755 --- a/bin/spark-daemon.sh +++ b/bin/spark-daemon.sh @@ -30,7 +30,7 @@ # SPARK_NICENESS The scheduling priority for daemons. Defaults to 0. ## -usage="Usage: spark-daemon.sh [--config <conf-dir>] [--hosts hostlistfile] (start|stop) <spark-command> <args...>" +usage="Usage: spark-daemon.sh [--config <conf-dir>] [--hosts hostlistfile] (start|stop) <spark-command> <spark-instance-number> <args...>" # if no args specified, show usage if [ $# -le 1 ]; then @@ -48,6 +48,8 @@ startStop=$1 shift command=$1 shift +instance=$1 +shift spark_rotate_log () { @@ -92,10 +94,10 @@ if [ "$SPARK_PID_DIR" = "" ]; then fi # some variables -export SPARK_LOGFILE=spark-$SPARK_IDENT_STRING-$command-$HOSTNAME.log +export SPARK_LOGFILE=spark-$SPARK_IDENT_STRING-$command-$instance-$HOSTNAME.log export SPARK_ROOT_LOGGER="INFO,DRFA" -log=$SPARK_LOG_DIR/spark-$SPARK_IDENT_STRING-$command-$HOSTNAME.out -pid=$SPARK_PID_DIR/spark-$SPARK_IDENT_STRING-$command.pid +log=$SPARK_LOG_DIR/spark-$SPARK_IDENT_STRING-$command-$instance-$HOSTNAME.out +pid=$SPARK_PID_DIR/spark-$SPARK_IDENT_STRING-$command-$instance.pid # Set default scheduling priority if [ "$SPARK_NICENESS" = "" ]; then diff --git a/bin/spark-daemons.sh b/bin/spark-daemons.sh index 4f9719ee80..0619097e4d 100755 --- a/bin/spark-daemons.sh +++ b/bin/spark-daemons.sh @@ -2,7 +2,7 @@ # Run a Spark command on all slave hosts. -usage="Usage: spark-daemons.sh [--config confdir] [--hosts hostlistfile] [start|stop] command args..." +usage="Usage: spark-daemons.sh [--config confdir] [--hosts hostlistfile] [start|stop] command instance-number args..." # if no args specified, show usage if [ $# -le 1 ]; then diff --git a/bin/start-master.sh b/bin/start-master.sh index 87feb261fe..83a3e1f3dc 100755 --- a/bin/start-master.sh +++ b/bin/start-master.sh @@ -32,4 +32,4 @@ if [ "$SPARK_PUBLIC_DNS" = "" ]; then fi fi -"$bin"/spark-daemon.sh start spark.deploy.master.Master --ip $SPARK_MASTER_IP --port $SPARK_MASTER_PORT --webui-port $SPARK_MASTER_WEBUI_PORT +"$bin"/spark-daemon.sh start spark.deploy.master.Master 1 --ip $SPARK_MASTER_IP --port $SPARK_MASTER_PORT --webui-port $SPARK_MASTER_WEBUI_PORT diff --git a/bin/start-slave.sh b/bin/start-slave.sh index 45a0cf7a6b..26b5b9d462 100755 --- a/bin/start-slave.sh +++ b/bin/start-slave.sh @@ -6,9 +6,10 @@ bin=`cd "$bin"; pwd` # Set SPARK_PUBLIC_DNS so slaves can be linked in master web UI if [ "$SPARK_PUBLIC_DNS" = "" ]; then # If we appear to be running on EC2, use the public address by default: - if [[ `hostname` == *ec2.internal ]]; then + # NOTE: ec2-metadata is installed on Amazon Linux AMI. Check based on that and hostname + if command -v ec2-metadata > /dev/null || [[ `hostname` == *ec2.internal ]]; then export SPARK_PUBLIC_DNS=`wget -q -O - http://instance-data.ec2.internal/latest/meta-data/public-hostname` fi fi -"$bin"/spark-daemon.sh start spark.deploy.worker.Worker $1 +"$bin"/spark-daemon.sh start spark.deploy.worker.Worker "$@" diff --git a/bin/start-slaves.sh b/bin/start-slaves.sh index 390247ca4a..4e05224190 100755 --- a/bin/start-slaves.sh +++ b/bin/start-slaves.sh @@ -21,4 +21,13 @@ fi echo "Master IP: $SPARK_MASTER_IP" # Launch the slaves -exec "$bin/slaves.sh" cd "$SPARK_HOME" \; "$bin/start-slave.sh" spark://$SPARK_MASTER_IP:$SPARK_MASTER_PORT +if [ "$SPARK_WORKER_INSTANCES" = "" ]; then + exec "$bin/slaves.sh" cd "$SPARK_HOME" \; "$bin/start-slave.sh" 1 spark://$SPARK_MASTER_IP:$SPARK_MASTER_PORT +else + if [ "$SPARK_WORKER_WEBUI_PORT" = "" ]; then + SPARK_WORKER_WEBUI_PORT=8081 + fi + for ((i=0; i<$SPARK_WORKER_INSTANCES; i++)); do + "$bin/slaves.sh" cd "$SPARK_HOME" \; "$bin/start-slave.sh" $(( $i + 1 )) spark://$SPARK_MASTER_IP:$SPARK_MASTER_PORT --webui-port $(( $SPARK_WORKER_WEBUI_PORT + $i )) + done +fi diff --git a/bin/stop-master.sh b/bin/stop-master.sh index f75167dd2c..172ee5891d 100755 --- a/bin/stop-master.sh +++ b/bin/stop-master.sh @@ -7,4 +7,4 @@ bin=`cd "$bin"; pwd` . "$bin/spark-config.sh" -"$bin"/spark-daemon.sh stop spark.deploy.master.Master
\ No newline at end of file +"$bin"/spark-daemon.sh stop spark.deploy.master.Master 1 diff --git a/bin/stop-slaves.sh b/bin/stop-slaves.sh index 21c9ebf324..fbfc594472 100755 --- a/bin/stop-slaves.sh +++ b/bin/stop-slaves.sh @@ -7,4 +7,14 @@ bin=`cd "$bin"; pwd` . "$bin/spark-config.sh" -"$bin"/spark-daemons.sh stop spark.deploy.worker.Worker
\ No newline at end of file +if [ -f "${SPARK_CONF_DIR}/spark-env.sh" ]; then + . "${SPARK_CONF_DIR}/spark-env.sh" +fi + +if [ "$SPARK_WORKER_INSTANCES" = "" ]; then + "$bin"/spark-daemons.sh stop spark.deploy.worker.Worker 1 +else + for ((i=0; i<$SPARK_WORKER_INSTANCES; i++)); do + "$bin"/spark-daemons.sh stop spark.deploy.worker.Worker $(( $i + 1 )) + done +fi diff --git a/conf/spark-env.sh.template b/conf/spark-env.sh.template index 6d71ec5691..37565ca827 100755 --- a/conf/spark-env.sh.template +++ b/conf/spark-env.sh.template @@ -12,6 +12,7 @@ # - SPARK_WORKER_CORES, to set the number of cores to use on this machine # - SPARK_WORKER_MEMORY, to set how much memory to use (e.g. 1000m, 2g) # - SPARK_WORKER_PORT / SPARK_WORKER_WEBUI_PORT +# - SPARK_WORKER_INSTANCES, to set the number of worker instances/processes to be spawned on every slave machine # # Finally, Spark also relies on the following variables, but these can be set # on just the *master* (i.e. in your driver program), and will automatically diff --git a/core/pom.xml b/core/pom.xml index 66c62151fe..88f0ed70f3 100644 --- a/core/pom.xml +++ b/core/pom.xml @@ -3,8 +3,8 @@ <modelVersion>4.0.0</modelVersion> <parent> <groupId>org.spark-project</groupId> - <artifactId>parent</artifactId> - <version>0.7.0-SNAPSHOT</version> + <artifactId>spark-parent</artifactId> + <version>0.8.0-SNAPSHOT</version> <relativePath>../pom.xml</relativePath> </parent> @@ -32,8 +32,8 @@ <artifactId>compress-lzf</artifactId> </dependency> <dependency> - <groupId>asm</groupId> - <artifactId>asm-all</artifactId> + <groupId>org.ow2.asm</groupId> + <artifactId>asm</artifactId> </dependency> <dependency> <groupId>com.google.protobuf</groupId> @@ -73,7 +73,7 @@ </dependency> <dependency> <groupId>cc.spray</groupId> - <artifactId>spray-json_${scala.version}</artifactId> + <artifactId>spray-json_2.9.2</artifactId> </dependency> <dependency> <groupId>org.tomdz.twirl</groupId> @@ -81,14 +81,27 @@ </dependency> <dependency> <groupId>com.github.scala-incubator.io</groupId> - <artifactId>scala-io-file_${scala.version}</artifactId> + <artifactId>scala-io-file_2.9.2</artifactId> </dependency> <dependency> <groupId>org.apache.mesos</groupId> <artifactId>mesos</artifactId> </dependency> + <dependency> + <groupId>io.netty</groupId> + <artifactId>netty-all</artifactId> + </dependency> + <dependency> + <groupId>log4j</groupId> + <artifactId>log4j</artifactId> + </dependency> <dependency> + <groupId>org.apache.derby</groupId> + <artifactId>derby</artifactId> + <scope>test</scope> + </dependency> + <dependency> <groupId>org.scalatest</groupId> <artifactId>scalatest_${scala.version}</artifactId> <scope>test</scope> @@ -275,5 +288,72 @@ </plugins> </build> </profile> + <profile> + <id>hadoop2-yarn</id> + <dependencies> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-api</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-client</artifactId> + <scope>provided</scope> + </dependency> + </dependencies> + <build> + <plugins> + <plugin> + <groupId>org.codehaus.mojo</groupId> + <artifactId>build-helper-maven-plugin</artifactId> + <executions> + <execution> + <id>add-source</id> + <phase>generate-sources</phase> + <goals> + <goal>add-source</goal> + </goals> + <configuration> + <sources> + <source>src/main/scala</source> + <source>src/hadoop2-yarn/scala</source> + </sources> + </configuration> + </execution> + <execution> + <id>add-scala-test-sources</id> + <phase>generate-test-sources</phase> + <goals> + <goal>add-test-source</goal> + </goals> + <configuration> + <sources> + <source>src/test/scala</source> + </sources> + </configuration> + </execution> + </executions> + </plugin> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-jar-plugin</artifactId> + <configuration> + <classifier>hadoop2-yarn</classifier> + </configuration> + </plugin> + </plugins> + </build> + </profile> </profiles> </project> diff --git a/core/src/hadoop1/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala b/core/src/hadoop1/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala index ca9f7219de..f286f2cf9c 100644 --- a/core/src/hadoop1/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala +++ b/core/src/hadoop1/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala @@ -4,4 +4,7 @@ trait HadoopMapRedUtil { def newJobContext(conf: JobConf, jobId: JobID): JobContext = new JobContext(conf, jobId) def newTaskAttemptContext(conf: JobConf, attemptId: TaskAttemptID): TaskAttemptContext = new TaskAttemptContext(conf, attemptId) + + def newTaskAttemptID(jtIdentifier: String, jobId: Int, isMap: Boolean, taskId: Int, attemptId: Int) = new TaskAttemptID(jtIdentifier, + jobId, isMap, taskId, attemptId) } diff --git a/core/src/hadoop1/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala b/core/src/hadoop1/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala index de7b0f81e3..264d421d14 100644 --- a/core/src/hadoop1/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala +++ b/core/src/hadoop1/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala @@ -6,4 +6,7 @@ trait HadoopMapReduceUtil { def newJobContext(conf: Configuration, jobId: JobID): JobContext = new JobContext(conf, jobId) def newTaskAttemptContext(conf: Configuration, attemptId: TaskAttemptID): TaskAttemptContext = new TaskAttemptContext(conf, attemptId) + + def newTaskAttemptID(jtIdentifier: String, jobId: Int, isMap: Boolean, taskId: Int, attemptId: Int) = new TaskAttemptID(jtIdentifier, + jobId, isMap, taskId, attemptId) } diff --git a/core/src/hadoop1/scala/spark/deploy/SparkHadoopUtil.scala b/core/src/hadoop1/scala/spark/deploy/SparkHadoopUtil.scala new file mode 100644 index 0000000000..a0fb4fe25d --- /dev/null +++ b/core/src/hadoop1/scala/spark/deploy/SparkHadoopUtil.scala @@ -0,0 +1,23 @@ +package spark.deploy +import org.apache.hadoop.conf.Configuration + + +/** + * Contains util methods to interact with Hadoop from spark. + */ +object SparkHadoopUtil { + + def getUserNameFromEnvironment(): String = { + // defaulting to -D ... + System.getProperty("user.name") + } + + def runAsUser(func: (Product) => Unit, args: Product) { + + // Add support, if exists - for now, simply run func ! + func(args) + } + + // Return an appropriate (subclass) of Configuration. Creating config can initializes some hadoop subsystems + def newConfiguration(): Configuration = new Configuration() +} diff --git a/core/src/hadoop2-yarn/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala b/core/src/hadoop2-yarn/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala new file mode 100644 index 0000000000..875c0a220b --- /dev/null +++ b/core/src/hadoop2-yarn/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala @@ -0,0 +1,13 @@ + +package org.apache.hadoop.mapred + +import org.apache.hadoop.mapreduce.TaskType + +trait HadoopMapRedUtil { + def newJobContext(conf: JobConf, jobId: JobID): JobContext = new JobContextImpl(conf, jobId) + + def newTaskAttemptContext(conf: JobConf, attemptId: TaskAttemptID): TaskAttemptContext = new TaskAttemptContextImpl(conf, attemptId) + + def newTaskAttemptID(jtIdentifier: String, jobId: Int, isMap: Boolean, taskId: Int, attemptId: Int) = + new TaskAttemptID(jtIdentifier, jobId, if (isMap) TaskType.MAP else TaskType.REDUCE, taskId, attemptId) +} diff --git a/core/src/hadoop2-yarn/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala b/core/src/hadoop2-yarn/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala new file mode 100644 index 0000000000..8bc6fb6dea --- /dev/null +++ b/core/src/hadoop2-yarn/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala @@ -0,0 +1,13 @@ +package org.apache.hadoop.mapreduce + +import org.apache.hadoop.conf.Configuration +import task.{TaskAttemptContextImpl, JobContextImpl} + +trait HadoopMapReduceUtil { + def newJobContext(conf: Configuration, jobId: JobID): JobContext = new JobContextImpl(conf, jobId) + + def newTaskAttemptContext(conf: Configuration, attemptId: TaskAttemptID): TaskAttemptContext = new TaskAttemptContextImpl(conf, attemptId) + + def newTaskAttemptID(jtIdentifier: String, jobId: Int, isMap: Boolean, taskId: Int, attemptId: Int) = + new TaskAttemptID(jtIdentifier, jobId, if (isMap) TaskType.MAP else TaskType.REDUCE, taskId, attemptId) +} diff --git a/core/src/hadoop2-yarn/scala/spark/deploy/SparkHadoopUtil.scala b/core/src/hadoop2-yarn/scala/spark/deploy/SparkHadoopUtil.scala new file mode 100644 index 0000000000..ab1ab9d8a7 --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/deploy/SparkHadoopUtil.scala @@ -0,0 +1,63 @@ +package spark.deploy + +import collection.mutable.HashMap +import org.apache.hadoop.security.UserGroupInformation +import org.apache.hadoop.yarn.conf.YarnConfiguration +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.yarn.api.ApplicationConstants.Environment +import java.security.PrivilegedExceptionAction + +/** + * Contains util methods to interact with Hadoop from spark. + */ +object SparkHadoopUtil { + + val yarnConf = newConfiguration() + + def getUserNameFromEnvironment(): String = { + // defaulting to env if -D is not present ... + val retval = System.getProperty(Environment.USER.name, System.getenv(Environment.USER.name)) + + // If nothing found, default to user we are running as + if (retval == null) System.getProperty("user.name") else retval + } + + def runAsUser(func: (Product) => Unit, args: Product) { + runAsUser(func, args, getUserNameFromEnvironment()) + } + + def runAsUser(func: (Product) => Unit, args: Product, user: String) { + + // println("running as user " + jobUserName) + + UserGroupInformation.setConfiguration(yarnConf) + val appMasterUgi: UserGroupInformation = UserGroupInformation.createRemoteUser(user) + appMasterUgi.doAs(new PrivilegedExceptionAction[AnyRef] { + def run: AnyRef = { + func(args) + // no return value ... + null + } + }) + } + + // Note that all params which start with SPARK are propagated all the way through, so if in yarn mode, this MUST be set to true. + def isYarnMode(): Boolean = { + val yarnMode = System.getProperty("SPARK_YARN_MODE", System.getenv("SPARK_YARN_MODE")) + java.lang.Boolean.valueOf(yarnMode) + } + + // Set an env variable indicating we are running in YARN mode. + // Note that anything with SPARK prefix gets propagated to all (remote) processes + def setYarnMode() { + System.setProperty("SPARK_YARN_MODE", "true") + } + + def setYarnMode(env: HashMap[String, String]) { + env("SPARK_YARN_MODE") = "true" + } + + // Return an appropriate (subclass) of Configuration. Creating config can initializes some hadoop subsystems + // Always create a new config, dont reuse yarnConf. + def newConfiguration(): Configuration = new YarnConfiguration(new Configuration()) +} diff --git a/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ApplicationMaster.scala b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ApplicationMaster.scala new file mode 100644 index 0000000000..aa72c1e5fe --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ApplicationMaster.scala @@ -0,0 +1,329 @@ +package spark.deploy.yarn + +import java.net.Socket +import java.util.concurrent.CopyOnWriteArrayList +import java.util.concurrent.atomic.{AtomicInteger, AtomicReference} +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.net.NetUtils +import org.apache.hadoop.yarn.api._ +import org.apache.hadoop.yarn.api.records._ +import org.apache.hadoop.yarn.api.protocolrecords._ +import org.apache.hadoop.yarn.conf.YarnConfiguration +import org.apache.hadoop.yarn.ipc.YarnRPC +import org.apache.hadoop.yarn.util.{ConverterUtils, Records} +import scala.collection.JavaConversions._ +import spark.{SparkContext, Logging, Utils} +import org.apache.hadoop.security.UserGroupInformation +import java.security.PrivilegedExceptionAction + +class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) extends Logging { + + def this(args: ApplicationMasterArguments) = this(args, new Configuration()) + + private var rpc: YarnRPC = YarnRPC.create(conf) + private var resourceManager: AMRMProtocol = null + private var appAttemptId: ApplicationAttemptId = null + private var userThread: Thread = null + private val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + + private var yarnAllocator: YarnAllocationHandler = null + + def run() { + + // Initialization + val jobUserName = Utils.getUserNameFromEnvironment() + logInfo("running as user " + jobUserName) + + // run as user ... + UserGroupInformation.setConfiguration(yarnConf) + val appMasterUgi: UserGroupInformation = UserGroupInformation.createRemoteUser(jobUserName) + appMasterUgi.doAs(new PrivilegedExceptionAction[AnyRef] { + def run: AnyRef = { + runImpl() + return null + } + }) + } + + private def runImpl() { + + appAttemptId = getApplicationAttemptId() + resourceManager = registerWithResourceManager() + val appMasterResponse: RegisterApplicationMasterResponse = registerApplicationMaster() + + // Compute number of threads for akka + val minimumMemory = appMasterResponse.getMinimumResourceCapability().getMemory() + + if (minimumMemory > 0) { + val mem = args.workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD + val numCore = (mem / minimumMemory) + (if (0 != (mem % minimumMemory)) 1 else 0) + + if (numCore > 0) { + // do not override - hits https://issues.apache.org/jira/browse/HADOOP-8406 + // TODO: Uncomment when hadoop is on a version which has this fixed. + // args.workerCores = numCore + } + } + + // Workaround until hadoop moves to something which has + // https://issues.apache.org/jira/browse/HADOOP-8406 + // ignore result + // This does not, unfortunately, always work reliably ... but alleviates the bug a lot of times + // Hence args.workerCores = numCore disabled above. Any better option ? + // org.apache.hadoop.io.compress.CompressionCodecFactory.getCodecClasses(conf) + + ApplicationMaster.register(this) + // Start the user's JAR + userThread = startUserClass() + + // This a bit hacky, but we need to wait until the spark.driver.port property has + // been set by the Thread executing the user class. + waitForSparkMaster() + + // Allocate all containers + allocateWorkers() + + // Wait for the user class to Finish + userThread.join() + + // Finish the ApplicationMaster + finishApplicationMaster() + // TODO: Exit based on success/failure + System.exit(0) + } + + private def getApplicationAttemptId(): ApplicationAttemptId = { + val envs = System.getenv() + val containerIdString = envs.get(ApplicationConstants.AM_CONTAINER_ID_ENV) + val containerId = ConverterUtils.toContainerId(containerIdString) + val appAttemptId = containerId.getApplicationAttemptId() + logInfo("ApplicationAttemptId: " + appAttemptId) + return appAttemptId + } + + private def registerWithResourceManager(): AMRMProtocol = { + val rmAddress = NetUtils.createSocketAddr(yarnConf.get( + YarnConfiguration.RM_SCHEDULER_ADDRESS, + YarnConfiguration.DEFAULT_RM_SCHEDULER_ADDRESS)) + logInfo("Connecting to ResourceManager at " + rmAddress) + return rpc.getProxy(classOf[AMRMProtocol], rmAddress, conf).asInstanceOf[AMRMProtocol] + } + + private def registerApplicationMaster(): RegisterApplicationMasterResponse = { + logInfo("Registering the ApplicationMaster") + val appMasterRequest = Records.newRecord(classOf[RegisterApplicationMasterRequest]) + .asInstanceOf[RegisterApplicationMasterRequest] + appMasterRequest.setApplicationAttemptId(appAttemptId) + // Setting this to master host,port - so that the ApplicationReport at client has some sensible info. + // Users can then monitor stderr/stdout on that node if required. + appMasterRequest.setHost(Utils.localHostName()) + appMasterRequest.setRpcPort(0) + // What do we provide here ? Might make sense to expose something sensible later ? + appMasterRequest.setTrackingUrl("") + return resourceManager.registerApplicationMaster(appMasterRequest) + } + + private def waitForSparkMaster() { + logInfo("Waiting for spark driver to be reachable.") + var driverUp = false + while(!driverUp) { + val driverHost = System.getProperty("spark.driver.host") + val driverPort = System.getProperty("spark.driver.port") + try { + val socket = new Socket(driverHost, driverPort.toInt) + socket.close() + logInfo("Master now available: " + driverHost + ":" + driverPort) + driverUp = true + } catch { + case e: Exception => + logError("Failed to connect to driver at " + driverHost + ":" + driverPort) + Thread.sleep(100) + } + } + } + + private def startUserClass(): Thread = { + logInfo("Starting the user JAR in a separate Thread") + val mainMethod = Class.forName(args.userClass, false, Thread.currentThread.getContextClassLoader) + .getMethod("main", classOf[Array[String]]) + val t = new Thread { + override def run() { + // Copy + var mainArgs: Array[String] = new Array[String](args.userArgs.size()) + args.userArgs.copyToArray(mainArgs, 0, args.userArgs.size()) + mainMethod.invoke(null, mainArgs) + } + } + t.start() + return t + } + + private def allocateWorkers() { + logInfo("Waiting for spark context initialization") + + try { + var sparkContext: SparkContext = null + ApplicationMaster.sparkContextRef.synchronized { + var count = 0 + while (ApplicationMaster.sparkContextRef.get() == null) { + logInfo("Waiting for spark context initialization ... " + count) + count = count + 1 + ApplicationMaster.sparkContextRef.wait(10000L) + } + sparkContext = ApplicationMaster.sparkContextRef.get() + assert(sparkContext != null) + this.yarnAllocator = YarnAllocationHandler.newAllocator(yarnConf, resourceManager, appAttemptId, args, sparkContext.preferredNodeLocationData) + } + + + logInfo("Allocating " + args.numWorkers + " workers.") + // Wait until all containers have finished + // TODO: This is a bit ugly. Can we make it nicer? + // TODO: Handle container failure + while(yarnAllocator.getNumWorkersRunning < args.numWorkers && + // If user thread exists, then quit ! + userThread.isAlive) { + + this.yarnAllocator.allocateContainers(math.max(args.numWorkers - yarnAllocator.getNumWorkersRunning, 0)) + ApplicationMaster.incrementAllocatorLoop(1) + Thread.sleep(100) + } + } finally { + // in case of exceptions, etc - ensure that count is atleast ALLOCATOR_LOOP_WAIT_COUNT : + // so that the loop (in ApplicationMaster.sparkContextInitialized) breaks + ApplicationMaster.incrementAllocatorLoop(ApplicationMaster.ALLOCATOR_LOOP_WAIT_COUNT) + } + logInfo("All workers have launched.") + + // Launch a progress reporter thread, else app will get killed after expiration (def: 10mins) timeout + if (userThread.isAlive){ + // ensure that progress is sent before YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS elapse. + + val timeoutInterval = yarnConf.getInt(YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS, 120000) + // must be <= timeoutInterval/ 2. + // On other hand, also ensure that we are reasonably responsive without causing too many requests to RM. + // so atleast 1 minute or timeoutInterval / 10 - whichever is higher. + val interval = math.min(timeoutInterval / 2, math.max(timeoutInterval/ 10, 60000L)) + launchReporterThread(interval) + } + } + + // TODO: We might want to extend this to allocate more containers in case they die ! + private def launchReporterThread(_sleepTime: Long): Thread = { + val sleepTime = if (_sleepTime <= 0 ) 0 else _sleepTime + + val t = new Thread { + override def run() { + while (userThread.isAlive){ + val missingWorkerCount = args.numWorkers - yarnAllocator.getNumWorkersRunning + if (missingWorkerCount > 0) { + logInfo("Allocating " + missingWorkerCount + " containers to make up for (potentially ?) lost containers") + yarnAllocator.allocateContainers(missingWorkerCount) + } + else sendProgress() + Thread.sleep(sleepTime) + } + } + } + // setting to daemon status, though this is usually not a good idea. + t.setDaemon(true) + t.start() + logInfo("Started progress reporter thread - sleep time : " + sleepTime) + return t + } + + private def sendProgress() { + logDebug("Sending progress") + // simulated with an allocate request with no nodes requested ... + yarnAllocator.allocateContainers(0) + } + + /* + def printContainers(containers: List[Container]) = { + for (container <- containers) { + logInfo("Launching shell command on a new container." + + ", containerId=" + container.getId() + + ", containerNode=" + container.getNodeId().getHost() + + ":" + container.getNodeId().getPort() + + ", containerNodeURI=" + container.getNodeHttpAddress() + + ", containerState" + container.getState() + + ", containerResourceMemory" + + container.getResource().getMemory()) + } + } + */ + + def finishApplicationMaster() { + val finishReq = Records.newRecord(classOf[FinishApplicationMasterRequest]) + .asInstanceOf[FinishApplicationMasterRequest] + finishReq.setAppAttemptId(appAttemptId) + // TODO: Check if the application has failed or succeeded + finishReq.setFinishApplicationStatus(FinalApplicationStatus.SUCCEEDED) + resourceManager.finishApplicationMaster(finishReq) + } + +} + +object ApplicationMaster { + // number of times to wait for the allocator loop to complete. + // each loop iteration waits for 100ms, so maximum of 3 seconds. + // This is to ensure that we have reasonable number of containers before we start + // TODO: Currently, task to container is computed once (TaskSetManager) - which need not be optimal as more + // containers are available. Might need to handle this better. + private val ALLOCATOR_LOOP_WAIT_COUNT = 30 + def incrementAllocatorLoop(by: Int) { + val count = yarnAllocatorLoop.getAndAdd(by) + if (count >= ALLOCATOR_LOOP_WAIT_COUNT){ + yarnAllocatorLoop.synchronized { + // to wake threads off wait ... + yarnAllocatorLoop.notifyAll() + } + } + } + + private val applicationMasters = new CopyOnWriteArrayList[ApplicationMaster]() + + def register(master: ApplicationMaster) { + applicationMasters.add(master) + } + + val sparkContextRef: AtomicReference[SparkContext] = new AtomicReference[SparkContext](null) + val yarnAllocatorLoop: AtomicInteger = new AtomicInteger(0) + + def sparkContextInitialized(sc: SparkContext): Boolean = { + var modified = false + sparkContextRef.synchronized { + modified = sparkContextRef.compareAndSet(null, sc) + sparkContextRef.notifyAll() + } + + // Add a shutdown hook - as a best case effort in case users do not call sc.stop or do System.exit + // Should not really have to do this, but it helps yarn to evict resources earlier. + // not to mention, prevent Client declaring failure even though we exit'ed properly. + if (modified) { + Runtime.getRuntime().addShutdownHook(new Thread with Logging { + // This is not just to log, but also to ensure that log system is initialized for this instance when we actually are 'run' + logInfo("Adding shutdown hook for context " + sc) + override def run() { + logInfo("Invoking sc stop from shutdown hook") + sc.stop() + // best case ... + for (master <- applicationMasters) master.finishApplicationMaster + } + } ) + } + + // Wait for initialization to complete and atleast 'some' nodes can get allocated + yarnAllocatorLoop.synchronized { + while (yarnAllocatorLoop.get() <= ALLOCATOR_LOOP_WAIT_COUNT){ + yarnAllocatorLoop.wait(1000L) + } + } + modified + } + + def main(argStrings: Array[String]) { + val args = new ApplicationMasterArguments(argStrings) + new ApplicationMaster(args).run() + } +} diff --git a/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ApplicationMasterArguments.scala b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ApplicationMasterArguments.scala new file mode 100644 index 0000000000..1b00208511 --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ApplicationMasterArguments.scala @@ -0,0 +1,77 @@ +package spark.deploy.yarn + +import spark.util.IntParam +import collection.mutable.ArrayBuffer + +class ApplicationMasterArguments(val args: Array[String]) { + var userJar: String = null + var userClass: String = null + var userArgs: Seq[String] = Seq[String]() + var workerMemory = 1024 + var workerCores = 1 + var numWorkers = 2 + + parseArgs(args.toList) + + private def parseArgs(inputArgs: List[String]): Unit = { + val userArgsBuffer = new ArrayBuffer[String]() + + var args = inputArgs + + while (! args.isEmpty) { + + args match { + case ("--jar") :: value :: tail => + userJar = value + args = tail + + case ("--class") :: value :: tail => + userClass = value + args = tail + + case ("--args") :: value :: tail => + userArgsBuffer += value + args = tail + + case ("--num-workers") :: IntParam(value) :: tail => + numWorkers = value + args = tail + + case ("--worker-memory") :: IntParam(value) :: tail => + workerMemory = value + args = tail + + case ("--worker-cores") :: IntParam(value) :: tail => + workerCores = value + args = tail + + case Nil => + if (userJar == null || userClass == null) { + printUsageAndExit(1) + } + + case _ => + printUsageAndExit(1, args) + } + } + + userArgs = userArgsBuffer.readOnly + } + + def printUsageAndExit(exitCode: Int, unknownParam: Any = null) { + if (unknownParam != null) { + System.err.println("Unknown/unsupported param " + unknownParam) + } + System.err.println( + "Usage: spark.deploy.yarn.ApplicationMaster [options] \n" + + "Options:\n" + + " --jar JAR_PATH Path to your application's JAR file (required)\n" + + " --class CLASS_NAME Name of your application's main class (required)\n" + + " --args ARGS Arguments to be passed to your application's main class.\n" + + " Mutliple invocations are possible, each will be passed in order.\n" + + " --num-workers NUM Number of workers to start (Default: 2)\n" + + " --worker-cores NUM Number of cores for the workers (Default: 1)\n" + + " --worker-memory MEM Memory per Worker (e.g. 1000M, 2G) (Default: 1G)\n") + System.exit(exitCode) + } +} diff --git a/core/src/hadoop2-yarn/scala/spark/deploy/yarn/Client.scala b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/Client.scala new file mode 100644 index 0000000000..7a881e26df --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/Client.scala @@ -0,0 +1,272 @@ +package spark.deploy.yarn + +import java.net.{InetSocketAddress, URI} +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.{FileStatus, FileSystem, Path} +import org.apache.hadoop.net.NetUtils +import org.apache.hadoop.yarn.api._ +import org.apache.hadoop.yarn.api.records._ +import org.apache.hadoop.yarn.api.protocolrecords._ +import org.apache.hadoop.yarn.client.YarnClientImpl +import org.apache.hadoop.yarn.conf.YarnConfiguration +import org.apache.hadoop.yarn.ipc.YarnRPC +import scala.collection.mutable.HashMap +import scala.collection.JavaConversions._ +import spark.{Logging, Utils} +import org.apache.hadoop.yarn.util.{Apps, Records, ConverterUtils} +import org.apache.hadoop.yarn.api.ApplicationConstants.Environment +import spark.deploy.SparkHadoopUtil + +class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl with Logging { + + def this(args: ClientArguments) = this(new Configuration(), args) + + var rpc: YarnRPC = YarnRPC.create(conf) + val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + + def run() { + init(yarnConf) + start() + logClusterResourceDetails() + + val newApp = super.getNewApplication() + val appId = newApp.getApplicationId() + + verifyClusterResources(newApp) + val appContext = createApplicationSubmissionContext(appId) + val localResources = prepareLocalResources(appId, "spark") + val env = setupLaunchEnv(localResources) + val amContainer = createContainerLaunchContext(newApp, localResources, env) + + appContext.setQueue(args.amQueue) + appContext.setAMContainerSpec(amContainer) + appContext.setUser(args.amUser) + + submitApp(appContext) + + monitorApplication(appId) + System.exit(0) + } + + + def logClusterResourceDetails() { + val clusterMetrics: YarnClusterMetrics = super.getYarnClusterMetrics + logInfo("Got Cluster metric info from ASM, numNodeManagers=" + clusterMetrics.getNumNodeManagers) + + val queueInfo: QueueInfo = super.getQueueInfo(args.amQueue) + logInfo("Queue info .. queueName=" + queueInfo.getQueueName + ", queueCurrentCapacity=" + queueInfo.getCurrentCapacity + + ", queueMaxCapacity=" + queueInfo.getMaximumCapacity + ", queueApplicationCount=" + queueInfo.getApplications.size + + ", queueChildQueueCount=" + queueInfo.getChildQueues.size) + } + + + def verifyClusterResources(app: GetNewApplicationResponse) = { + val maxMem = app.getMaximumResourceCapability().getMemory() + logInfo("Max mem capabililty of resources in this cluster " + maxMem) + + // If the cluster does not have enough memory resources, exit. + val requestedMem = (args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + args.numWorkers * args.workerMemory + if (requestedMem > maxMem) { + logError("Cluster cannot satisfy memory resource request of " + requestedMem) + System.exit(1) + } + } + + def createApplicationSubmissionContext(appId: ApplicationId): ApplicationSubmissionContext = { + logInfo("Setting up application submission context for ASM") + val appContext = Records.newRecord(classOf[ApplicationSubmissionContext]) + appContext.setApplicationId(appId) + appContext.setApplicationName("Spark") + return appContext + } + + def prepareLocalResources(appId: ApplicationId, appName: String): HashMap[String, LocalResource] = { + logInfo("Preparing Local resources") + val locaResources = HashMap[String, LocalResource]() + // Upload Spark and the application JAR to the remote file system + // Add them as local resources to the AM + val fs = FileSystem.get(conf) + Map("spark.jar" -> System.getenv("SPARK_JAR"), "app.jar" -> args.userJar, "log4j.properties" -> System.getenv("SPARK_LOG4J_CONF")) + .foreach { case(destName, _localPath) => + val localPath: String = if (_localPath != null) _localPath.trim() else "" + if (! localPath.isEmpty()) { + val src = new Path(localPath) + val pathSuffix = appName + "/" + appId.getId() + destName + val dst = new Path(fs.getHomeDirectory(), pathSuffix) + logInfo("Uploading " + src + " to " + dst) + fs.copyFromLocalFile(false, true, src, dst) + val destStatus = fs.getFileStatus(dst) + + val amJarRsrc = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource] + amJarRsrc.setType(LocalResourceType.FILE) + amJarRsrc.setVisibility(LocalResourceVisibility.APPLICATION) + amJarRsrc.setResource(ConverterUtils.getYarnUrlFromPath(dst)) + amJarRsrc.setTimestamp(destStatus.getModificationTime()) + amJarRsrc.setSize(destStatus.getLen()) + locaResources(destName) = amJarRsrc + } + } + return locaResources + } + + def setupLaunchEnv(localResources: HashMap[String, LocalResource]): HashMap[String, String] = { + logInfo("Setting up the launch environment") + val log4jConfLocalRes = localResources.getOrElse("log4j.properties", null) + + val env = new HashMap[String, String]() + Apps.addToEnvironment(env, Environment.USER.name, args.amUser) + + // If log4j present, ensure ours overrides all others + if (log4jConfLocalRes != null) Apps.addToEnvironment(env, Environment.CLASSPATH.name, "./") + + Apps.addToEnvironment(env, Environment.CLASSPATH.name, "./*") + Apps.addToEnvironment(env, Environment.CLASSPATH.name, "$CLASSPATH") + Client.populateHadoopClasspath(yarnConf, env) + SparkHadoopUtil.setYarnMode(env) + env("SPARK_YARN_JAR_PATH") = + localResources("spark.jar").getResource().getScheme.toString() + "://" + + localResources("spark.jar").getResource().getFile().toString() + env("SPARK_YARN_JAR_TIMESTAMP") = localResources("spark.jar").getTimestamp().toString() + env("SPARK_YARN_JAR_SIZE") = localResources("spark.jar").getSize().toString() + + env("SPARK_YARN_USERJAR_PATH") = + localResources("app.jar").getResource().getScheme.toString() + "://" + + localResources("app.jar").getResource().getFile().toString() + env("SPARK_YARN_USERJAR_TIMESTAMP") = localResources("app.jar").getTimestamp().toString() + env("SPARK_YARN_USERJAR_SIZE") = localResources("app.jar").getSize().toString() + + if (log4jConfLocalRes != null) { + env("SPARK_YARN_LOG4J_PATH") = + log4jConfLocalRes.getResource().getScheme.toString() + "://" + log4jConfLocalRes.getResource().getFile().toString() + env("SPARK_YARN_LOG4J_TIMESTAMP") = log4jConfLocalRes.getTimestamp().toString() + env("SPARK_YARN_LOG4J_SIZE") = log4jConfLocalRes.getSize().toString() + } + + // Add each SPARK-* key to the environment + System.getenv().filterKeys(_.startsWith("SPARK")).foreach { case (k,v) => env(k) = v } + return env + } + + def userArgsToString(clientArgs: ClientArguments): String = { + val prefix = " --args " + val args = clientArgs.userArgs + val retval = new StringBuilder() + for (arg <- args){ + retval.append(prefix).append(" '").append(arg).append("' ") + } + + retval.toString + } + + def createContainerLaunchContext(newApp: GetNewApplicationResponse, + localResources: HashMap[String, LocalResource], + env: HashMap[String, String]): ContainerLaunchContext = { + logInfo("Setting up container launch context") + val amContainer = Records.newRecord(classOf[ContainerLaunchContext]) + amContainer.setLocalResources(localResources) + amContainer.setEnvironment(env) + + val minResMemory: Int = newApp.getMinimumResourceCapability().getMemory() + + var amMemory = ((args.amMemory / minResMemory) * minResMemory) + + (if (0 != (args.amMemory % minResMemory)) minResMemory else 0) - YarnAllocationHandler.MEMORY_OVERHEAD + + // Extra options for the JVM + var JAVA_OPTS = "" + + // Add Xmx for am memory + JAVA_OPTS += "-Xmx" + amMemory + "m " + + // Commenting it out for now - so that people can refer to the properties if required. Remove it once cpuset version is pushed out. + // The context is, default gc for server class machines end up using all cores to do gc - hence if there are multiple containers in same + // node, spark gc effects all other containers performance (which can also be other spark containers) + // Instead of using this, rely on cpusets by YARN to enforce spark behaves 'properly' in multi-tenant environments. Not sure how default java gc behaves if it is + // limited to subset of cores on a node. + if (env.isDefinedAt("SPARK_USE_CONC_INCR_GC") && java.lang.Boolean.parseBoolean(env("SPARK_USE_CONC_INCR_GC"))) { + // In our expts, using (default) throughput collector has severe perf ramnifications in multi-tenant machines + JAVA_OPTS += " -XX:+UseConcMarkSweepGC " + JAVA_OPTS += " -XX:+CMSIncrementalMode " + JAVA_OPTS += " -XX:+CMSIncrementalPacing " + JAVA_OPTS += " -XX:CMSIncrementalDutyCycleMin=0 " + JAVA_OPTS += " -XX:CMSIncrementalDutyCycle=10 " + } + if (env.isDefinedAt("SPARK_JAVA_OPTS")) { + JAVA_OPTS += env("SPARK_JAVA_OPTS") + " " + } + + // Command for the ApplicationMaster + val commands = List[String]("java " + + " -server " + + JAVA_OPTS + + " spark.deploy.yarn.ApplicationMaster" + + " --class " + args.userClass + + " --jar " + args.userJar + + userArgsToString(args) + + " --worker-memory " + args.workerMemory + + " --worker-cores " + args.workerCores + + " --num-workers " + args.numWorkers + + " 1> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stdout" + + " 2> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stderr") + logInfo("Command for the ApplicationMaster: " + commands(0)) + amContainer.setCommands(commands) + + val capability = Records.newRecord(classOf[Resource]).asInstanceOf[Resource] + // Memory for the ApplicationMaster + capability.setMemory(args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + amContainer.setResource(capability) + + return amContainer + } + + def submitApp(appContext: ApplicationSubmissionContext) = { + // Submit the application to the applications manager + logInfo("Submitting application to ASM") + super.submitApplication(appContext) + } + + def monitorApplication(appId: ApplicationId): Boolean = { + while(true) { + Thread.sleep(1000) + val report = super.getApplicationReport(appId) + + logInfo("Application report from ASM: \n" + + "\t application identifier: " + appId.toString() + "\n" + + "\t appId: " + appId.getId() + "\n" + + "\t clientToken: " + report.getClientToken() + "\n" + + "\t appDiagnostics: " + report.getDiagnostics() + "\n" + + "\t appMasterHost: " + report.getHost() + "\n" + + "\t appQueue: " + report.getQueue() + "\n" + + "\t appMasterRpcPort: " + report.getRpcPort() + "\n" + + "\t appStartTime: " + report.getStartTime() + "\n" + + "\t yarnAppState: " + report.getYarnApplicationState() + "\n" + + "\t distributedFinalState: " + report.getFinalApplicationStatus() + "\n" + + "\t appTrackingUrl: " + report.getTrackingUrl() + "\n" + + "\t appUser: " + report.getUser() + ) + + val state = report.getYarnApplicationState() + val dsStatus = report.getFinalApplicationStatus() + if (state == YarnApplicationState.FINISHED || + state == YarnApplicationState.FAILED || + state == YarnApplicationState.KILLED) { + return true + } + } + return true + } +} + +object Client { + def main(argStrings: Array[String]) { + val args = new ClientArguments(argStrings) + SparkHadoopUtil.setYarnMode() + new Client(args).run + } + + // Based on code from org.apache.hadoop.mapreduce.v2.util.MRApps + def populateHadoopClasspath(conf: Configuration, env: HashMap[String, String]) { + for (c <- conf.getStrings(YarnConfiguration.YARN_APPLICATION_CLASSPATH)) { + Apps.addToEnvironment(env, Environment.CLASSPATH.name, c.trim) + } + } +} diff --git a/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ClientArguments.scala b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ClientArguments.scala new file mode 100644 index 0000000000..24110558e7 --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/ClientArguments.scala @@ -0,0 +1,105 @@ +package spark.deploy.yarn + +import spark.util.MemoryParam +import spark.util.IntParam +import collection.mutable.{ArrayBuffer, HashMap} +import spark.scheduler.{InputFormatInfo, SplitInfo} + +// TODO: Add code and support for ensuring that yarn resource 'asks' are location aware ! +class ClientArguments(val args: Array[String]) { + var userJar: String = null + var userClass: String = null + var userArgs: Seq[String] = Seq[String]() + var workerMemory = 1024 + var workerCores = 1 + var numWorkers = 2 + var amUser = System.getProperty("user.name") + var amQueue = System.getProperty("QUEUE", "default") + var amMemory: Int = 512 + // TODO + var inputFormatInfo: List[InputFormatInfo] = null + + parseArgs(args.toList) + + private def parseArgs(inputArgs: List[String]): Unit = { + val userArgsBuffer: ArrayBuffer[String] = new ArrayBuffer[String]() + val inputFormatMap: HashMap[String, InputFormatInfo] = new HashMap[String, InputFormatInfo]() + + var args = inputArgs + + while (! args.isEmpty) { + + args match { + case ("--jar") :: value :: tail => + userJar = value + args = tail + + case ("--class") :: value :: tail => + userClass = value + args = tail + + case ("--args") :: value :: tail => + userArgsBuffer += value + args = tail + + case ("--master-memory") :: MemoryParam(value) :: tail => + amMemory = value + args = tail + + case ("--num-workers") :: IntParam(value) :: tail => + numWorkers = value + args = tail + + case ("--worker-memory") :: MemoryParam(value) :: tail => + workerMemory = value + args = tail + + case ("--worker-cores") :: IntParam(value) :: tail => + workerCores = value + args = tail + + case ("--user") :: value :: tail => + amUser = value + args = tail + + case ("--queue") :: value :: tail => + amQueue = value + args = tail + + case Nil => + if (userJar == null || userClass == null) { + printUsageAndExit(1) + } + + case _ => + printUsageAndExit(1, args) + } + } + + userArgs = userArgsBuffer.readOnly + inputFormatInfo = inputFormatMap.values.toList + } + + + def printUsageAndExit(exitCode: Int, unknownParam: Any = null) { + if (unknownParam != null) { + System.err.println("Unknown/unsupported param " + unknownParam) + } + System.err.println( + "Usage: spark.deploy.yarn.Client [options] \n" + + "Options:\n" + + " --jar JAR_PATH Path to your application's JAR file (required)\n" + + " --class CLASS_NAME Name of your application's main class (required)\n" + + " --args ARGS Arguments to be passed to your application's main class.\n" + + " Mutliple invocations are possible, each will be passed in order.\n" + + " --num-workers NUM Number of workers to start (Default: 2)\n" + + " --worker-cores NUM Number of cores for the workers (Default: 1). This is unsused right now.\n" + + " --master-memory MEM Memory for Master (e.g. 1000M, 2G) (Default: 512 Mb)\n" + + " --worker-memory MEM Memory per Worker (e.g. 1000M, 2G) (Default: 1G)\n" + + " --queue QUEUE The hadoop queue to use for allocation requests (Default: 'default')\n" + + " --user USERNAME Run the ApplicationMaster (and slaves) as a different user\n" + ) + System.exit(exitCode) + } + +} diff --git a/core/src/hadoop2-yarn/scala/spark/deploy/yarn/WorkerRunnable.scala b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/WorkerRunnable.scala new file mode 100644 index 0000000000..a2bf0af762 --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/WorkerRunnable.scala @@ -0,0 +1,171 @@ +package spark.deploy.yarn + +import java.net.URI + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.{FileStatus, FileSystem, Path} +import org.apache.hadoop.net.NetUtils +import org.apache.hadoop.security.UserGroupInformation +import org.apache.hadoop.yarn.api._ +import org.apache.hadoop.yarn.api.records._ +import org.apache.hadoop.yarn.api.protocolrecords._ +import org.apache.hadoop.yarn.conf.YarnConfiguration +import org.apache.hadoop.yarn.ipc.YarnRPC +import org.apache.hadoop.yarn.util.{Apps, ConverterUtils, Records} +import org.apache.hadoop.yarn.api.ApplicationConstants.Environment + +import scala.collection.JavaConversions._ +import scala.collection.mutable.HashMap + +import spark.{Logging, Utils} + +class WorkerRunnable(container: Container, conf: Configuration, masterAddress: String, + slaveId: String, hostname: String, workerMemory: Int, workerCores: Int) + extends Runnable with Logging { + + var rpc: YarnRPC = YarnRPC.create(conf) + var cm: ContainerManager = null + val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + + def run = { + logInfo("Starting Worker Container") + cm = connectToCM + startContainer + } + + def startContainer = { + logInfo("Setting up ContainerLaunchContext") + + val ctx = Records.newRecord(classOf[ContainerLaunchContext]) + .asInstanceOf[ContainerLaunchContext] + + ctx.setContainerId(container.getId()) + ctx.setResource(container.getResource()) + val localResources = prepareLocalResources + ctx.setLocalResources(localResources) + + val env = prepareEnvironment + ctx.setEnvironment(env) + + // Extra options for the JVM + var JAVA_OPTS = "" + // Set the JVM memory + val workerMemoryString = workerMemory + "m" + JAVA_OPTS += "-Xms" + workerMemoryString + " -Xmx" + workerMemoryString + " " + if (env.isDefinedAt("SPARK_JAVA_OPTS")) { + JAVA_OPTS += env("SPARK_JAVA_OPTS") + " " + } + // Commenting it out for now - so that people can refer to the properties if required. Remove it once cpuset version is pushed out. + // The context is, default gc for server class machines end up using all cores to do gc - hence if there are multiple containers in same + // node, spark gc effects all other containers performance (which can also be other spark containers) + // Instead of using this, rely on cpusets by YARN to enforce spark behaves 'properly' in multi-tenant environments. Not sure how default java gc behaves if it is + // limited to subset of cores on a node. +/* + else { + // If no java_opts specified, default to using -XX:+CMSIncrementalMode + // It might be possible that other modes/config is being done in SPARK_JAVA_OPTS, so we dont want to mess with it. + // In our expts, using (default) throughput collector has severe perf ramnifications in multi-tennent machines + // The options are based on + // http://www.oracle.com/technetwork/java/gc-tuning-5-138395.html#0.0.0.%20When%20to%20Use%20the%20Concurrent%20Low%20Pause%20Collector|outline + JAVA_OPTS += " -XX:+UseConcMarkSweepGC " + JAVA_OPTS += " -XX:+CMSIncrementalMode " + JAVA_OPTS += " -XX:+CMSIncrementalPacing " + JAVA_OPTS += " -XX:CMSIncrementalDutyCycleMin=0 " + JAVA_OPTS += " -XX:CMSIncrementalDutyCycle=10 " + } +*/ + + ctx.setUser(UserGroupInformation.getCurrentUser().getShortUserName()) + val commands = List[String]("java " + + " -server " + + // Kill if OOM is raised - leverage yarn's failure handling to cause rescheduling. + // Not killing the task leaves various aspects of the worker and (to some extent) the jvm in an inconsistent state. + // TODO: If the OOM is not recoverable by rescheduling it on different node, then do 'something' to fail job ... akin to blacklisting trackers in mapred ? + " -XX:OnOutOfMemoryError='kill %p' " + + JAVA_OPTS + + " spark.executor.StandaloneExecutorBackend " + + masterAddress + " " + + slaveId + " " + + hostname + " " + + workerCores + + " 1> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stdout" + + " 2> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stderr") + logInfo("Setting up worker with commands: " + commands) + ctx.setCommands(commands) + + // Send the start request to the ContainerManager + val startReq = Records.newRecord(classOf[StartContainerRequest]) + .asInstanceOf[StartContainerRequest] + startReq.setContainerLaunchContext(ctx) + cm.startContainer(startReq) + } + + + def prepareLocalResources: HashMap[String, LocalResource] = { + logInfo("Preparing Local resources") + val locaResources = HashMap[String, LocalResource]() + + // Spark JAR + val sparkJarResource = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource] + sparkJarResource.setType(LocalResourceType.FILE) + sparkJarResource.setVisibility(LocalResourceVisibility.APPLICATION) + sparkJarResource.setResource(ConverterUtils.getYarnUrlFromURI( + new URI(System.getenv("SPARK_YARN_JAR_PATH")))) + sparkJarResource.setTimestamp(System.getenv("SPARK_YARN_JAR_TIMESTAMP").toLong) + sparkJarResource.setSize(System.getenv("SPARK_YARN_JAR_SIZE").toLong) + locaResources("spark.jar") = sparkJarResource + // User JAR + val userJarResource = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource] + userJarResource.setType(LocalResourceType.FILE) + userJarResource.setVisibility(LocalResourceVisibility.APPLICATION) + userJarResource.setResource(ConverterUtils.getYarnUrlFromURI( + new URI(System.getenv("SPARK_YARN_USERJAR_PATH")))) + userJarResource.setTimestamp(System.getenv("SPARK_YARN_USERJAR_TIMESTAMP").toLong) + userJarResource.setSize(System.getenv("SPARK_YARN_USERJAR_SIZE").toLong) + locaResources("app.jar") = userJarResource + + // Log4j conf - if available + if (System.getenv("SPARK_YARN_LOG4J_PATH") != null) { + val log4jConfResource = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource] + log4jConfResource.setType(LocalResourceType.FILE) + log4jConfResource.setVisibility(LocalResourceVisibility.APPLICATION) + log4jConfResource.setResource(ConverterUtils.getYarnUrlFromURI( + new URI(System.getenv("SPARK_YARN_LOG4J_PATH")))) + log4jConfResource.setTimestamp(System.getenv("SPARK_YARN_LOG4J_TIMESTAMP").toLong) + log4jConfResource.setSize(System.getenv("SPARK_YARN_LOG4J_SIZE").toLong) + locaResources("log4j.properties") = log4jConfResource + } + + + logInfo("Prepared Local resources " + locaResources) + return locaResources + } + + def prepareEnvironment: HashMap[String, String] = { + val env = new HashMap[String, String]() + // should we add this ? + Apps.addToEnvironment(env, Environment.USER.name, Utils.getUserNameFromEnvironment()) + + // If log4j present, ensure ours overrides all others + if (System.getenv("SPARK_YARN_LOG4J_PATH") != null) { + // Which is correct ? + Apps.addToEnvironment(env, Environment.CLASSPATH.name, "./log4j.properties") + Apps.addToEnvironment(env, Environment.CLASSPATH.name, "./") + } + + Apps.addToEnvironment(env, Environment.CLASSPATH.name, "./*") + Apps.addToEnvironment(env, Environment.CLASSPATH.name, "$CLASSPATH") + Client.populateHadoopClasspath(yarnConf, env) + + System.getenv().filterKeys(_.startsWith("SPARK")).foreach { case (k,v) => env(k) = v } + return env + } + + def connectToCM: ContainerManager = { + val cmHostPortStr = container.getNodeId().getHost() + ":" + container.getNodeId().getPort() + val cmAddress = NetUtils.createSocketAddr(cmHostPortStr) + logInfo("Connecting to ContainerManager at " + cmHostPortStr) + return rpc.getProxy(classOf[ContainerManager], cmAddress, conf).asInstanceOf[ContainerManager] + } + +} diff --git a/core/src/hadoop2-yarn/scala/spark/deploy/yarn/YarnAllocationHandler.scala b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/YarnAllocationHandler.scala new file mode 100644 index 0000000000..61dd72a651 --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/deploy/yarn/YarnAllocationHandler.scala @@ -0,0 +1,547 @@ +package spark.deploy.yarn + +import spark.{Logging, Utils} +import spark.scheduler.SplitInfo +import scala.collection +import org.apache.hadoop.yarn.api.records.{AMResponse, ApplicationAttemptId, ContainerId, Priority, Resource, ResourceRequest, ContainerStatus, Container} +import spark.scheduler.cluster.{ClusterScheduler, StandaloneSchedulerBackend} +import org.apache.hadoop.yarn.api.protocolrecords.{AllocateRequest, AllocateResponse} +import org.apache.hadoop.yarn.util.{RackResolver, Records} +import java.util.concurrent.{CopyOnWriteArrayList, ConcurrentHashMap} +import java.util.concurrent.atomic.AtomicInteger +import org.apache.hadoop.yarn.api.AMRMProtocol +import collection.JavaConversions._ +import collection.mutable.{ArrayBuffer, HashMap, HashSet} +import org.apache.hadoop.conf.Configuration +import java.util.{Collections, Set => JSet} +import java.lang.{Boolean => JBoolean} + +object AllocationType extends Enumeration ("HOST", "RACK", "ANY") { + type AllocationType = Value + val HOST, RACK, ANY = Value +} + +// too many params ? refactor it 'somehow' ? +// needs to be mt-safe +// Need to refactor this to make it 'cleaner' ... right now, all computation is reactive : should make it +// more proactive and decoupled. +// Note that right now, we assume all node asks as uniform in terms of capabilities and priority +// Refer to http://developer.yahoo.com/blogs/hadoop/posts/2011/03/mapreduce-nextgen-scheduler/ for more info +// on how we are requesting for containers. +private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceManager: AMRMProtocol, + val appAttemptId: ApplicationAttemptId, + val maxWorkers: Int, val workerMemory: Int, val workerCores: Int, + val preferredHostToCount: Map[String, Int], + val preferredRackToCount: Map[String, Int]) + extends Logging { + + + // These three are locked on allocatedHostToContainersMap. Complementary data structures + // allocatedHostToContainersMap : containers which are running : host, Set<containerid> + // allocatedContainerToHostMap: container to host mapping + private val allocatedHostToContainersMap = new HashMap[String, collection.mutable.Set[ContainerId]]() + private val allocatedContainerToHostMap = new HashMap[ContainerId, String]() + // allocatedRackCount is populated ONLY if allocation happens (or decremented if this is an allocated node) + // As with the two data structures above, tightly coupled with them, and to be locked on allocatedHostToContainersMap + private val allocatedRackCount = new HashMap[String, Int]() + + // containers which have been released. + private val releasedContainerList = new CopyOnWriteArrayList[ContainerId]() + // containers to be released in next request to RM + private val pendingReleaseContainers = new ConcurrentHashMap[ContainerId, Boolean] + + private val numWorkersRunning = new AtomicInteger() + // Used to generate a unique id per worker + private val workerIdCounter = new AtomicInteger() + private val lastResponseId = new AtomicInteger() + + def getNumWorkersRunning: Int = numWorkersRunning.intValue + + + def isResourceConstraintSatisfied(container: Container): Boolean = { + container.getResource.getMemory >= (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + } + + def allocateContainers(workersToRequest: Int) { + // We need to send the request only once from what I understand ... but for now, not modifying this much. + + // Keep polling the Resource Manager for containers + val amResp = allocateWorkerResources(workersToRequest).getAMResponse + + val _allocatedContainers = amResp.getAllocatedContainers() + if (_allocatedContainers.size > 0) { + + + logDebug("Allocated " + _allocatedContainers.size + " containers, current count " + + numWorkersRunning.get() + ", to-be-released " + releasedContainerList + + ", pendingReleaseContainers : " + pendingReleaseContainers) + logDebug("Cluster Resources: " + amResp.getAvailableResources) + + val hostToContainers = new HashMap[String, ArrayBuffer[Container]]() + + // ignore if not satisfying constraints { + for (container <- _allocatedContainers) { + if (isResourceConstraintSatisfied(container)) { + // allocatedContainers += container + + val host = container.getNodeId.getHost + val containers = hostToContainers.getOrElseUpdate(host, new ArrayBuffer[Container]()) + + containers += container + } + // Add all ignored containers to released list + else releasedContainerList.add(container.getId()) + } + + // Find the appropriate containers to use + // Slightly non trivial groupBy I guess ... + val dataLocalContainers = new HashMap[String, ArrayBuffer[Container]]() + val rackLocalContainers = new HashMap[String, ArrayBuffer[Container]]() + val offRackContainers = new HashMap[String, ArrayBuffer[Container]]() + + for (candidateHost <- hostToContainers.keySet) + { + val maxExpectedHostCount = preferredHostToCount.getOrElse(candidateHost, 0) + val requiredHostCount = maxExpectedHostCount - allocatedContainersOnHost(candidateHost) + + var remainingContainers = hostToContainers.get(candidateHost).getOrElse(null) + assert(remainingContainers != null) + + if (requiredHostCount >= remainingContainers.size){ + // Since we got <= required containers, add all to dataLocalContainers + dataLocalContainers.put(candidateHost, remainingContainers) + // all consumed + remainingContainers = null + } + else if (requiredHostCount > 0) { + // container list has more containers than we need for data locality. + // Split into two : data local container count of (remainingContainers.size - requiredHostCount) + // and rest as remainingContainer + val (dataLocal, remaining) = remainingContainers.splitAt(remainingContainers.size - requiredHostCount) + dataLocalContainers.put(candidateHost, dataLocal) + // remainingContainers = remaining + + // yarn has nasty habit of allocating a tonne of containers on a host - discourage this : + // add remaining to release list. If we have insufficient containers, next allocation cycle + // will reallocate (but wont treat it as data local) + for (container <- remaining) releasedContainerList.add(container.getId()) + remainingContainers = null + } + + // now rack local + if (remainingContainers != null){ + val rack = YarnAllocationHandler.lookupRack(conf, candidateHost) + + if (rack != null){ + val maxExpectedRackCount = preferredRackToCount.getOrElse(rack, 0) + val requiredRackCount = maxExpectedRackCount - allocatedContainersOnRack(rack) - + rackLocalContainers.get(rack).getOrElse(List()).size + + + if (requiredRackCount >= remainingContainers.size){ + // Add all to dataLocalContainers + dataLocalContainers.put(rack, remainingContainers) + // all consumed + remainingContainers = null + } + else if (requiredRackCount > 0) { + // container list has more containers than we need for data locality. + // Split into two : data local container count of (remainingContainers.size - requiredRackCount) + // and rest as remainingContainer + val (rackLocal, remaining) = remainingContainers.splitAt(remainingContainers.size - requiredRackCount) + val existingRackLocal = rackLocalContainers.getOrElseUpdate(rack, new ArrayBuffer[Container]()) + + existingRackLocal ++= rackLocal + remainingContainers = remaining + } + } + } + + // If still not consumed, then it is off rack host - add to that list. + if (remainingContainers != null){ + offRackContainers.put(candidateHost, remainingContainers) + } + } + + // Now that we have split the containers into various groups, go through them in order : + // first host local, then rack local and then off rack (everything else). + // Note that the list we create below tries to ensure that not all containers end up within a host + // if there are sufficiently large number of hosts/containers. + + val allocatedContainers = new ArrayBuffer[Container](_allocatedContainers.size) + allocatedContainers ++= ClusterScheduler.prioritizeContainers(dataLocalContainers) + allocatedContainers ++= ClusterScheduler.prioritizeContainers(rackLocalContainers) + allocatedContainers ++= ClusterScheduler.prioritizeContainers(offRackContainers) + + // Run each of the allocated containers + for (container <- allocatedContainers) { + val numWorkersRunningNow = numWorkersRunning.incrementAndGet() + val workerHostname = container.getNodeId.getHost + val containerId = container.getId + + assert (container.getResource.getMemory >= (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)) + + if (numWorkersRunningNow > maxWorkers) { + logInfo("Ignoring container " + containerId + " at host " + workerHostname + + " .. we already have required number of containers") + releasedContainerList.add(containerId) + // reset counter back to old value. + numWorkersRunning.decrementAndGet() + } + else { + // deallocate + allocate can result in reusing id's wrongly - so use a different counter (workerIdCounter) + val workerId = workerIdCounter.incrementAndGet().toString + val driverUrl = "akka://spark@%s:%s/user/%s".format( + System.getProperty("spark.driver.host"), System.getProperty("spark.driver.port"), + StandaloneSchedulerBackend.ACTOR_NAME) + + logInfo("launching container on " + containerId + " host " + workerHostname) + // just to be safe, simply remove it from pendingReleaseContainers. Should not be there, but .. + pendingReleaseContainers.remove(containerId) + + val rack = YarnAllocationHandler.lookupRack(conf, workerHostname) + allocatedHostToContainersMap.synchronized { + val containerSet = allocatedHostToContainersMap.getOrElseUpdate(workerHostname, new HashSet[ContainerId]()) + + containerSet += containerId + allocatedContainerToHostMap.put(containerId, workerHostname) + if (rack != null) allocatedRackCount.put(rack, allocatedRackCount.getOrElse(rack, 0) + 1) + } + + new Thread( + new WorkerRunnable(container, conf, driverUrl, workerId, + workerHostname, workerMemory, workerCores) + ).start() + } + } + logDebug("After allocated " + allocatedContainers.size + " containers (orig : " + + _allocatedContainers.size + "), current count " + numWorkersRunning.get() + + ", to-be-released " + releasedContainerList + ", pendingReleaseContainers : " + pendingReleaseContainers) + } + + + val completedContainers = amResp.getCompletedContainersStatuses() + if (completedContainers.size > 0){ + logDebug("Completed " + completedContainers.size + " containers, current count " + numWorkersRunning.get() + + ", to-be-released " + releasedContainerList + ", pendingReleaseContainers : " + pendingReleaseContainers) + + for (completedContainer <- completedContainers){ + val containerId = completedContainer.getContainerId + + // Was this released by us ? If yes, then simply remove from containerSet and move on. + if (pendingReleaseContainers.containsKey(containerId)) { + pendingReleaseContainers.remove(containerId) + } + else { + // simply decrement count - next iteration of ReporterThread will take care of allocating ! + numWorkersRunning.decrementAndGet() + logInfo("Container completed ? nodeId: " + containerId + ", state " + completedContainer.getState + + " httpaddress: " + completedContainer.getDiagnostics) + } + + allocatedHostToContainersMap.synchronized { + if (allocatedContainerToHostMap.containsKey(containerId)) { + val host = allocatedContainerToHostMap.get(containerId).getOrElse(null) + assert (host != null) + + val containerSet = allocatedHostToContainersMap.get(host).getOrElse(null) + assert (containerSet != null) + + containerSet -= containerId + if (containerSet.isEmpty) allocatedHostToContainersMap.remove(host) + else allocatedHostToContainersMap.update(host, containerSet) + + allocatedContainerToHostMap -= containerId + + // doing this within locked context, sigh ... move to outside ? + val rack = YarnAllocationHandler.lookupRack(conf, host) + if (rack != null) { + val rackCount = allocatedRackCount.getOrElse(rack, 0) - 1 + if (rackCount > 0) allocatedRackCount.put(rack, rackCount) + else allocatedRackCount.remove(rack) + } + } + } + } + logDebug("After completed " + completedContainers.size + " containers, current count " + + numWorkersRunning.get() + ", to-be-released " + releasedContainerList + + ", pendingReleaseContainers : " + pendingReleaseContainers) + } + } + + def createRackResourceRequests(hostContainers: List[ResourceRequest]): List[ResourceRequest] = { + // First generate modified racks and new set of hosts under it : then issue requests + val rackToCounts = new HashMap[String, Int]() + + // Within this lock - used to read/write to the rack related maps too. + for (container <- hostContainers) { + val candidateHost = container.getHostName + val candidateNumContainers = container.getNumContainers + assert(YarnAllocationHandler.ANY_HOST != candidateHost) + + val rack = YarnAllocationHandler.lookupRack(conf, candidateHost) + if (rack != null) { + var count = rackToCounts.getOrElse(rack, 0) + count += candidateNumContainers + rackToCounts.put(rack, count) + } + } + + val requestedContainers: ArrayBuffer[ResourceRequest] = + new ArrayBuffer[ResourceRequest](rackToCounts.size) + for ((rack, count) <- rackToCounts){ + requestedContainers += + createResourceRequest(AllocationType.RACK, rack, count, YarnAllocationHandler.PRIORITY) + } + + requestedContainers.toList + } + + def allocatedContainersOnHost(host: String): Int = { + var retval = 0 + allocatedHostToContainersMap.synchronized { + retval = allocatedHostToContainersMap.getOrElse(host, Set()).size + } + retval + } + + def allocatedContainersOnRack(rack: String): Int = { + var retval = 0 + allocatedHostToContainersMap.synchronized { + retval = allocatedRackCount.getOrElse(rack, 0) + } + retval + } + + private def allocateWorkerResources(numWorkers: Int): AllocateResponse = { + + var resourceRequests: List[ResourceRequest] = null + + // default. + if (numWorkers <= 0 || preferredHostToCount.isEmpty) { + logDebug("numWorkers: " + numWorkers + ", host preferences ? " + preferredHostToCount.isEmpty) + resourceRequests = List( + createResourceRequest(AllocationType.ANY, null, numWorkers, YarnAllocationHandler.PRIORITY)) + } + else { + // request for all hosts in preferred nodes and for numWorkers - + // candidates.size, request by default allocation policy. + val hostContainerRequests: ArrayBuffer[ResourceRequest] = + new ArrayBuffer[ResourceRequest](preferredHostToCount.size) + for ((candidateHost, candidateCount) <- preferredHostToCount) { + val requiredCount = candidateCount - allocatedContainersOnHost(candidateHost) + + if (requiredCount > 0) { + hostContainerRequests += + createResourceRequest(AllocationType.HOST, candidateHost, requiredCount, YarnAllocationHandler.PRIORITY) + } + } + val rackContainerRequests: List[ResourceRequest] = createRackResourceRequests(hostContainerRequests.toList) + + val anyContainerRequests: ResourceRequest = + createResourceRequest(AllocationType.ANY, null, numWorkers, YarnAllocationHandler.PRIORITY) + + val containerRequests: ArrayBuffer[ResourceRequest] = + new ArrayBuffer[ResourceRequest](hostContainerRequests.size() + rackContainerRequests.size() + 1) + + containerRequests ++= hostContainerRequests + containerRequests ++= rackContainerRequests + containerRequests += anyContainerRequests + + resourceRequests = containerRequests.toList + } + + val req = Records.newRecord(classOf[AllocateRequest]) + req.setResponseId(lastResponseId.incrementAndGet) + req.setApplicationAttemptId(appAttemptId) + + req.addAllAsks(resourceRequests) + + val releasedContainerList = createReleasedContainerList() + req.addAllReleases(releasedContainerList) + + + + if (numWorkers > 0) { + logInfo("Allocating " + numWorkers + " worker containers with " + (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + " of memory each.") + } + else { + logDebug("Empty allocation req .. release : " + releasedContainerList) + } + + for (req <- resourceRequests) { + logInfo("rsrcRequest ... host : " + req.getHostName + ", numContainers : " + req.getNumContainers + + ", p = " + req.getPriority().getPriority + ", capability: " + req.getCapability) + } + resourceManager.allocate(req) + } + + + private def createResourceRequest(requestType: AllocationType.AllocationType, + resource:String, numWorkers: Int, priority: Int): ResourceRequest = { + + // If hostname specified, we need atleast two requests - node local and rack local. + // There must be a third request - which is ANY : that will be specially handled. + requestType match { + case AllocationType.HOST => { + assert (YarnAllocationHandler.ANY_HOST != resource) + + val hostname = resource + val nodeLocal = createResourceRequestImpl(hostname, numWorkers, priority) + + // add to host->rack mapping + YarnAllocationHandler.populateRackInfo(conf, hostname) + + nodeLocal + } + + case AllocationType.RACK => { + val rack = resource + createResourceRequestImpl(rack, numWorkers, priority) + } + + case AllocationType.ANY => { + createResourceRequestImpl(YarnAllocationHandler.ANY_HOST, numWorkers, priority) + } + + case _ => throw new IllegalArgumentException("Unexpected/unsupported request type .. " + requestType) + } + } + + private def createResourceRequestImpl(hostname:String, numWorkers: Int, priority: Int): ResourceRequest = { + + val rsrcRequest = Records.newRecord(classOf[ResourceRequest]) + val memCapability = Records.newRecord(classOf[Resource]) + // There probably is some overhead here, let's reserve a bit more memory. + memCapability.setMemory(workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + rsrcRequest.setCapability(memCapability) + + val pri = Records.newRecord(classOf[Priority]) + pri.setPriority(priority) + rsrcRequest.setPriority(pri) + + rsrcRequest.setHostName(hostname) + + rsrcRequest.setNumContainers(java.lang.Math.max(numWorkers, 0)) + rsrcRequest + } + + def createReleasedContainerList(): ArrayBuffer[ContainerId] = { + + val retval = new ArrayBuffer[ContainerId](1) + // iterator on COW list ... + for (container <- releasedContainerList.iterator()){ + retval += container + } + // remove from the original list. + if (! retval.isEmpty) { + releasedContainerList.removeAll(retval) + for (v <- retval) pendingReleaseContainers.put(v, true) + logInfo("Releasing " + retval.size + " containers. pendingReleaseContainers : " + + pendingReleaseContainers) + } + + retval + } +} + +object YarnAllocationHandler { + + val ANY_HOST = "*" + // all requests are issued with same priority : we do not (yet) have any distinction between + // request types (like map/reduce in hadoop for example) + val PRIORITY = 1 + + // Additional memory overhead - in mb + val MEMORY_OVERHEAD = 384 + + // host to rack map - saved from allocation requests + // We are expecting this not to change. + // Note that it is possible for this to change : and RM will indicate that to us via update + // response to allocate. But we are punting on handling that for now. + private val hostToRack = new ConcurrentHashMap[String, String]() + private val rackToHostSet = new ConcurrentHashMap[String, JSet[String]]() + + def newAllocator(conf: Configuration, + resourceManager: AMRMProtocol, appAttemptId: ApplicationAttemptId, + args: ApplicationMasterArguments, + map: collection.Map[String, collection.Set[SplitInfo]]): YarnAllocationHandler = { + + val (hostToCount, rackToCount) = generateNodeToWeight(conf, map) + + + new YarnAllocationHandler(conf, resourceManager, appAttemptId, args.numWorkers, + args.workerMemory, args.workerCores, hostToCount, rackToCount) + } + + def newAllocator(conf: Configuration, + resourceManager: AMRMProtocol, appAttemptId: ApplicationAttemptId, + maxWorkers: Int, workerMemory: Int, workerCores: Int, + map: collection.Map[String, collection.Set[SplitInfo]]): YarnAllocationHandler = { + + val (hostToCount, rackToCount) = generateNodeToWeight(conf, map) + + new YarnAllocationHandler(conf, resourceManager, appAttemptId, maxWorkers, + workerMemory, workerCores, hostToCount, rackToCount) + } + + // A simple method to copy the split info map. + private def generateNodeToWeight(conf: Configuration, input: collection.Map[String, collection.Set[SplitInfo]]) : + // host to count, rack to count + (Map[String, Int], Map[String, Int]) = { + + if (input == null) return (Map[String, Int](), Map[String, Int]()) + + val hostToCount = new HashMap[String, Int] + val rackToCount = new HashMap[String, Int] + + for ((host, splits) <- input) { + val hostCount = hostToCount.getOrElse(host, 0) + hostToCount.put(host, hostCount + splits.size) + + val rack = lookupRack(conf, host) + if (rack != null){ + val rackCount = rackToCount.getOrElse(host, 0) + rackToCount.put(host, rackCount + splits.size) + } + } + + (hostToCount.toMap, rackToCount.toMap) + } + + def lookupRack(conf: Configuration, host: String): String = { + if (! hostToRack.contains(host)) populateRackInfo(conf, host) + hostToRack.get(host) + } + + def fetchCachedHostsForRack(rack: String): Option[Set[String]] = { + val set = rackToHostSet.get(rack) + if (set == null) return None + + // No better way to get a Set[String] from JSet ? + val convertedSet: collection.mutable.Set[String] = set + Some(convertedSet.toSet) + } + + def populateRackInfo(conf: Configuration, hostname: String) { + Utils.checkHost(hostname) + + if (!hostToRack.containsKey(hostname)) { + // If there are repeated failures to resolve, all to an ignore list ? + val rackInfo = RackResolver.resolve(conf, hostname) + if (rackInfo != null && rackInfo.getNetworkLocation != null) { + val rack = rackInfo.getNetworkLocation + hostToRack.put(hostname, rack) + if (! rackToHostSet.containsKey(rack)) { + rackToHostSet.putIfAbsent(rack, Collections.newSetFromMap(new ConcurrentHashMap[String, JBoolean]())) + } + rackToHostSet.get(rack).add(hostname) + + // Since RackResolver caches, we are disabling this for now ... + } /* else { + // right ? Else we will keep calling rack resolver in case we cant resolve rack info ... + hostToRack.put(hostname, null) + } */ + } + } +} diff --git a/core/src/hadoop2-yarn/scala/spark/scheduler/cluster/YarnClusterScheduler.scala b/core/src/hadoop2-yarn/scala/spark/scheduler/cluster/YarnClusterScheduler.scala new file mode 100644 index 0000000000..ed732d36bf --- /dev/null +++ b/core/src/hadoop2-yarn/scala/spark/scheduler/cluster/YarnClusterScheduler.scala @@ -0,0 +1,42 @@ +package spark.scheduler.cluster + +import spark._ +import spark.deploy.yarn.{ApplicationMaster, YarnAllocationHandler} +import org.apache.hadoop.conf.Configuration + +/** + * + * This is a simple extension to ClusterScheduler - to ensure that appropriate initialization of ApplicationMaster, etc is done + */ +private[spark] class YarnClusterScheduler(sc: SparkContext, conf: Configuration) extends ClusterScheduler(sc) { + + def this(sc: SparkContext) = this(sc, new Configuration()) + + // Nothing else for now ... initialize application master : which needs sparkContext to determine how to allocate + // Note that only the first creation of SparkContext influences (and ideally, there must be only one SparkContext, right ?) + // Subsequent creations are ignored - since nodes are already allocated by then. + + + // By default, rack is unknown + override def getRackForHost(hostPort: String): Option[String] = { + val host = Utils.parseHostPort(hostPort)._1 + val retval = YarnAllocationHandler.lookupRack(conf, host) + if (retval != null) Some(retval) else None + } + + // By default, if rack is unknown, return nothing + override def getCachedHostsForRack(rack: String): Option[Set[String]] = { + if (rack == None || rack == null) return None + + YarnAllocationHandler.fetchCachedHostsForRack(rack) + } + + override def postStartHook() { + val sparkContextInitialized = ApplicationMaster.sparkContextInitialized(sc) + if (sparkContextInitialized){ + // Wait for a few seconds for the slaves to bootstrap and register with master - best case attempt + Thread.sleep(3000L) + } + logInfo("YarnClusterScheduler.postStartHook done") + } +} diff --git a/core/src/hadoop2/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala b/core/src/hadoop2/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala index 35300cea58..a0652d7fc7 100644 --- a/core/src/hadoop2/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala +++ b/core/src/hadoop2/scala/org/apache/hadoop/mapred/HadoopMapRedUtil.scala @@ -4,4 +4,7 @@ trait HadoopMapRedUtil { def newJobContext(conf: JobConf, jobId: JobID): JobContext = new JobContextImpl(conf, jobId) def newTaskAttemptContext(conf: JobConf, attemptId: TaskAttemptID): TaskAttemptContext = new TaskAttemptContextImpl(conf, attemptId) + + def newTaskAttemptID(jtIdentifier: String, jobId: Int, isMap: Boolean, taskId: Int, attemptId: Int) = new TaskAttemptID(jtIdentifier, + jobId, isMap, taskId, attemptId) } diff --git a/core/src/hadoop2/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala b/core/src/hadoop2/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala index 7afdbff320..7fdbe322fd 100644 --- a/core/src/hadoop2/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala +++ b/core/src/hadoop2/scala/org/apache/hadoop/mapreduce/HadoopMapReduceUtil.scala @@ -7,4 +7,7 @@ trait HadoopMapReduceUtil { def newJobContext(conf: Configuration, jobId: JobID): JobContext = new JobContextImpl(conf, jobId) def newTaskAttemptContext(conf: Configuration, attemptId: TaskAttemptID): TaskAttemptContext = new TaskAttemptContextImpl(conf, attemptId) + + def newTaskAttemptID(jtIdentifier: String, jobId: Int, isMap: Boolean, taskId: Int, attemptId: Int) = new TaskAttemptID(jtIdentifier, + jobId, isMap, taskId, attemptId) } diff --git a/core/src/hadoop2/scala/spark/deploy/SparkHadoopUtil.scala b/core/src/hadoop2/scala/spark/deploy/SparkHadoopUtil.scala new file mode 100644 index 0000000000..a0fb4fe25d --- /dev/null +++ b/core/src/hadoop2/scala/spark/deploy/SparkHadoopUtil.scala @@ -0,0 +1,23 @@ +package spark.deploy +import org.apache.hadoop.conf.Configuration + + +/** + * Contains util methods to interact with Hadoop from spark. + */ +object SparkHadoopUtil { + + def getUserNameFromEnvironment(): String = { + // defaulting to -D ... + System.getProperty("user.name") + } + + def runAsUser(func: (Product) => Unit, args: Product) { + + // Add support, if exists - for now, simply run func ! + func(args) + } + + // Return an appropriate (subclass) of Configuration. Creating config can initializes some hadoop subsystems + def newConfiguration(): Configuration = new Configuration() +} diff --git a/core/src/main/java/spark/network/netty/FileClient.java b/core/src/main/java/spark/network/netty/FileClient.java new file mode 100644 index 0000000000..a4bb4bc701 --- /dev/null +++ b/core/src/main/java/spark/network/netty/FileClient.java @@ -0,0 +1,72 @@ +package spark.network.netty; + +import io.netty.bootstrap.Bootstrap; +import io.netty.channel.Channel; +import io.netty.channel.ChannelFuture; +import io.netty.channel.ChannelFutureListener; +import io.netty.channel.ChannelOption; +import io.netty.channel.oio.OioEventLoopGroup; +import io.netty.channel.socket.oio.OioSocketChannel; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +class FileClient { + + private Logger LOG = LoggerFactory.getLogger(this.getClass().getName()); + private FileClientHandler handler = null; + private Channel channel = null; + private Bootstrap bootstrap = null; + private int connectTimeout = 60*1000; // 1 min + + public FileClient(FileClientHandler handler, int connectTimeout) { + this.handler = handler; + this.connectTimeout = connectTimeout; + } + + public void init() { + bootstrap = new Bootstrap(); + bootstrap.group(new OioEventLoopGroup()) + .channel(OioSocketChannel.class) + .option(ChannelOption.SO_KEEPALIVE, true) + .option(ChannelOption.TCP_NODELAY, true) + .option(ChannelOption.CONNECT_TIMEOUT_MILLIS, connectTimeout) + .handler(new FileClientChannelInitializer(handler)); + } + + public void connect(String host, int port) { + try { + // Start the connection attempt. + channel = bootstrap.connect(host, port).sync().channel(); + // ChannelFuture cf = channel.closeFuture(); + //cf.addListener(new ChannelCloseListener(this)); + } catch (InterruptedException e) { + close(); + } + } + + public void waitForClose() { + try { + channel.closeFuture().sync(); + } catch (InterruptedException e) { + LOG.warn("FileClient interrupted", e); + } + } + + public void sendRequest(String file) { + //assert(file == null); + //assert(channel == null); + channel.write(file + "\r\n"); + } + + public void close() { + if(channel != null) { + channel.close(); + channel = null; + } + if ( bootstrap!=null) { + bootstrap.shutdown(); + bootstrap = null; + } + } +} diff --git a/core/src/main/java/spark/network/netty/FileClientChannelInitializer.java b/core/src/main/java/spark/network/netty/FileClientChannelInitializer.java new file mode 100644 index 0000000000..af25baf641 --- /dev/null +++ b/core/src/main/java/spark/network/netty/FileClientChannelInitializer.java @@ -0,0 +1,24 @@ +package spark.network.netty; + +import io.netty.buffer.BufType; +import io.netty.channel.ChannelInitializer; +import io.netty.channel.socket.SocketChannel; +import io.netty.handler.codec.string.StringEncoder; + + +class FileClientChannelInitializer extends ChannelInitializer<SocketChannel> { + + private FileClientHandler fhandler; + + public FileClientChannelInitializer(FileClientHandler handler) { + fhandler = handler; + } + + @Override + public void initChannel(SocketChannel channel) { + // file no more than 2G + channel.pipeline() + .addLast("encoder", new StringEncoder(BufType.BYTE)) + .addLast("handler", fhandler); + } +} diff --git a/core/src/main/java/spark/network/netty/FileClientHandler.java b/core/src/main/java/spark/network/netty/FileClientHandler.java new file mode 100644 index 0000000000..9fc9449827 --- /dev/null +++ b/core/src/main/java/spark/network/netty/FileClientHandler.java @@ -0,0 +1,43 @@ +package spark.network.netty; + +import io.netty.buffer.ByteBuf; +import io.netty.channel.ChannelHandlerContext; +import io.netty.channel.ChannelInboundByteHandlerAdapter; + + +abstract class FileClientHandler extends ChannelInboundByteHandlerAdapter { + + private FileHeader currentHeader = null; + + private volatile boolean handlerCalled = false; + + public boolean isComplete() { + return handlerCalled; + } + + public abstract void handle(ChannelHandlerContext ctx, ByteBuf in, FileHeader header); + public abstract void handleError(String blockId); + + @Override + public ByteBuf newInboundBuffer(ChannelHandlerContext ctx) { + // Use direct buffer if possible. + return ctx.alloc().ioBuffer(); + } + + @Override + public void inboundBufferUpdated(ChannelHandlerContext ctx, ByteBuf in) { + // get header + if (currentHeader == null && in.readableBytes() >= FileHeader.HEADER_SIZE()) { + currentHeader = FileHeader.create(in.readBytes(FileHeader.HEADER_SIZE())); + } + // get file + if(in.readableBytes() >= currentHeader.fileLen()) { + handle(ctx, in, currentHeader); + handlerCalled = true; + currentHeader = null; + ctx.close(); + } + } + +} + diff --git a/core/src/main/java/spark/network/netty/FileServer.java b/core/src/main/java/spark/network/netty/FileServer.java new file mode 100644 index 0000000000..dd3a557ae5 --- /dev/null +++ b/core/src/main/java/spark/network/netty/FileServer.java @@ -0,0 +1,86 @@ +package spark.network.netty; + +import java.net.InetSocketAddress; + +import io.netty.bootstrap.ServerBootstrap; +import io.netty.channel.Channel; +import io.netty.channel.ChannelFuture; +import io.netty.channel.ChannelOption; +import io.netty.channel.oio.OioEventLoopGroup; +import io.netty.channel.socket.oio.OioServerSocketChannel; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + + +/** + * Server that accept the path of a file an echo back its content. + */ +class FileServer { + + private Logger LOG = LoggerFactory.getLogger(this.getClass().getName()); + + private ServerBootstrap bootstrap = null; + private ChannelFuture channelFuture = null; + private int port = 0; + private Thread blockingThread = null; + + public FileServer(PathResolver pResolver, int port) { + InetSocketAddress addr = new InetSocketAddress(port); + + // Configure the server. + bootstrap = new ServerBootstrap(); + bootstrap.group(new OioEventLoopGroup(), new OioEventLoopGroup()) + .channel(OioServerSocketChannel.class) + .option(ChannelOption.SO_BACKLOG, 100) + .option(ChannelOption.SO_RCVBUF, 1500) + .childHandler(new FileServerChannelInitializer(pResolver)); + // Start the server. + channelFuture = bootstrap.bind(addr); + try { + // Get the address we bound to. + InetSocketAddress boundAddress = + ((InetSocketAddress) channelFuture.sync().channel().localAddress()); + this.port = boundAddress.getPort(); + } catch (InterruptedException ie) { + this.port = 0; + } + } + + /** + * Start the file server asynchronously in a new thread. + */ + public void start() { + blockingThread = new Thread() { + public void run() { + try { + channelFuture.channel().closeFuture().sync(); + LOG.info("FileServer exiting"); + } catch (InterruptedException e) { + LOG.error("File server start got interrupted", e); + } + // NOTE: bootstrap is shutdown in stop() + } + }; + blockingThread.setDaemon(true); + blockingThread.start(); + } + + public int getPort() { + return port; + } + + public void stop() { + // Close the bound channel. + if (channelFuture != null) { + channelFuture.channel().close(); + channelFuture = null; + } + // Shutdown bootstrap. + if (bootstrap != null) { + bootstrap.shutdown(); + bootstrap = null; + } + // TODO: Shutdown all accepted channels as well ? + } +} diff --git a/core/src/main/java/spark/network/netty/FileServerChannelInitializer.java b/core/src/main/java/spark/network/netty/FileServerChannelInitializer.java new file mode 100644 index 0000000000..8f1f5c65cd --- /dev/null +++ b/core/src/main/java/spark/network/netty/FileServerChannelInitializer.java @@ -0,0 +1,25 @@ +package spark.network.netty; + +import io.netty.channel.ChannelInitializer; +import io.netty.channel.socket.SocketChannel; +import io.netty.handler.codec.DelimiterBasedFrameDecoder; +import io.netty.handler.codec.Delimiters; +import io.netty.handler.codec.string.StringDecoder; + + +class FileServerChannelInitializer extends ChannelInitializer<SocketChannel> { + + PathResolver pResolver; + + public FileServerChannelInitializer(PathResolver pResolver) { + this.pResolver = pResolver; + } + + @Override + public void initChannel(SocketChannel channel) { + channel.pipeline() + .addLast("framer", new DelimiterBasedFrameDecoder(8192, Delimiters.lineDelimiter())) + .addLast("strDecoder", new StringDecoder()) + .addLast("handler", new FileServerHandler(pResolver)); + } +} diff --git a/core/src/main/java/spark/network/netty/FileServerHandler.java b/core/src/main/java/spark/network/netty/FileServerHandler.java new file mode 100644 index 0000000000..a78eddb1b5 --- /dev/null +++ b/core/src/main/java/spark/network/netty/FileServerHandler.java @@ -0,0 +1,65 @@ +package spark.network.netty; + +import java.io.File; +import java.io.FileInputStream; + +import io.netty.channel.ChannelHandlerContext; +import io.netty.channel.ChannelInboundMessageHandlerAdapter; +import io.netty.channel.DefaultFileRegion; + + +class FileServerHandler extends ChannelInboundMessageHandlerAdapter<String> { + + PathResolver pResolver; + + public FileServerHandler(PathResolver pResolver){ + this.pResolver = pResolver; + } + + @Override + public void messageReceived(ChannelHandlerContext ctx, String blockId) { + String path = pResolver.getAbsolutePath(blockId); + // if getFilePath returns null, close the channel + if (path == null) { + //ctx.close(); + return; + } + File file = new File(path); + if (file.exists()) { + if (!file.isFile()) { + //logger.info("Not a file : " + file.getAbsolutePath()); + ctx.write(new FileHeader(0, blockId).buffer()); + ctx.flush(); + return; + } + long length = file.length(); + if (length > Integer.MAX_VALUE || length <= 0) { + //logger.info("too large file : " + file.getAbsolutePath() + " of size "+ length); + ctx.write(new FileHeader(0, blockId).buffer()); + ctx.flush(); + return; + } + int len = new Long(length).intValue(); + //logger.info("Sending block "+blockId+" filelen = "+len); + //logger.info("header = "+ (new FileHeader(len, blockId)).buffer()); + ctx.write((new FileHeader(len, blockId)).buffer()); + try { + ctx.sendFile(new DefaultFileRegion(new FileInputStream(file) + .getChannel(), 0, file.length())); + } catch (Exception e) { + //logger.warning("Exception when sending file : " + file.getAbsolutePath()); + e.printStackTrace(); + } + } else { + //logger.warning("File not found: " + file.getAbsolutePath()); + ctx.write(new FileHeader(0, blockId).buffer()); + } + ctx.flush(); + } + + @Override + public void exceptionCaught(ChannelHandlerContext ctx, Throwable cause) { + cause.printStackTrace(); + ctx.close(); + } +} diff --git a/core/src/main/java/spark/network/netty/PathResolver.java b/core/src/main/java/spark/network/netty/PathResolver.java new file mode 100755 index 0000000000..302411672c --- /dev/null +++ b/core/src/main/java/spark/network/netty/PathResolver.java @@ -0,0 +1,12 @@ +package spark.network.netty;
+
+
+public interface PathResolver {
+ /**
+ * Get the absolute path of the file
+ *
+ * @param fileId
+ * @return the absolute path of file
+ */
+ public String getAbsolutePath(String fileId);
+}
diff --git a/core/src/main/scala/spark/BlockStoreShuffleFetcher.scala b/core/src/main/scala/spark/BlockStoreShuffleFetcher.scala index 86432d0127..3239f4c385 100644 --- a/core/src/main/scala/spark/BlockStoreShuffleFetcher.scala +++ b/core/src/main/scala/spark/BlockStoreShuffleFetcher.scala @@ -3,18 +3,25 @@ package spark import scala.collection.mutable.ArrayBuffer import scala.collection.mutable.HashMap +import spark.executor.{ShuffleReadMetrics, TaskMetrics} +import spark.serializer.Serializer import spark.storage.BlockManagerId +import spark.util.CompletionIterator + private[spark] class BlockStoreShuffleFetcher extends ShuffleFetcher with Logging { - override def fetch[K, V](shuffleId: Int, reduceId: Int) = { + + override def fetch[K, V]( + shuffleId: Int, reduceId: Int, metrics: TaskMetrics, serializer: Serializer) = { + logDebug("Fetching outputs for shuffle %d, reduce %d".format(shuffleId, reduceId)) val blockManager = SparkEnv.get.blockManager - + val startTime = System.currentTimeMillis val statuses = SparkEnv.get.mapOutputTracker.getServerStatuses(shuffleId, reduceId) logDebug("Fetching map output location for shuffle %d, reduce %d took %d ms".format( shuffleId, reduceId, System.currentTimeMillis - startTime)) - + val splitsByAddress = new HashMap[BlockManagerId, ArrayBuffer[(Int, Long)]] for (((address, size), index) <- statuses.zipWithIndex) { splitsByAddress.getOrElseUpdate(address, ArrayBuffer()) += ((index, size)) @@ -45,6 +52,20 @@ private[spark] class BlockStoreShuffleFetcher extends ShuffleFetcher with Loggin } } } - blockManager.getMultiple(blocksByAddress).flatMap(unpackBlock) + + val blockFetcherItr = blockManager.getMultiple(blocksByAddress, serializer) + val itr = blockFetcherItr.flatMap(unpackBlock) + + CompletionIterator[(K,V), Iterator[(K,V)]](itr, { + val shuffleMetrics = new ShuffleReadMetrics + shuffleMetrics.shuffleFinishTime = System.currentTimeMillis + shuffleMetrics.remoteFetchTime = blockFetcherItr.remoteFetchTime + shuffleMetrics.fetchWaitTime = blockFetcherItr.fetchWaitTime + shuffleMetrics.remoteBytesRead = blockFetcherItr.remoteBytesRead + shuffleMetrics.totalBlocksFetched = blockFetcherItr.totalBlocks + shuffleMetrics.localBlocksFetched = blockFetcherItr.numLocalBlocks + shuffleMetrics.remoteBlocksFetched = blockFetcherItr.numRemoteBlocks + metrics.shuffleReadMetrics = Some(shuffleMetrics) + }) } } diff --git a/core/src/main/scala/spark/CacheManager.scala b/core/src/main/scala/spark/CacheManager.scala index c7b379a3fb..f7a2b7e802 100644 --- a/core/src/main/scala/spark/CacheManager.scala +++ b/core/src/main/scala/spark/CacheManager.scala @@ -27,7 +27,7 @@ private[spark] class CacheManager(blockManager: BlockManager) extends Logging { if (loading.contains(key)) { logInfo("Loading contains " + key + ", waiting...") while (loading.contains(key)) { - try {loading.wait()} catch {case _ =>} + try {loading.wait()} catch {case _ : Throwable =>} } logInfo("Loading no longer contains " + key + ", so returning cached result") // See whether someone else has successfully loaded it. The main way this would fail diff --git a/core/src/main/scala/spark/ClosureCleaner.scala b/core/src/main/scala/spark/ClosureCleaner.scala index 98525b99c8..d5e7132ff9 100644 --- a/core/src/main/scala/spark/ClosureCleaner.scala +++ b/core/src/main/scala/spark/ClosureCleaner.scala @@ -5,15 +5,22 @@ import java.lang.reflect.Field import scala.collection.mutable.Map import scala.collection.mutable.Set -import org.objectweb.asm.{ClassReader, MethodVisitor, Type} -import org.objectweb.asm.commons.EmptyVisitor +import org.objectweb.asm.{ClassReader, ClassVisitor, MethodVisitor, Type} import org.objectweb.asm.Opcodes._ +import java.io.{InputStream, IOException, ByteArrayOutputStream, ByteArrayInputStream, BufferedInputStream} private[spark] object ClosureCleaner extends Logging { // Get an ASM class reader for a given class from the JAR that loaded it private def getClassReader(cls: Class[_]): ClassReader = { - new ClassReader(cls.getResourceAsStream( - cls.getName.replaceFirst("^.*\\.", "") + ".class")) + // Copy data over, before delegating to ClassReader - else we can run out of open file handles. + val className = cls.getName.replaceFirst("^.*\\.", "") + ".class" + val resourceStream = cls.getResourceAsStream(className) + // todo: Fixme - continuing with earlier behavior ... + if (resourceStream == null) return new ClassReader(resourceStream) + + val baos = new ByteArrayOutputStream(128) + Utils.copyStream(resourceStream, baos, true) + new ClassReader(new ByteArrayInputStream(baos.toByteArray)) } // Check whether a class represents a Scala closure @@ -154,10 +161,10 @@ private[spark] object ClosureCleaner extends Logging { } } -private[spark] class FieldAccessFinder(output: Map[Class[_], Set[String]]) extends EmptyVisitor { +private[spark] class FieldAccessFinder(output: Map[Class[_], Set[String]]) extends ClassVisitor(ASM4) { override def visitMethod(access: Int, name: String, desc: String, sig: String, exceptions: Array[String]): MethodVisitor = { - return new EmptyVisitor { + return new MethodVisitor(ASM4) { override def visitFieldInsn(op: Int, owner: String, name: String, desc: String) { if (op == GETFIELD) { for (cl <- output.keys if cl.getName == owner.replace('/', '.')) { @@ -180,7 +187,7 @@ private[spark] class FieldAccessFinder(output: Map[Class[_], Set[String]]) exten } } -private[spark] class InnerClosureFinder(output: Set[Class[_]]) extends EmptyVisitor { +private[spark] class InnerClosureFinder(output: Set[Class[_]]) extends ClassVisitor(ASM4) { var myName: String = null override def visit(version: Int, access: Int, name: String, sig: String, @@ -190,7 +197,7 @@ private[spark] class InnerClosureFinder(output: Set[Class[_]]) extends EmptyVisi override def visitMethod(access: Int, name: String, desc: String, sig: String, exceptions: Array[String]): MethodVisitor = { - return new EmptyVisitor { + return new MethodVisitor(ASM4) { override def visitMethodInsn(op: Int, owner: String, name: String, desc: String) { val argTypes = Type.getArgumentTypes(desc) diff --git a/core/src/main/scala/spark/Dependency.scala b/core/src/main/scala/spark/Dependency.scala index 5eea907322..2af44aa383 100644 --- a/core/src/main/scala/spark/Dependency.scala +++ b/core/src/main/scala/spark/Dependency.scala @@ -25,10 +25,12 @@ abstract class NarrowDependency[T](rdd: RDD[T]) extends Dependency(rdd) { * @param shuffleId the shuffle id * @param rdd the parent RDD * @param partitioner partitioner used to partition the shuffle output + * @param serializerClass class name of the serializer to use */ class ShuffleDependency[K, V]( @transient rdd: RDD[(K, V)], - val partitioner: Partitioner) + val partitioner: Partitioner, + val serializerClass: String = null) extends Dependency(rdd) { val shuffleId: Int = rdd.context.newShuffleId() diff --git a/core/src/main/scala/spark/FetchFailedException.scala b/core/src/main/scala/spark/FetchFailedException.scala index a953081d24..40b0193f19 100644 --- a/core/src/main/scala/spark/FetchFailedException.scala +++ b/core/src/main/scala/spark/FetchFailedException.scala @@ -3,18 +3,25 @@ package spark import spark.storage.BlockManagerId private[spark] class FetchFailedException( - val bmAddress: BlockManagerId, - val shuffleId: Int, - val mapId: Int, - val reduceId: Int, + taskEndReason: TaskEndReason, + message: String, cause: Throwable) extends Exception { - - override def getMessage(): String = - "Fetch failed: %s %d %d %d".format(bmAddress, shuffleId, mapId, reduceId) + + def this (bmAddress: BlockManagerId, shuffleId: Int, mapId: Int, reduceId: Int, cause: Throwable) = + this(FetchFailed(bmAddress, shuffleId, mapId, reduceId), + "Fetch failed: %s %d %d %d".format(bmAddress, shuffleId, mapId, reduceId), + cause) + + def this (shuffleId: Int, reduceId: Int, cause: Throwable) = + this(FetchFailed(null, shuffleId, -1, reduceId), + "Unable to fetch locations from master: %d %d".format(shuffleId, reduceId), cause) + + override def getMessage(): String = message + override def getCause(): Throwable = cause - def toTaskEndReason: TaskEndReason = - FetchFailed(bmAddress, shuffleId, mapId, reduceId) + def toTaskEndReason: TaskEndReason = taskEndReason + } diff --git a/core/src/main/scala/spark/HadoopWriter.scala b/core/src/main/scala/spark/HadoopWriter.scala index afcf9f6db4..5e8396edb9 100644 --- a/core/src/main/scala/spark/HadoopWriter.scala +++ b/core/src/main/scala/spark/HadoopWriter.scala @@ -2,14 +2,10 @@ package org.apache.hadoop.mapred import org.apache.hadoop.fs.FileSystem import org.apache.hadoop.fs.Path -import org.apache.hadoop.util.ReflectionUtils -import org.apache.hadoop.io.NullWritable -import org.apache.hadoop.io.Text import java.text.SimpleDateFormat import java.text.NumberFormat import java.io.IOException -import java.net.URI import java.util.Date import spark.Logging @@ -24,7 +20,7 @@ import spark.SerializableWritable * a filename to write to, etc, exactly like in a Hadoop MapReduce job. */ class HadoopWriter(@transient jobConf: JobConf) extends Logging with HadoopMapRedUtil with Serializable { - + private val now = new Date() private val conf = new SerializableWritable(jobConf) @@ -106,6 +102,12 @@ class HadoopWriter(@transient jobConf: JobConf) extends Logging with HadoopMapRe } } + def commitJob() { + // always ? Or if cmtr.needsTaskCommit ? + val cmtr = getOutputCommitter() + cmtr.commitJob(getJobContext()) + } + def cleanup() { getOutputCommitter().cleanupJob(getJobContext()) } diff --git a/core/src/main/scala/spark/KryoSerializer.scala b/core/src/main/scala/spark/KryoSerializer.scala index 0bd73e936b..d723ab7b1e 100644 --- a/core/src/main/scala/spark/KryoSerializer.scala +++ b/core/src/main/scala/spark/KryoSerializer.scala @@ -157,27 +157,34 @@ class KryoSerializer extends spark.serializer.Serializer with Logging { // Register maps with a special serializer since they have complex internal structure class ScalaMapSerializer(buildMap: Array[(Any, Any)] => scala.collection.Map[Any, Any]) - extends KSerializer[Array[(Any, Any)] => scala.collection.Map[Any, Any]] { + extends KSerializer[Array[(Any, Any)] => scala.collection.Map[Any, Any]] { + + //hack, look at https://groups.google.com/forum/#!msg/kryo-users/Eu5V4bxCfws/k-8UQ22y59AJ + private final val FAKE_REFERENCE = new Object() override def write( - kryo: Kryo, - output: KryoOutput, - obj: Array[(Any, Any)] => scala.collection.Map[Any, Any]) { + kryo: Kryo, + output: KryoOutput, + obj: Array[(Any, Any)] => scala.collection.Map[Any, Any]) { val map = obj.asInstanceOf[scala.collection.Map[Any, Any]] - kryo.writeObject(output, map.size.asInstanceOf[java.lang.Integer]) + output.writeInt(map.size) for ((k, v) <- map) { kryo.writeClassAndObject(output, k) kryo.writeClassAndObject(output, v) } } override def read ( - kryo: Kryo, - input: KryoInput, - cls: Class[Array[(Any, Any)] => scala.collection.Map[Any, Any]]) + kryo: Kryo, + input: KryoInput, + cls: Class[Array[(Any, Any)] => scala.collection.Map[Any, Any]]) : Array[(Any, Any)] => scala.collection.Map[Any, Any] = { - val size = kryo.readObject(input, classOf[java.lang.Integer]).intValue + kryo.reference(FAKE_REFERENCE) + val size = input.readInt() val elems = new Array[(Any, Any)](size) - for (i <- 0 until size) - elems(i) = (kryo.readClassAndObject(input), kryo.readClassAndObject(input)) + for (i <- 0 until size) { + val k = kryo.readClassAndObject(input) + val v = kryo.readClassAndObject(input) + elems(i)=(k,v) + } buildMap(elems).asInstanceOf[Array[(Any, Any)] => scala.collection.Map[Any, Any]] } } diff --git a/core/src/main/scala/spark/Logging.scala b/core/src/main/scala/spark/Logging.scala index 7c1c1bb144..0fc8c31463 100644 --- a/core/src/main/scala/spark/Logging.scala +++ b/core/src/main/scala/spark/Logging.scala @@ -68,6 +68,10 @@ trait Logging { if (log.isErrorEnabled) log.error(msg, throwable) } + protected def isTraceEnabled(): Boolean = { + log.isTraceEnabled + } + // Method for ensuring that logging is initialized, to avoid having multiple // threads do it concurrently (as SLF4J initialization is not thread safe). protected def initLogging() { log } diff --git a/core/src/main/scala/spark/MapOutputTracker.scala b/core/src/main/scala/spark/MapOutputTracker.scala index 4735207585..fde597ffd1 100644 --- a/core/src/main/scala/spark/MapOutputTracker.scala +++ b/core/src/main/scala/spark/MapOutputTracker.scala @@ -1,7 +1,6 @@ package spark import java.io._ -import java.util.concurrent.ConcurrentHashMap import java.util.zip.{GZIPInputStream, GZIPOutputStream} import scala.collection.mutable.HashMap @@ -12,8 +11,7 @@ import akka.dispatch._ import akka.pattern.ask import akka.remote._ import akka.util.Duration -import akka.util.Timeout -import akka.util.duration._ + import spark.scheduler.MapStatus import spark.storage.BlockManagerId @@ -38,11 +36,14 @@ private[spark] class MapOutputTrackerActor(tracker: MapOutputTracker) extends Ac } } -private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolean) extends Logging { +private[spark] class MapOutputTracker extends Logging { - val timeout = 10.seconds + private val timeout = Duration.create(System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds") + + // Set to the MapOutputTrackerActor living on the driver + var trackerActor: ActorRef = _ - var mapStatuses = new TimeStampedHashMap[Int, Array[MapStatus]] + private var mapStatuses = new TimeStampedHashMap[Int, Array[MapStatus]] // Incremented every time a fetch fails so that client nodes know to clear // their cache of map output locations if this happens. @@ -51,19 +52,7 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea // Cache a serialized version of the output statuses for each shuffle to send them out faster var cacheGeneration = generation - val cachedSerializedStatuses = new TimeStampedHashMap[Int, Array[Byte]] - - val actorName: String = "MapOutputTracker" - var trackerActor: ActorRef = if (isDriver) { - val actor = actorSystem.actorOf(Props(new MapOutputTrackerActor(this)), name = actorName) - logInfo("Registered MapOutputTrackerActor actor") - actor - } else { - val ip = System.getProperty("spark.driver.host", "localhost") - val port = System.getProperty("spark.driver.port", "7077").toInt - val url = "akka://spark@%s:%s/user/%s".format(ip, port, actorName) - actorSystem.actorFor(url) - } + private val cachedSerializedStatuses = new TimeStampedHashMap[Int, Array[Byte]] val metadataCleaner = new MetadataCleaner("MapOutputTracker", this.cleanup) @@ -87,10 +76,9 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea } def registerShuffle(shuffleId: Int, numMaps: Int) { - if (mapStatuses.get(shuffleId) != None) { + if (mapStatuses.putIfAbsent(shuffleId, new Array[MapStatus](numMaps)).isDefined) { throw new IllegalArgumentException("Shuffle ID " + shuffleId + " registered twice") } - mapStatuses.put(shuffleId, new Array[MapStatus](numMaps)) } def registerMapOutput(shuffleId: Int, mapId: Int, status: MapStatus) { @@ -111,8 +99,9 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea } def unregisterMapOutput(shuffleId: Int, mapId: Int, bmAddress: BlockManagerId) { - var array = mapStatuses(shuffleId) - if (array != null) { + var arrayOpt = mapStatuses.get(shuffleId) + if (arrayOpt.isDefined && arrayOpt.get != null) { + var array = arrayOpt.get array.synchronized { if (array(mapId) != null && array(mapId).location == bmAddress) { array(mapId) = null @@ -125,13 +114,14 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea } // Remembers which map output locations are currently being fetched on a worker - val fetching = new HashSet[Int] + private val fetching = new HashSet[Int] // Called on possibly remote nodes to get the server URIs and output sizes for a given shuffle def getServerStatuses(shuffleId: Int, reduceId: Int): Array[(BlockManagerId, Long)] = { val statuses = mapStatuses.get(shuffleId).orNull if (statuses == null) { logInfo("Don't have map outputs for shuffle " + shuffleId + ", fetching them") + var fetchedStatuses: Array[MapStatus] = null fetching.synchronized { if (fetching.contains(shuffleId)) { // Someone else is fetching it; wait for them to be done @@ -142,31 +132,48 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea case e: InterruptedException => } } - return MapOutputTracker.convertMapStatuses(shuffleId, reduceId, mapStatuses(shuffleId)) - } else { + } + + // Either while we waited the fetch happened successfully, or + // someone fetched it in between the get and the fetching.synchronized. + fetchedStatuses = mapStatuses.get(shuffleId).orNull + if (fetchedStatuses == null) { + // We have to do the fetch, get others to wait for us. fetching += shuffleId } } - // We won the race to fetch the output locs; do so - logInfo("Doing the fetch; tracker actor = " + trackerActor) - val host = System.getProperty("spark.hostname", Utils.localHostName) - // This try-finally prevents hangs due to timeouts: - var fetchedStatuses: Array[MapStatus] = null - try { - val fetchedBytes = - askTracker(GetMapOutputStatuses(shuffleId, host)).asInstanceOf[Array[Byte]] - fetchedStatuses = deserializeStatuses(fetchedBytes) - logInfo("Got the output locations") - mapStatuses.put(shuffleId, fetchedStatuses) - } finally { - fetching.synchronized { - fetching -= shuffleId - fetching.notifyAll() + + if (fetchedStatuses == null) { + // We won the race to fetch the output locs; do so + logInfo("Doing the fetch; tracker actor = " + trackerActor) + val hostPort = Utils.localHostPort() + // This try-finally prevents hangs due to timeouts: + try { + val fetchedBytes = + askTracker(GetMapOutputStatuses(shuffleId, hostPort)).asInstanceOf[Array[Byte]] + fetchedStatuses = deserializeStatuses(fetchedBytes) + logInfo("Got the output locations") + mapStatuses.put(shuffleId, fetchedStatuses) + } finally { + fetching.synchronized { + fetching -= shuffleId + fetching.notifyAll() + } + } + } + if (fetchedStatuses != null) { + fetchedStatuses.synchronized { + return MapOutputTracker.convertMapStatuses(shuffleId, reduceId, fetchedStatuses) } } - return MapOutputTracker.convertMapStatuses(shuffleId, reduceId, fetchedStatuses) + else{ + throw new FetchFailedException(null, shuffleId, -1, reduceId, + new Exception("Missing all output locations for shuffle " + shuffleId)) + } } else { - return MapOutputTracker.convertMapStatuses(shuffleId, reduceId, statuses) + statuses.synchronized { + return MapOutputTracker.convertMapStatuses(shuffleId, reduceId, statuses) + } } } @@ -204,7 +211,8 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea generationLock.synchronized { if (newGen > generation) { logInfo("Updating generation to " + newGen + " and clearing cache") - mapStatuses = new TimeStampedHashMap[Int, Array[MapStatus]] + // mapStatuses = new TimeStampedHashMap[Int, Array[MapStatus]] + mapStatuses.clear() generation = newGen } } @@ -242,10 +250,13 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea // Serialize an array of map output locations into an efficient byte format so that we can send // it to reduce tasks. We do this by compressing the serialized bytes using GZIP. They will // generally be pretty compressible because many map outputs will be on the same hostname. - def serializeStatuses(statuses: Array[MapStatus]): Array[Byte] = { + private def serializeStatuses(statuses: Array[MapStatus]): Array[Byte] = { val out = new ByteArrayOutputStream val objOut = new ObjectOutputStream(new GZIPOutputStream(out)) - objOut.writeObject(statuses) + // Since statuses can be modified in parallel, sync on it + statuses.synchronized { + objOut.writeObject(statuses) + } objOut.close() out.toByteArray } @@ -253,7 +264,10 @@ private[spark] class MapOutputTracker(actorSystem: ActorSystem, isDriver: Boolea // Opposite of serializeStatuses. def deserializeStatuses(bytes: Array[Byte]): Array[MapStatus] = { val objIn = new ObjectInputStream(new GZIPInputStream(new ByteArrayInputStream(bytes))) - objIn.readObject().asInstanceOf[Array[MapStatus]] + objIn.readObject(). + // // drop all null's from status - not sure why they are occuring though. Causes NPE downstream in slave if present + // comment this out - nulls could be due to missing location ? + asInstanceOf[Array[MapStatus]] // .filter( _ != null ) } } @@ -263,14 +277,11 @@ private[spark] object MapOutputTracker { // Convert an array of MapStatuses to locations and sizes for a given reduce ID. If // any of the statuses is null (indicating a missing location due to a failed mapper), // throw a FetchFailedException. - def convertMapStatuses( + private def convertMapStatuses( shuffleId: Int, reduceId: Int, statuses: Array[MapStatus]): Array[(BlockManagerId, Long)] = { - if (statuses == null) { - throw new FetchFailedException(null, shuffleId, -1, reduceId, - new Exception("Missing all output locations for shuffle " + shuffleId)) - } + assert (statuses != null) statuses.map { status => if (status == null) { diff --git a/core/src/main/scala/spark/PairRDDFunctions.scala b/core/src/main/scala/spark/PairRDDFunctions.scala index 4319cbd892..7630fe7803 100644 --- a/core/src/main/scala/spark/PairRDDFunctions.scala +++ b/core/src/main/scala/spark/PairRDDFunctions.scala @@ -1,5 +1,6 @@ package spark +import java.nio.ByteBuffer import java.util.{Date, HashMap => JHashMap} import java.text.SimpleDateFormat @@ -10,6 +11,8 @@ import scala.collection.JavaConversions._ import org.apache.hadoop.conf.Configuration import org.apache.hadoop.fs.Path +import org.apache.hadoop.io.compress.CompressionCodec +import org.apache.hadoop.io.SequenceFile.CompressionType import org.apache.hadoop.mapred.FileOutputCommitter import org.apache.hadoop.mapred.FileOutputFormat import org.apache.hadoop.mapred.HadoopWriter @@ -17,12 +20,13 @@ import org.apache.hadoop.mapred.JobConf import org.apache.hadoop.mapred.OutputFormat import org.apache.hadoop.mapreduce.lib.output.{FileOutputFormat => NewFileOutputFormat} -import org.apache.hadoop.mapreduce.{OutputFormat => NewOutputFormat, RecordWriter => NewRecordWriter, Job => NewAPIHadoopJob, HadoopMapReduceUtil, TaskAttemptID, TaskAttemptContext} +import org.apache.hadoop.mapreduce.{OutputFormat => NewOutputFormat, RecordWriter => NewRecordWriter, Job => NewAPIHadoopJob, HadoopMapReduceUtil} import spark.partial.BoundedDouble import spark.partial.PartialResult import spark.rdd._ import spark.SparkContext._ +import spark.Partitioner._ /** * Extra functions available on RDDs of (key, value) pairs through an implicit conversion. @@ -51,7 +55,8 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( mergeValue: (C, V) => C, mergeCombiners: (C, C) => C, partitioner: Partitioner, - mapSideCombine: Boolean = true): RDD[(K, C)] = { + mapSideCombine: Boolean = true, + serializerClass: String = null): RDD[(K, C)] = { if (getKeyClass().isArray) { if (mapSideCombine) { throw new SparkException("Cannot use map-side combining with array keys.") @@ -60,19 +65,18 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( throw new SparkException("Default partitioner cannot partition array keys.") } } - val aggregator = - new Aggregator[K, V, C](createCombiner, mergeValue, mergeCombiners) + val aggregator = new Aggregator[K, V, C](createCombiner, mergeValue, mergeCombiners) if (self.partitioner == Some(partitioner)) { self.mapPartitions(aggregator.combineValuesByKey(_), true) } else if (mapSideCombine) { val mapSideCombined = self.mapPartitions(aggregator.combineValuesByKey(_), true) - val partitioned = new ShuffledRDD[K, C](mapSideCombined, partitioner) + val partitioned = new ShuffledRDD[K, C](mapSideCombined, partitioner, serializerClass) partitioned.mapPartitions(aggregator.combineCombinersByKey(_), true) } else { // Don't apply map-side combiner. // A sanity check to make sure mergeCombiners is not defined. assert(mergeCombiners == null) - val values = new ShuffledRDD[K, V](self, partitioner) + val values = new ShuffledRDD[K, V](self, partitioner, serializerClass) values.mapPartitions(aggregator.combineValuesByKey(_), true) } } @@ -88,6 +92,42 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( } /** + * Merge the values for each key using an associative function and a neutral "zero value" which may + * be added to the result an arbitrary number of times, and must not change the result (e.g., Nil for + * list concatenation, 0 for addition, or 1 for multiplication.). + */ + def foldByKey(zeroValue: V, partitioner: Partitioner)(func: (V, V) => V): RDD[(K, V)] = { + // Serialize the zero value to a byte array so that we can get a new clone of it on each key + val zeroBuffer = SparkEnv.get.closureSerializer.newInstance().serialize(zeroValue) + val zeroArray = new Array[Byte](zeroBuffer.limit) + zeroBuffer.get(zeroArray) + + // When deserializing, use a lazy val to create just one instance of the serializer per task + lazy val cachedSerializer = SparkEnv.get.closureSerializer.newInstance() + def createZero() = cachedSerializer.deserialize[V](ByteBuffer.wrap(zeroArray)) + + combineByKey[V]((v: V) => func(createZero(), v), func, func, partitioner) + } + + /** + * Merge the values for each key using an associative function and a neutral "zero value" which may + * be added to the result an arbitrary number of times, and must not change the result (e.g., Nil for + * list concatenation, 0 for addition, or 1 for multiplication.). + */ + def foldByKey(zeroValue: V, numPartitions: Int)(func: (V, V) => V): RDD[(K, V)] = { + foldByKey(zeroValue, new HashPartitioner(numPartitions))(func) + } + + /** + * Merge the values for each key using an associative function and a neutral "zero value" which may + * be added to the result an arbitrary number of times, and must not change the result (e.g., Nil for + * list concatenation, 0 for addition, or 1 for multiplication.). + */ + def foldByKey(zeroValue: V)(func: (V, V) => V): RDD[(K, V)] = { + foldByKey(zeroValue, defaultPartitioner(self))(func) + } + + /** * Merge the values for each key using an associative reduce function. This will also perform * the merging locally on each mapper before sending results to a reducer, similarly to a * "combiner" in MapReduce. @@ -156,11 +196,13 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( * partitioning of the resulting key-value pair RDD by passing a Partitioner. */ def groupByKey(partitioner: Partitioner): RDD[(K, Seq[V])] = { + // groupByKey shouldn't use map side combine because map side combine does not + // reduce the amount of data shuffled and requires all map side data be inserted + // into a hash table, leading to more objects in the old gen. def createCombiner(v: V) = ArrayBuffer(v) def mergeValue(buf: ArrayBuffer[V], v: V) = buf += v - def mergeCombiners(b1: ArrayBuffer[V], b2: ArrayBuffer[V]) = b1 ++= b2 val bufs = combineByKey[ArrayBuffer[V]]( - createCombiner _, mergeValue _, mergeCombiners _, partitioner) + createCombiner _, mergeValue _, null, partitioner, mapSideCombine=false) bufs.asInstanceOf[RDD[(K, Seq[V])]] } @@ -248,8 +290,8 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( } /** - * Simplified version of combineByKey that hash-partitions the resulting RDD using the default - * parallelism level. + * Simplified version of combineByKey that hash-partitions the resulting RDD using the + * existing partitioner/parallelism level. */ def combineByKey[C](createCombiner: V => C, mergeValue: (C, V) => C, mergeCombiners: (C, C) => C) : RDD[(K, C)] = { @@ -259,7 +301,8 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( /** * Merge the values for each key using an associative reduce function. This will also perform * the merging locally on each mapper before sending results to a reducer, similarly to a - * "combiner" in MapReduce. Output will be hash-partitioned with the default parallelism level. + * "combiner" in MapReduce. Output will be hash-partitioned with the existing partitioner/ + * parallelism level. */ def reduceByKey(func: (V, V) => V): RDD[(K, V)] = { reduceByKey(defaultPartitioner(self), func) @@ -267,7 +310,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( /** * Group the values for each key in the RDD into a single sequence. Hash-partitions the - * resulting RDD with the default parallelism level. + * resulting RDD with the existing partitioner/parallelism level. */ def groupByKey(): RDD[(K, Seq[V])] = { groupByKey(defaultPartitioner(self)) @@ -295,7 +338,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( * Perform a left outer join of `this` and `other`. For each element (k, v) in `this`, the * resulting RDD will either contain all pairs (k, (v, Some(w))) for w in `other`, or the * pair (k, (v, None)) if no elements in `other` have key k. Hash-partitions the output - * using the default level of parallelism. + * using the existing partitioner/parallelism level. */ def leftOuterJoin[W](other: RDD[(K, W)]): RDD[(K, (V, Option[W]))] = { leftOuterJoin(other, defaultPartitioner(self, other)) @@ -315,7 +358,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( * Perform a right outer join of `this` and `other`. For each element (k, w) in `other`, the * resulting RDD will either contain all pairs (k, (Some(v), w)) for v in `this`, or the * pair (k, (None, w)) if no elements in `this` have key k. Hash-partitions the resulting - * RDD using the default parallelism level. + * RDD using the existing partitioner/parallelism level. */ def rightOuterJoin[W](other: RDD[(K, W)]): RDD[(K, (Option[V], W))] = { rightOuterJoin(other, defaultPartitioner(self, other)) @@ -439,15 +482,21 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( } /** - * Choose a partitioner to use for a cogroup-like operation between a number of RDDs. If any of - * the RDDs already has a partitioner, choose that one, otherwise use a default HashPartitioner. + * Return an RDD with the pairs from `this` whose keys are not in `other`. + * + * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting + * RDD will be <= us. */ - def defaultPartitioner(rdds: RDD[_]*): Partitioner = { - for (r <- rdds if r.partitioner != None) { - return r.partitioner.get - } - return new HashPartitioner(self.context.defaultParallelism) - } + def subtractByKey[W: ClassManifest](other: RDD[(K, W)]): RDD[(K, V)] = + subtractByKey(other, self.partitioner.getOrElse(new HashPartitioner(self.partitions.size))) + + /** Return an RDD with the pairs from `this` whose keys are not in `other`. */ + def subtractByKey[W: ClassManifest](other: RDD[(K, W)], numPartitions: Int): RDD[(K, V)] = + subtractByKey(other, new HashPartitioner(numPartitions)) + + /** Return an RDD with the pairs from `this` whose keys are not in `other`. */ + def subtractByKey[W: ClassManifest](other: RDD[(K, W)], p: Partitioner): RDD[(K, V)] = + new SubtractedRDD[K, V, W](self, other, p) /** * Return the list of values in the RDD for key `key`. This operation is done efficiently if the @@ -480,6 +529,16 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( } /** + * Output the RDD to any Hadoop-supported file system, using a Hadoop `OutputFormat` class + * supporting the key and value types K and V in this RDD. Compress the result with the + * supplied codec. + */ + def saveAsHadoopFile[F <: OutputFormat[K, V]]( + path: String, codec: Class[_ <: CompressionCodec]) (implicit fm: ClassManifest[F]) { + saveAsHadoopFile(path, getKeyClass, getValueClass, fm.erasure.asInstanceOf[Class[F]], codec) + } + + /** * Output the RDD to any Hadoop-supported file system, using a new Hadoop API `OutputFormat` * (mapreduce.OutputFormat) object supporting the key and value types K and V in this RDD. */ @@ -510,8 +569,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( // around by taking a mod. We expect that no task will be attempted 2 billion times. val attemptNumber = (context.attemptId % Int.MaxValue).toInt /* "reduce task" <split #> <attempt # = spark task #> */ - val attemptId = new TaskAttemptID(jobtrackerID, - stageId, false, context.splitId, attemptNumber) + val attemptId = newTaskAttemptID(jobtrackerID, stageId, false, context.splitId, attemptNumber) val hadoopContext = newTaskAttemptContext(wrappedConf.value, attemptId) val format = outputFormatClass.newInstance val committer = format.getOutputCommitter(hadoopContext) @@ -530,16 +588,31 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( * however we're only going to use this local OutputCommitter for * setupJob/commitJob, so we just use a dummy "map" task. */ - val jobAttemptId = new TaskAttemptID(jobtrackerID, stageId, true, 0, 0) + val jobAttemptId = newTaskAttemptID(jobtrackerID, stageId, true, 0, 0) val jobTaskContext = newTaskAttemptContext(wrappedConf.value, jobAttemptId) val jobCommitter = jobFormat.getOutputCommitter(jobTaskContext) jobCommitter.setupJob(jobTaskContext) val count = self.context.runJob(self, writeShard _).sum + jobCommitter.commitJob(jobTaskContext) jobCommitter.cleanupJob(jobTaskContext) } /** * Output the RDD to any Hadoop-supported file system, using a Hadoop `OutputFormat` class + * supporting the key and value types K and V in this RDD. Compress with the supplied codec. + */ + def saveAsHadoopFile( + path: String, + keyClass: Class[_], + valueClass: Class[_], + outputFormatClass: Class[_ <: OutputFormat[_, _]], + codec: Class[_ <: CompressionCodec]) { + saveAsHadoopFile(path, keyClass, valueClass, outputFormatClass, + new JobConf(self.context.hadoopConfiguration), Some(codec)) + } + + /** + * Output the RDD to any Hadoop-supported file system, using a Hadoop `OutputFormat` class * supporting the key and value types K and V in this RDD. */ def saveAsHadoopFile( @@ -547,11 +620,19 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( keyClass: Class[_], valueClass: Class[_], outputFormatClass: Class[_ <: OutputFormat[_, _]], - conf: JobConf = new JobConf(self.context.hadoopConfiguration)) { + conf: JobConf = new JobConf(self.context.hadoopConfiguration), + codec: Option[Class[_ <: CompressionCodec]] = None) { conf.setOutputKeyClass(keyClass) conf.setOutputValueClass(valueClass) // conf.setOutputFormat(outputFormatClass) // Doesn't work in Scala 2.9 due to what may be a generics bug conf.set("mapred.output.format.class", outputFormatClass.getName) + for (c <- codec) { + conf.setCompressMapOutput(true) + conf.set("mapred.output.compress", "true") + conf.setMapOutputCompressorClass(c) + conf.set("mapred.output.compression.codec", c.getCanonicalName) + conf.set("mapred.output.compression.type", CompressionType.BLOCK.toString) + } conf.setOutputCommitter(classOf[FileOutputCommitter]) FileOutputFormat.setOutputPath(conf, HadoopWriter.createPathFromString(path, conf)) saveAsHadoopDataset(conf) @@ -602,6 +683,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( } self.context.runJob(self, writeToFile _) + writer.commitJob() writer.cleanup() } @@ -609,7 +691,7 @@ class PairRDDFunctions[K: ClassManifest, V: ClassManifest]( * Return an RDD with the keys of each tuple. */ def keys: RDD[K] = self.map(_._1) - + /** * Return an RDD with the values of each tuple. */ diff --git a/core/src/main/scala/spark/Partitioner.scala b/core/src/main/scala/spark/Partitioner.scala index 9d5b966e1e..6f8cd17c88 100644 --- a/core/src/main/scala/spark/Partitioner.scala +++ b/core/src/main/scala/spark/Partitioner.scala @@ -9,6 +9,35 @@ abstract class Partitioner extends Serializable { def getPartition(key: Any): Int } +object Partitioner { + /** + * Choose a partitioner to use for a cogroup-like operation between a number of RDDs. + * + * If any of the RDDs already has a partitioner, choose that one. + * + * Otherwise, we use a default HashPartitioner. For the number of partitions, if + * spark.default.parallelism is set, then we'll use the value from SparkContext + * defaultParallelism, otherwise we'll use the max number of upstream partitions. + * + * Unless spark.default.parallelism is set, He number of partitions will be the + * same as the number of partitions in the largest upstream RDD, as this should + * be least likely to cause out-of-memory errors. + * + * We use two method parameters (rdd, others) to enforce callers passing at least 1 RDD. + */ + def defaultPartitioner(rdd: RDD[_], others: RDD[_]*): Partitioner = { + val bySize = (Seq(rdd) ++ others).sortBy(_.partitions.size).reverse + for (r <- bySize if r.partitioner != None) { + return r.partitioner.get + } + if (System.getProperty("spark.default.parallelism") != null) { + return new HashPartitioner(rdd.context.defaultParallelism) + } else { + return new HashPartitioner(bySize.head.partitions.size) + } + } +} + /** * A [[spark.Partitioner]] that implements hash-based partitioning using Java's `Object.hashCode`. * diff --git a/core/src/main/scala/spark/RDD.scala b/core/src/main/scala/spark/RDD.scala index da82dfd10f..f336c2ea1e 100644 --- a/core/src/main/scala/spark/RDD.scala +++ b/core/src/main/scala/spark/RDD.scala @@ -1,21 +1,21 @@ package spark -import java.net.URL -import java.util.{Date, Random} -import java.util.{HashMap => JHashMap} +import java.util.Random import scala.collection.Map import scala.collection.JavaConversions.mapAsScalaMap import scala.collection.mutable.ArrayBuffer -import scala.collection.mutable.HashMap import org.apache.hadoop.io.BytesWritable +import org.apache.hadoop.io.compress.CompressionCodec import org.apache.hadoop.io.NullWritable import org.apache.hadoop.io.Text import org.apache.hadoop.mapred.TextOutputFormat import it.unimi.dsi.fastutil.objects.{Object2LongOpenHashMap => OLMap} +import spark.broadcast.Broadcast +import spark.Partitioner._ import spark.partial.BoundedDouble import spark.partial.CountEvaluator import spark.partial.GroupedCountEvaluator @@ -30,9 +30,14 @@ import spark.rdd.MapPartitionsRDD import spark.rdd.MapPartitionsWithIndexRDD import spark.rdd.PipedRDD import spark.rdd.SampledRDD +import spark.rdd.ShuffledRDD import spark.rdd.UnionRDD import spark.rdd.ZippedRDD +import spark.rdd.ZippedPartitionsRDD2 +import spark.rdd.ZippedPartitionsRDD3 +import spark.rdd.ZippedPartitionsRDD4 import spark.storage.StorageLevel +import spark.util.BoundedPriorityQueue import SparkContext._ @@ -101,7 +106,7 @@ abstract class RDD[T: ClassManifest]( // ======================================================================= /** A unique ID for this RDD (within its SparkContext). */ - val id = sc.newRddId() + val id: Int = sc.newRddId() /** A friendly name for this RDD */ var name: String = null @@ -112,9 +117,18 @@ abstract class RDD[T: ClassManifest]( this } + /** User-defined generator of this RDD*/ + var generator = Utils.getCallSiteInfo.firstUserClass + + /** Reset generator*/ + def setGenerator(_generator: String) = { + generator = _generator + } + /** * Set this RDD's storage level to persist its values across operations after the first time - * it is computed. Can only be called once on each RDD. + * it is computed. This can only be used to assign a new storage level if the RDD does not + * have a storage level set yet.. */ def persist(newLevel: StorageLevel): RDD[T] = { // TODO: Handle changes of StorageLevel @@ -134,6 +148,20 @@ abstract class RDD[T: ClassManifest]( /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */ def cache(): RDD[T] = persist() + /** + * Mark the RDD as non-persistent, and remove all blocks for it from memory and disk. + * + * @param blocking Whether to block until all blocks are deleted. + * @return This RDD. + */ + def unpersist(blocking: Boolean = true): RDD[T] = { + logInfo("Removing RDD " + id + " from persistence list") + sc.env.blockManager.master.removeRdd(id, blocking) + sc.persistentRdds.remove(id) + storageLevel = StorageLevel.NONE + this + } + /** Get the RDD's current storage level, or StorageLevel.NONE if none is set. */ def getStorageLevel = storageLevel @@ -235,7 +263,14 @@ abstract class RDD[T: ClassManifest]( /** * Return a new RDD that is reduced into `numPartitions` partitions. */ - def coalesce(numPartitions: Int): RDD[T] = new CoalescedRDD(this, numPartitions) + def coalesce(numPartitions: Int, shuffle: Boolean = false): RDD[T] = { + if (shuffle) { + // include a shuffle step so that our upstream tasks are still distributed + new CoalescedRDD(new ShuffledRDD(map(x => (x, null)), new HashPartitioner(numPartitions)), numPartitions).keys + } else { + new CoalescedRDD(this, numPartitions) + } + } /** * Return a sampled subset of this RDD. @@ -246,8 +281,8 @@ abstract class RDD[T: ClassManifest]( def takeSample(withReplacement: Boolean, num: Int, seed: Int): Array[T] = { var fraction = 0.0 var total = 0 - var multiplier = 3.0 - var initialCount = count() + val multiplier = 3.0 + val initialCount = count() var maxSelected = 0 if (initialCount > Integer.MAX_VALUE - 1) { @@ -300,18 +335,25 @@ abstract class RDD[T: ClassManifest]( def cartesian[U: ClassManifest](other: RDD[U]): RDD[(T, U)] = new CartesianRDD(sc, this, other) /** + * Return an RDD of grouped items. + */ + def groupBy[K: ClassManifest](f: T => K): RDD[(K, Seq[T])] = + groupBy[K](f, defaultPartitioner(this)) + + /** * Return an RDD of grouped elements. Each group consists of a key and a sequence of elements * mapping to that key. */ - def groupBy[K: ClassManifest](f: T => K, numPartitions: Int): RDD[(K, Seq[T])] = { - val cleanF = sc.clean(f) - this.map(t => (cleanF(t), t)).groupByKey(numPartitions) - } + def groupBy[K: ClassManifest](f: T => K, numPartitions: Int): RDD[(K, Seq[T])] = + groupBy(f, new HashPartitioner(numPartitions)) /** * Return an RDD of grouped items. */ - def groupBy[K: ClassManifest](f: T => K): RDD[(K, Seq[T])] = groupBy[K](f, sc.defaultParallelism) + def groupBy[K: ClassManifest](f: T => K, p: Partitioner): RDD[(K, Seq[T])] = { + val cleanF = sc.clean(f) + this.map(t => (cleanF(t), t)).groupByKey(p) + } /** * Return an RDD created by piping elements to a forked external process. @@ -321,13 +363,36 @@ abstract class RDD[T: ClassManifest]( /** * Return an RDD created by piping elements to a forked external process. */ - def pipe(command: Seq[String]): RDD[String] = new PipedRDD(this, command) + def pipe(command: String, env: Map[String, String]): RDD[String] = + new PipedRDD(this, command, env) + /** * Return an RDD created by piping elements to a forked external process. - */ - def pipe(command: Seq[String], env: Map[String, String]): RDD[String] = - new PipedRDD(this, command, env) + * The print behavior can be customized by providing two functions. + * + * @param command command to run in forked process. + * @param env environment variables to set. + * @param printPipeContext Before piping elements, this function is called as an oppotunity + * to pipe context data. Print line function (like out.println) will be + * passed as printPipeContext's parameter. + * @param printRDDElement Use this function to customize how to pipe elements. This function + * will be called with each RDD element as the 1st parameter, and the + * print line function (like out.println()) as the 2nd parameter. + * An example of pipe the RDD data of groupBy() in a streaming way, + * instead of constructing a huge String to concat all the elements: + * def printRDDElement(record:(String, Seq[String]), f:String=>Unit) = + * for (e <- record._2){f(e)} + * @return the result RDD + */ + def pipe( + command: Seq[String], + env: Map[String, String] = Map(), + printPipeContext: (String => Unit) => Unit = null, + printRDDElement: (T, String => Unit) => Unit = null): RDD[String] = + new PipedRDD(this, command, env, + if (printPipeContext ne null) sc.clean(printPipeContext) else null, + if (printRDDElement ne null) sc.clean(printRDDElement) else null) /** * Return a new RDD by applying a function to each partition of this RDD. @@ -349,13 +414,69 @@ abstract class RDD[T: ClassManifest]( * Return a new RDD by applying a function to each partition of this RDD, while tracking the index * of the original partition. */ - @deprecated("use mapPartitionsWithIndex") + @deprecated("use mapPartitionsWithIndex", "0.7.0") def mapPartitionsWithSplit[U: ClassManifest]( f: (Int, Iterator[T]) => Iterator[U], preservesPartitioning: Boolean = false): RDD[U] = new MapPartitionsWithIndexRDD(this, sc.clean(f), preservesPartitioning) /** + * Maps f over this RDD, where f takes an additional parameter of type A. This + * additional parameter is produced by constructA, which is called in each + * partition with the index of that partition. + */ + def mapWith[A: ClassManifest, U: ClassManifest](constructA: Int => A, preservesPartitioning: Boolean = false) + (f:(T, A) => U): RDD[U] = { + def iterF(index: Int, iter: Iterator[T]): Iterator[U] = { + val a = constructA(index) + iter.map(t => f(t, a)) + } + new MapPartitionsWithIndexRDD(this, sc.clean(iterF _), preservesPartitioning) + } + + /** + * FlatMaps f over this RDD, where f takes an additional parameter of type A. This + * additional parameter is produced by constructA, which is called in each + * partition with the index of that partition. + */ + def flatMapWith[A: ClassManifest, U: ClassManifest](constructA: Int => A, preservesPartitioning: Boolean = false) + (f:(T, A) => Seq[U]): RDD[U] = { + def iterF(index: Int, iter: Iterator[T]): Iterator[U] = { + val a = constructA(index) + iter.flatMap(t => f(t, a)) + } + new MapPartitionsWithIndexRDD(this, sc.clean(iterF _), preservesPartitioning) + } + + /** + * Applies f to each element of this RDD, where f takes an additional parameter of type A. + * This additional parameter is produced by constructA, which is called in each + * partition with the index of that partition. + */ + def foreachWith[A: ClassManifest](constructA: Int => A) + (f:(T, A) => Unit) { + def iterF(index: Int, iter: Iterator[T]): Iterator[T] = { + val a = constructA(index) + iter.map(t => {f(t, a); t}) + } + (new MapPartitionsWithIndexRDD(this, sc.clean(iterF _), true)).foreach(_ => {}) + } + + /** + * Filters this RDD with p, where p takes an additional parameter of type A. This + * additional parameter is produced by constructA, which is called in each + * partition with the index of that partition. + */ + def filterWith[A: ClassManifest](constructA: Int => A) + (p:(T, A) => Boolean): RDD[T] = { + def iterF(index: Int, iter: Iterator[T]): Iterator[T] = { + val a = constructA(index) + iter.filter(t => p(t, a)) + } + new MapPartitionsWithIndexRDD(this, sc.clean(iterF _), true) + } + + /** * Zips this RDD with another one, returning key-value pairs with the first element in each RDD, * second element in each RDD, etc. Assumes that the two RDDs have the *same number of * partitions* and the *same number of elements in each partition* (e.g. one was made through @@ -363,6 +484,31 @@ abstract class RDD[T: ClassManifest]( */ def zip[U: ClassManifest](other: RDD[U]): RDD[(T, U)] = new ZippedRDD(sc, this, other) + /** + * Zip this RDD's partitions with one (or more) RDD(s) and return a new RDD by + * applying a function to the zipped partitions. Assumes that all the RDDs have the + * *same number of partitions*, but does *not* require them to have the same number + * of elements in each partition. + */ + def zipPartitions[B: ClassManifest, V: ClassManifest]( + f: (Iterator[T], Iterator[B]) => Iterator[V], + rdd2: RDD[B]): RDD[V] = + new ZippedPartitionsRDD2(sc, sc.clean(f), this, rdd2) + + def zipPartitions[B: ClassManifest, C: ClassManifest, V: ClassManifest]( + f: (Iterator[T], Iterator[B], Iterator[C]) => Iterator[V], + rdd2: RDD[B], + rdd3: RDD[C]): RDD[V] = + new ZippedPartitionsRDD3(sc, sc.clean(f), this, rdd2, rdd3) + + def zipPartitions[B: ClassManifest, C: ClassManifest, D: ClassManifest, V: ClassManifest]( + f: (Iterator[T], Iterator[B], Iterator[C], Iterator[D]) => Iterator[V], + rdd2: RDD[B], + rdd3: RDD[C], + rdd4: RDD[D]): RDD[V] = + new ZippedPartitionsRDD4(sc, sc.clean(f), this, rdd2, rdd3, rdd4) + + // Actions (launch a job to return a value to the user program) /** @@ -374,6 +520,14 @@ abstract class RDD[T: ClassManifest]( } /** + * Applies a function f to each partition of this RDD. + */ + def foreachPartition(f: Iterator[T] => Unit) { + val cleanF = sc.clean(f) + sc.runJob(this, (iter: Iterator[T]) => cleanF(iter)) + } + + /** * Return an array that contains all of the elements in this RDD. */ def collect(): Array[T] = { @@ -394,6 +548,42 @@ abstract class RDD[T: ClassManifest]( } /** + * Return an RDD with the elements from `this` that are not in `other`. + * + * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting + * RDD will be <= us. + */ + def subtract(other: RDD[T]): RDD[T] = + subtract(other, partitioner.getOrElse(new HashPartitioner(partitions.size))) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: RDD[T], numPartitions: Int): RDD[T] = + subtract(other, new HashPartitioner(numPartitions)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: RDD[T], p: Partitioner): RDD[T] = { + if (partitioner == Some(p)) { + // Our partitioner knows how to handle T (which, since we have a partitioner, is + // really (K, V)) so make a new Partitioner that will de-tuple our fake tuples + val p2 = new Partitioner() { + override def numPartitions = p.numPartitions + override def getPartition(k: Any) = p.getPartition(k.asInstanceOf[(Any, _)]._1) + } + // Unfortunately, since we're making a new p2, we'll get ShuffleDependencies + // anyway, and when calling .keys, will not have a partitioner set, even though + // the SubtractedRDD will, thanks to p2's de-tupled partitioning, already be + // partitioned by the right/real keys (e.g. p). + this.map(x => (x, null)).subtractByKey(other.map((_, null)), p2).keys + } else { + this.map(x => (x, null)).subtractByKey(other.map((_, null)), p).keys + } + } + + /** * Reduces the elements of this RDD using the specified commutative and associative binary operator. */ def reduce(f: (T, T) => T): T = { @@ -567,6 +757,24 @@ abstract class RDD[T: ClassManifest]( } /** + * Returns the top K elements from this RDD as defined by + * the specified implicit Ordering[T]. + * @param num the number of top elements to return + * @param ord the implicit ordering for T + * @return an array of top elements + */ + def top(num: Int)(implicit ord: Ordering[T]): Array[T] = { + mapPartitions { items => + val queue = new BoundedPriorityQueue[T](num) + queue ++= items + Iterator.single(queue) + }.reduce { (queue1, queue2) => + queue1 ++= queue2 + queue1 + }.toArray + } + + /** * Save this RDD as a text file, using string representations of elements. */ def saveAsTextFile(path: String) { @@ -575,6 +783,14 @@ abstract class RDD[T: ClassManifest]( } /** + * Save this RDD as a compressed text file, using string representations of elements. + */ + def saveAsTextFile(path: String, codec: Class[_ <: CompressionCodec]) { + this.map(x => (NullWritable.get(), new Text(x.toString))) + .saveAsHadoopFile[TextOutputFormat[NullWritable, Text]](path, codec) + } + + /** * Save this RDD as a SequenceFile of serialized objects. */ def saveAsObjectFile(path: String) { @@ -632,7 +848,7 @@ abstract class RDD[T: ClassManifest]( private var storageLevel: StorageLevel = StorageLevel.NONE /** Record user function generating this RDD. */ - private[spark] val origin = Utils.getSparkCallSite + private[spark] val origin = Utils.formatSparkCallSite private[spark] def elementClassManifest: ClassManifest[T] = classManifest[T] diff --git a/core/src/main/scala/spark/RDDCheckpointData.scala b/core/src/main/scala/spark/RDDCheckpointData.scala index d00092e984..57e0405fb4 100644 --- a/core/src/main/scala/spark/RDDCheckpointData.scala +++ b/core/src/main/scala/spark/RDDCheckpointData.scala @@ -1,6 +1,7 @@ package spark import org.apache.hadoop.fs.Path +import org.apache.hadoop.conf.Configuration import rdd.{CheckpointRDD, CoalescedRDD} import scheduler.{ResultTask, ShuffleMapTask} @@ -62,14 +63,20 @@ private[spark] class RDDCheckpointData[T: ClassManifest](rdd: RDD[T]) } } + // Create the output path for the checkpoint + val path = new Path(rdd.context.checkpointDir.get, "rdd-" + rdd.id) + val fs = path.getFileSystem(new Configuration()) + if (!fs.mkdirs(path)) { + throw new SparkException("Failed to create checkpoint path " + path) + } + // Save to file, and reload it as an RDD - val path = new Path(rdd.context.checkpointDir.get, "rdd-" + rdd.id).toString - rdd.context.runJob(rdd, CheckpointRDD.writeToFile(path) _) - val newRDD = new CheckpointRDD[T](rdd.context, path) + rdd.context.runJob(rdd, CheckpointRDD.writeToFile(path.toString) _) + val newRDD = new CheckpointRDD[T](rdd.context, path.toString) // Change the dependencies and partitions of the RDD RDDCheckpointData.synchronized { - cpFile = Some(path) + cpFile = Some(path.toString) cpRDD = Some(newRDD) rdd.markCheckpointed(newRDD) // Update the RDD's dependencies and partitions cpState = Checkpointed diff --git a/core/src/main/scala/spark/SequenceFileRDDFunctions.scala b/core/src/main/scala/spark/SequenceFileRDDFunctions.scala index 6b4a11d6d3..2911f9036e 100644 --- a/core/src/main/scala/spark/SequenceFileRDDFunctions.scala +++ b/core/src/main/scala/spark/SequenceFileRDDFunctions.scala @@ -18,6 +18,7 @@ import org.apache.hadoop.mapred.TextOutputFormat import org.apache.hadoop.mapred.SequenceFileOutputFormat import org.apache.hadoop.mapred.OutputCommitter import org.apache.hadoop.mapred.FileOutputCommitter +import org.apache.hadoop.io.compress.CompressionCodec import org.apache.hadoop.io.Writable import org.apache.hadoop.io.NullWritable import org.apache.hadoop.io.BytesWritable @@ -36,17 +37,17 @@ class SequenceFileRDDFunctions[K <% Writable: ClassManifest, V <% Writable : Cla self: RDD[(K, V)]) extends Logging with Serializable { - + private def getWritableClass[T <% Writable: ClassManifest](): Class[_ <: Writable] = { val c = { - if (classOf[Writable].isAssignableFrom(classManifest[T].erasure)) { + if (classOf[Writable].isAssignableFrom(classManifest[T].erasure)) { classManifest[T].erasure } else { // We get the type of the Writable class by looking at the apply method which converts // from T to Writable. Since we have two apply methods we filter out the one which - // is of the form "java.lang.Object apply(java.lang.Object)" + // is not of the form "java.lang.Object apply(java.lang.Object)" implicitly[T => Writable].getClass.getDeclaredMethods().filter( - m => m.getReturnType().toString != "java.lang.Object" && + m => m.getReturnType().toString != "class java.lang.Object" && m.getName() == "apply")(0).getReturnType } @@ -62,24 +63,28 @@ class SequenceFileRDDFunctions[K <% Writable: ClassManifest, V <% Writable : Cla * byte arrays to BytesWritable, and Strings to Text. The `path` can be on any Hadoop-supported * file system. */ - def saveAsSequenceFile(path: String) { + def saveAsSequenceFile(path: String, codec: Option[Class[_ <: CompressionCodec]] = None) { def anyToWritable[U <% Writable](u: U): Writable = u val keyClass = getWritableClass[K] val valueClass = getWritableClass[V] val convertKey = !classOf[Writable].isAssignableFrom(self.getKeyClass) val convertValue = !classOf[Writable].isAssignableFrom(self.getValueClass) - - logInfo("Saving as sequence file of type (" + keyClass.getSimpleName + "," + valueClass.getSimpleName + ")" ) + + logInfo("Saving as sequence file of type (" + keyClass.getSimpleName + "," + valueClass.getSimpleName + ")" ) val format = classOf[SequenceFileOutputFormat[Writable, Writable]] + val jobConf = new JobConf(self.context.hadoopConfiguration) if (!convertKey && !convertValue) { - self.saveAsHadoopFile(path, keyClass, valueClass, format) + self.saveAsHadoopFile(path, keyClass, valueClass, format, jobConf, codec) } else if (!convertKey && convertValue) { - self.map(x => (x._1,anyToWritable(x._2))).saveAsHadoopFile(path, keyClass, valueClass, format) + self.map(x => (x._1,anyToWritable(x._2))).saveAsHadoopFile( + path, keyClass, valueClass, format, jobConf, codec) } else if (convertKey && !convertValue) { - self.map(x => (anyToWritable(x._1),x._2)).saveAsHadoopFile(path, keyClass, valueClass, format) + self.map(x => (anyToWritable(x._1),x._2)).saveAsHadoopFile( + path, keyClass, valueClass, format, jobConf, codec) } else if (convertKey && convertValue) { - self.map(x => (anyToWritable(x._1),anyToWritable(x._2))).saveAsHadoopFile(path, keyClass, valueClass, format) - } + self.map(x => (anyToWritable(x._1),anyToWritable(x._2))).saveAsHadoopFile( + path, keyClass, valueClass, format, jobConf, codec) + } } } diff --git a/core/src/main/scala/spark/ShuffleFetcher.scala b/core/src/main/scala/spark/ShuffleFetcher.scala index d9a94d4021..9513a00126 100644 --- a/core/src/main/scala/spark/ShuffleFetcher.scala +++ b/core/src/main/scala/spark/ShuffleFetcher.scala @@ -1,11 +1,16 @@ package spark +import spark.executor.TaskMetrics +import spark.serializer.Serializer + + private[spark] abstract class ShuffleFetcher { /** * Fetch the shuffle outputs for a given ShuffleDependency. * @return An iterator over the elements of the fetched shuffle outputs. */ - def fetch[K, V](shuffleId: Int, reduceId: Int) : Iterator[(K, V)] + def fetch[K, V](shuffleId: Int, reduceId: Int, metrics: TaskMetrics, + serializer: Serializer = SparkEnv.get.serializerManager.default): Iterator[(K,V)] /** Stop the fetcher */ def stop() {} diff --git a/core/src/main/scala/spark/SizeEstimator.scala b/core/src/main/scala/spark/SizeEstimator.scala index d4e1157250..f8a4c4e489 100644 --- a/core/src/main/scala/spark/SizeEstimator.scala +++ b/core/src/main/scala/spark/SizeEstimator.scala @@ -198,7 +198,7 @@ private[spark] object SizeEstimator extends Logging { val elem = JArray.get(array, index) size += SizeEstimator.estimate(elem, state.visited) } - state.size += ((length / 100.0) * size).toLong + state.size += ((length / (ARRAY_SAMPLE_SIZE * 1.0)) * size).toLong } } } diff --git a/core/src/main/scala/spark/SparkContext.scala b/core/src/main/scala/spark/SparkContext.scala index d39767c3b3..70a9d7698c 100644 --- a/core/src/main/scala/spark/SparkContext.scala +++ b/core/src/main/scala/spark/SparkContext.scala @@ -1,52 +1,53 @@ package spark import java.io._ +import java.net.URI +import java.util.Properties import java.util.concurrent.ConcurrentHashMap import java.util.concurrent.atomic.AtomicInteger -import java.net.{URI, URLClassLoader} -import java.lang.ref.WeakReference +import scala.collection.JavaConversions._ import scala.collection.Map import scala.collection.generic.Growable -import scala.collection.mutable.{ArrayBuffer, HashMap} +import scala.collection.mutable.HashMap import scala.collection.JavaConversions._ +import scala.util.DynamicVariable +import scala.collection.mutable.{ConcurrentMap, HashMap} -import akka.actor.Actor import akka.actor.Actor._ -import org.apache.hadoop.fs.{FileUtil, Path} + import org.apache.hadoop.conf.Configuration -import org.apache.hadoop.mapred.InputFormat -import org.apache.hadoop.mapred.SequenceFileInputFormat -import org.apache.hadoop.io.Writable -import org.apache.hadoop.io.IntWritable -import org.apache.hadoop.io.LongWritable -import org.apache.hadoop.io.FloatWritable -import org.apache.hadoop.io.DoubleWritable +import org.apache.hadoop.fs.Path +import org.apache.hadoop.io.ArrayWritable import org.apache.hadoop.io.BooleanWritable import org.apache.hadoop.io.BytesWritable -import org.apache.hadoop.io.ArrayWritable +import org.apache.hadoop.io.DoubleWritable +import org.apache.hadoop.io.FloatWritable +import org.apache.hadoop.io.IntWritable +import org.apache.hadoop.io.LongWritable import org.apache.hadoop.io.NullWritable import org.apache.hadoop.io.Text +import org.apache.hadoop.io.Writable import org.apache.hadoop.mapred.FileInputFormat +import org.apache.hadoop.mapred.InputFormat import org.apache.hadoop.mapred.JobConf +import org.apache.hadoop.mapred.SequenceFileInputFormat import org.apache.hadoop.mapred.TextInputFormat import org.apache.hadoop.mapreduce.{InputFormat => NewInputFormat} -import org.apache.hadoop.mapreduce.lib.input.{FileInputFormat => NewFileInputFormat} import org.apache.hadoop.mapreduce.{Job => NewHadoopJob} -import org.apache.mesos.{Scheduler, MesosNativeLibrary} - -import spark.broadcast._ -import spark.deploy.LocalSparkCluster -import spark.partial.ApproximateEvaluator -import spark.partial.PartialResult -import rdd.{CheckpointRDD, HadoopRDD, NewHadoopRDD, UnionRDD, ParallelCollectionRDD} -import scheduler.{ResultTask, ShuffleMapTask, DAGScheduler, TaskScheduler} +import org.apache.hadoop.mapreduce.lib.input.{FileInputFormat => NewFileInputFormat} + +import org.apache.mesos.MesosNativeLibrary + +import spark.deploy.{LocalSparkCluster, SparkHadoopUtil} +import spark.partial.{ApproximateEvaluator, PartialResult} +import spark.rdd.{CheckpointRDD, HadoopRDD, NewHadoopRDD, UnionRDD, ParallelCollectionRDD} +import spark.scheduler.{DAGScheduler, ResultTask, ShuffleMapTask, SparkListener, SplitInfo, Stage, StageInfo, TaskScheduler} +import spark.scheduler.cluster.{StandaloneSchedulerBackend, SparkDeploySchedulerBackend, ClusterScheduler} import spark.scheduler.local.LocalScheduler -import spark.scheduler.cluster.{SparkDeploySchedulerBackend, SchedulerBackend, ClusterScheduler} import spark.scheduler.mesos.{CoarseMesosSchedulerBackend, MesosSchedulerBackend} -import storage.BlockManagerUI -import util.{MetadataCleaner, TimeStampedHashMap} -import storage.{StorageStatus, StorageUtils, RDDInfo} +import spark.storage.{BlockManagerUI, StorageStatus, StorageUtils, RDDInfo} +import spark.util.{MetadataCleaner, TimeStampedHashMap} /** * Main entry point for Spark functionality. A SparkContext represents the connection to a Spark @@ -64,7 +65,10 @@ class SparkContext( val appName: String, val sparkHome: String = null, val jars: Seq[String] = Nil, - environment: Map[String, String] = Map()) + val environment: Map[String, String] = Map(), + // This is used only by yarn for now, but should be relevant to other cluster types (mesos, etc) too. + // This is typically generated from InputFormatInfo.computePreferredLocations .. host, set of data-local splits on host + val preferredNodeLocationData: scala.collection.Map[String, scala.collection.Set[SplitInfo]] = scala.collection.immutable.Map()) extends Logging { // Ensure logging is initialized before we spawn any threads @@ -72,7 +76,7 @@ class SparkContext( // Set Spark driver host and port system properties if (System.getProperty("spark.driver.host") == null) { - System.setProperty("spark.driver.host", Utils.localIpAddress) + System.setProperty("spark.driver.host", Utils.localHostName()) } if (System.getProperty("spark.driver.port") == null) { System.setProperty("spark.driver.port", "0") @@ -99,12 +103,14 @@ class SparkContext( private[spark] val addedJars = HashMap[String, Long]() // Keeps track of all persisted RDDs - private[spark] val persistentRdds = new TimeStampedHashMap[Int, RDD[_]]() + private[spark] val persistentRdds = new TimeStampedHashMap[Int, RDD[_]] private[spark] val metadataCleaner = new MetadataCleaner("SparkContext", this.cleanup) // Add each JAR given through the constructor - jars.foreach { addJar(_) } + if (jars != null) { + jars.foreach { addJar(_) } + } // Environment variables to pass to our executors private[spark] val executorEnvs = HashMap[String, String]() @@ -116,7 +122,9 @@ class SparkContext( executorEnvs(key) = value } } - executorEnvs ++= environment + if (environment != null) { + executorEnvs ++= environment + } // Create and start the scheduler private var taskScheduler: TaskScheduler = { @@ -169,6 +177,22 @@ class SparkContext( } scheduler + case "yarn-standalone" => + val scheduler = try { + val clazz = Class.forName("spark.scheduler.cluster.YarnClusterScheduler") + val cons = clazz.getConstructor(classOf[SparkContext]) + cons.newInstance(this).asInstanceOf[ClusterScheduler] + } catch { + // TODO: Enumerate the exact reasons why it can fail + // But irrespective of it, it means we cannot proceed ! + case th: Throwable => { + throw new SparkException("YARN mode not available ?", th) + } + } + val backend = new StandaloneSchedulerBackend(scheduler, this.env.actorSystem) + scheduler.initialize(backend) + scheduler + case _ => if (MESOS_REGEX.findFirstIn(master).isEmpty) { logWarning("Master %s does not match expected format, parsing as Mesos URL".format(master)) @@ -188,12 +212,12 @@ class SparkContext( } taskScheduler.start() - private var dagScheduler = new DAGScheduler(taskScheduler) + @volatile private var dagScheduler = new DAGScheduler(taskScheduler) dagScheduler.start() /** A default Hadoop Configuration for the Hadoop code (e.g. file systems) that we reuse. */ val hadoopConfiguration = { - val conf = new Configuration() + val conf = SparkHadoopUtil.newConfiguration() // Explicitly check for S3 environment variables if (System.getenv("AWS_ACCESS_KEY_ID") != null && System.getenv("AWS_SECRET_ACCESS_KEY") != null) { conf.set("fs.s3.awsAccessKeyId", System.getenv("AWS_ACCESS_KEY_ID")) @@ -212,6 +236,22 @@ class SparkContext( private[spark] var checkpointDir: Option[String] = None + // Thread Local variable that can be used by users to pass information down the stack + private val localProperties = new DynamicVariable[Properties](null) + + def initLocalProperties() { + localProperties.value = new Properties() + } + + def addLocalProperties(key: String, value: String) { + if(localProperties.value == null) { + localProperties.value = new Properties() + } + localProperties.value.setProperty(key,value) + } + // Post init + taskScheduler.postStartHook() + // Methods for creating RDDs /** Distribute a local Scala collection to form an RDD. */ @@ -439,7 +479,7 @@ class SparkContext( } /** - * Broadcast a read-only variable to the cluster, returning a [[spark.Broadcast]] object for + * Broadcast a read-only variable to the cluster, returning a [[spark.broadcast.Broadcast]] object for * reading it in distributed functions. The variable will be sent to each cluster only once. */ def broadcast[T](value: T) = env.broadcastManager.newBroadcast[T](value, isLocal) @@ -466,13 +506,17 @@ class SparkContext( logInfo("Added file " + path + " at " + key + " with timestamp " + addedFiles(key)) } + def addSparkListener(listener: SparkListener) { + dagScheduler.sparkListeners += listener + } + /** * Return a map from the slave to the max memory available for caching and the remaining * memory available for caching. */ def getExecutorMemoryStatus: Map[String, (Long, Long)] = { env.blockManager.master.getMemoryStatus.map { case(blockManagerId, mem) => - (blockManagerId.ip + ":" + blockManagerId.port, mem) + (blockManagerId.host + ":" + blockManagerId.port, mem) } } @@ -480,14 +524,18 @@ class SparkContext( * Return information about what RDDs are cached, if they are in mem or on disk, how much space * they take, etc. */ - def getRDDStorageInfo : Array[RDDInfo] = { + def getRDDStorageInfo: Array[RDDInfo] = { StorageUtils.rddInfoFromStorageStatus(getExecutorStorageStatus, this) } + def getStageInfo: Map[Stage,StageInfo] = { + dagScheduler.stageToInfos + } + /** * Return information about blocks stored in all of the slaves */ - def getExecutorStorageStatus : Array[StorageStatus] = { + def getExecutorStorageStatus: Array[StorageStatus] = { env.blockManager.master.getStorageStatus } @@ -505,13 +553,18 @@ class SparkContext( * filesystems), or an HTTP, HTTPS or FTP URI. */ def addJar(path: String) { - val uri = new URI(path) - val key = uri.getScheme match { - case null | "file" => env.httpFileServer.addJar(new File(uri.getPath)) - case _ => path + if (null == path) { + logWarning("null specified as parameter to addJar", + new SparkException("null specified as parameter to addJar")) + } else { + val uri = new URI(path) + val key = uri.getScheme match { + case null | "file" => env.httpFileServer.addJar(new File(uri.getPath)) + case _ => path + } + addedJars(key) = System.currentTimeMillis + logInfo("Added JAR " + path + " at " + key + " with timestamp " + addedJars(key)) } - addedJars(key) = System.currentTimeMillis - logInfo("Added JAR " + path + " at " + key + " with timestamp " + addedJars(key)) } /** @@ -524,10 +577,13 @@ class SparkContext( /** Shut down the SparkContext. */ def stop() { - if (dagScheduler != null) { + // Do this only if not stopped already - best case effort. + // prevent NPE if stopped more than once. + val dagSchedulerCopy = dagScheduler + dagScheduler = null + if (dagSchedulerCopy != null) { metadataCleaner.cancel() - dagScheduler.stop() - dagScheduler = null + dagSchedulerCopy.stop() taskScheduler = null // TODO: Cache.stop()? env.stop() @@ -543,6 +599,7 @@ class SparkContext( } } + /** * Get Spark's home location from either a value set through the constructor, * or the spark.home Java property, or the SPARK_HOME environment variable @@ -572,10 +629,10 @@ class SparkContext( partitions: Seq[Int], allowLocal: Boolean, resultHandler: (Int, U) => Unit) { - val callSite = Utils.getSparkCallSite + val callSite = Utils.formatSparkCallSite logInfo("Starting job: " + callSite) val start = System.nanoTime - val result = dagScheduler.runJob(rdd, func, partitions, callSite, allowLocal, resultHandler) + val result = dagScheduler.runJob(rdd, func, partitions, callSite, allowLocal, resultHandler, localProperties.value) logInfo("Job finished: " + callSite + ", took " + (System.nanoTime - start) / 1e9 + " s") rdd.doCheckpoint() result @@ -654,12 +711,11 @@ class SparkContext( rdd: RDD[T], func: (TaskContext, Iterator[T]) => U, evaluator: ApproximateEvaluator[U, R], - timeout: Long - ): PartialResult[R] = { - val callSite = Utils.getSparkCallSite + timeout: Long): PartialResult[R] = { + val callSite = Utils.formatSparkCallSite logInfo("Starting job: " + callSite) val start = System.nanoTime - val result = dagScheduler.runApproximateJob(rdd, func, evaluator, callSite, timeout) + val result = dagScheduler.runApproximateJob(rdd, func, evaluator, callSite, timeout, localProperties.value) logInfo("Job finished: " + callSite + ", took " + (System.nanoTime - start) / 1e9 + " s") result } @@ -682,7 +738,7 @@ class SparkContext( */ def setCheckpointDir(dir: String, useExisting: Boolean = false) { val path = new Path(dir) - val fs = path.getFileSystem(new Configuration()) + val fs = path.getFileSystem(SparkHadoopUtil.newConfiguration()) if (!useExisting) { if (fs.exists(path)) { throw new Exception("Checkpoint directory '" + path + "' already exists.") @@ -693,7 +749,7 @@ class SparkContext( checkpointDir = Some(dir) } - /** Default level of parallelism to use when not given by user (e.g. for reduce tasks) */ + /** Default level of parallelism to use when not given by user (e.g. parallelize and makeRDD). */ def defaultParallelism: Int = taskScheduler.defaultParallelism /** Default min number of partitions for Hadoop RDDs when not given by user */ diff --git a/core/src/main/scala/spark/SparkEnv.scala b/core/src/main/scala/spark/SparkEnv.scala index d2193ae72b..7ccde2e818 100644 --- a/core/src/main/scala/spark/SparkEnv.scala +++ b/core/src/main/scala/spark/SparkEnv.scala @@ -1,15 +1,19 @@ package spark -import akka.actor.ActorSystem -import akka.actor.ActorSystemImpl +import collection.mutable +import serializer.Serializer + +import akka.actor.{Actor, ActorRef, Props, ActorSystemImpl, ActorSystem} import akka.remote.RemoteActorRefProvider -import serializer.Serializer import spark.broadcast.BroadcastManager import spark.storage.BlockManager import spark.storage.BlockManagerMaster import spark.network.ConnectionManager +import spark.serializer.{Serializer, SerializerManager} import spark.util.AkkaUtils +import spark.api.python.PythonWorkerFactory + /** * Holds all the runtime environment objects for a running Spark instance (either master or worker), @@ -21,6 +25,7 @@ import spark.util.AkkaUtils class SparkEnv ( val executorId: String, val actorSystem: ActorSystem, + val serializerManager: SerializerManager, val serializer: Serializer, val closureSerializer: Serializer, val cacheManager: CacheManager, @@ -30,10 +35,16 @@ class SparkEnv ( val blockManager: BlockManager, val connectionManager: ConnectionManager, val httpFileServer: HttpFileServer, - val sparkFilesDir: String - ) { + val sparkFilesDir: String, + // To be set only as part of initialization of SparkContext. + // (executorId, defaultHostPort) => executorHostPort + // If executorId is NOT found, return defaultHostPort + var executorIdToHostPort: Option[(String, String) => String]) { + + private val pythonWorkers = mutable.HashMap[(String, Map[String, String]), PythonWorkerFactory]() def stop() { + pythonWorkers.foreach { case(key, worker) => worker.stop() } httpFileServer.stop() mapOutputTracker.stop() shuffleFetcher.stop() @@ -45,6 +56,22 @@ class SparkEnv ( // down, but let's call it anyway in case it gets fixed in a later release actorSystem.awaitTermination() } + + def createPythonWorker(pythonExec: String, envVars: Map[String, String]): java.net.Socket = { + synchronized { + pythonWorkers.getOrElseUpdate((pythonExec, envVars), new PythonWorkerFactory(pythonExec, envVars)).create() + } + } + + def resolveExecutorIdToHostPort(executorId: String, defaultHostPort: String): String = { + val env = SparkEnv.get + if (env.executorIdToHostPort.isEmpty) { + // default to using host, not host port. Relevant to non cluster modes. + return defaultHostPort + } + + env.executorIdToHostPort.get(executorId, defaultHostPort) + } } object SparkEnv extends Logging { @@ -73,6 +100,16 @@ object SparkEnv extends Logging { System.setProperty("spark.driver.port", boundPort.toString) } + // set only if unset until now. + if (System.getProperty("spark.hostPort", null) == null) { + if (!isDriver){ + // unexpected + Utils.logErrorWithStack("Unexpected NOT to have spark.hostPort set") + } + Utils.checkHost(hostname) + System.setProperty("spark.hostPort", hostname + ":" + boundPort) + } + val classLoader = Thread.currentThread.getContextClassLoader // Create an instance of the class named by the given Java system property, or by @@ -82,24 +119,45 @@ object SparkEnv extends Logging { Class.forName(name, true, classLoader).newInstance().asInstanceOf[T] } - val serializer = instantiateClass[Serializer]("spark.serializer", "spark.JavaSerializer") + val serializerManager = new SerializerManager + + val serializer = serializerManager.setDefault( + System.getProperty("spark.serializer", "spark.JavaSerializer")) + + val closureSerializer = serializerManager.get( + System.getProperty("spark.closure.serializer", "spark.JavaSerializer")) + + def registerOrLookup(name: String, newActor: => Actor): ActorRef = { + if (isDriver) { + logInfo("Registering " + name) + actorSystem.actorOf(Props(newActor), name = name) + } else { + val driverHost: String = System.getProperty("spark.driver.host", "localhost") + val driverPort: Int = System.getProperty("spark.driver.port", "7077").toInt + Utils.checkHost(driverHost, "Expected hostname") + val url = "akka://spark@%s:%s/user/%s".format(driverHost, driverPort, name) + logInfo("Connecting to " + name + ": " + url) + actorSystem.actorFor(url) + } + } - val driverIp: String = System.getProperty("spark.driver.host", "localhost") - val driverPort: Int = System.getProperty("spark.driver.port", "7077").toInt - val blockManagerMaster = new BlockManagerMaster( - actorSystem, isDriver, isLocal, driverIp, driverPort) + val blockManagerMaster = new BlockManagerMaster(registerOrLookup( + "BlockManagerMaster", + new spark.storage.BlockManagerMasterActor(isLocal))) val blockManager = new BlockManager(executorId, actorSystem, blockManagerMaster, serializer) val connectionManager = blockManager.connectionManager val broadcastManager = new BroadcastManager(isDriver) - val closureSerializer = instantiateClass[Serializer]( - "spark.closure.serializer", "spark.JavaSerializer") - val cacheManager = new CacheManager(blockManager) - val mapOutputTracker = new MapOutputTracker(actorSystem, isDriver) + // Have to assign trackerActor after initialization as MapOutputTrackerActor + // requires the MapOutputTracker itself + val mapOutputTracker = new MapOutputTracker() + mapOutputTracker.trackerActor = registerOrLookup( + "MapOutputTracker", + new MapOutputTrackerActor(mapOutputTracker)) val shuffleFetcher = instantiateClass[ShuffleFetcher]( "spark.shuffle.fetcher", "spark.BlockStoreShuffleFetcher") @@ -126,6 +184,7 @@ object SparkEnv extends Logging { new SparkEnv( executorId, actorSystem, + serializerManager, serializer, closureSerializer, cacheManager, @@ -135,6 +194,7 @@ object SparkEnv extends Logging { blockManager, connectionManager, httpFileServer, - sparkFilesDir) + sparkFilesDir, + None) } } diff --git a/core/src/main/scala/spark/TaskContext.scala b/core/src/main/scala/spark/TaskContext.scala index eab85f85a2..dd0609026a 100644 --- a/core/src/main/scala/spark/TaskContext.scala +++ b/core/src/main/scala/spark/TaskContext.scala @@ -1,9 +1,14 @@ package spark +import executor.TaskMetrics import scala.collection.mutable.ArrayBuffer - -class TaskContext(val stageId: Int, val splitId: Int, val attemptId: Long) extends Serializable { +class TaskContext( + val stageId: Int, + val splitId: Int, + val attemptId: Long, + val taskMetrics: TaskMetrics = TaskMetrics.empty() +) extends Serializable { @transient val onCompleteCallbacks = new ArrayBuffer[() => Unit] diff --git a/core/src/main/scala/spark/TaskEndReason.scala b/core/src/main/scala/spark/TaskEndReason.scala index 420c54bc9a..8140cba084 100644 --- a/core/src/main/scala/spark/TaskEndReason.scala +++ b/core/src/main/scala/spark/TaskEndReason.scala @@ -14,9 +14,19 @@ private[spark] case object Success extends TaskEndReason private[spark] case object Resubmitted extends TaskEndReason // Task was finished earlier but we've now lost it -private[spark] -case class FetchFailed(bmAddress: BlockManagerId, shuffleId: Int, mapId: Int, reduceId: Int) extends TaskEndReason +private[spark] case class FetchFailed( + bmAddress: BlockManagerId, + shuffleId: Int, + mapId: Int, + reduceId: Int) + extends TaskEndReason -private[spark] case class ExceptionFailure(exception: Throwable) extends TaskEndReason +private[spark] case class ExceptionFailure( + className: String, + description: String, + stackTrace: Array[StackTraceElement]) + extends TaskEndReason private[spark] case class OtherFailure(message: String) extends TaskEndReason + +private[spark] case class TaskResultTooBigFailure() extends TaskEndReason diff --git a/core/src/main/scala/spark/Utils.scala b/core/src/main/scala/spark/Utils.scala index 81daacf958..f3621c6bee 100644 --- a/core/src/main/scala/spark/Utils.scala +++ b/core/src/main/scala/spark/Utils.scala @@ -1,23 +1,29 @@ package spark import java.io._ -import java.net._ +import java.net.{InetAddress, URL, URI, NetworkInterface, Inet4Address, ServerSocket} import java.util.{Locale, Random, UUID} -import java.util.concurrent.{Executors, ThreadFactory, ThreadPoolExecutor} -import org.apache.hadoop.conf.Configuration -import org.apache.hadoop.fs.{Path, FileSystem, FileUtil} -import scala.collection.mutable.ArrayBuffer +import java.util.concurrent.{ConcurrentHashMap, Executors, ThreadFactory, ThreadPoolExecutor} +import java.util.regex.Pattern + +import scala.collection.mutable.{ArrayBuffer, HashMap} import scala.collection.JavaConversions._ import scala.io.Source + import com.google.common.io.Files import com.google.common.util.concurrent.ThreadFactoryBuilder -import scala.Some + +import org.apache.hadoop.fs.{Path, FileSystem, FileUtil} + import spark.serializer.SerializerInstance +import spark.deploy.SparkHadoopUtil + /** * Various utility methods used by Spark. */ private object Utils extends Logging { + /** Serialize an object using Java serialization */ def serialize[T](o: T): Array[Byte] = { val bos = new ByteArrayOutputStream() @@ -68,6 +74,40 @@ private object Utils extends Logging { return buf } + private val shutdownDeletePaths = new collection.mutable.HashSet[String]() + + // Register the path to be deleted via shutdown hook + def registerShutdownDeleteDir(file: File) { + val absolutePath = file.getAbsolutePath() + shutdownDeletePaths.synchronized { + shutdownDeletePaths += absolutePath + } + } + + // Is the path already registered to be deleted via a shutdown hook ? + def hasShutdownDeleteDir(file: File): Boolean = { + val absolutePath = file.getAbsolutePath() + shutdownDeletePaths.synchronized { + shutdownDeletePaths.contains(absolutePath) + } + } + + // Note: if file is child of some registered path, while not equal to it, then return true; + // else false. This is to ensure that two shutdown hooks do not try to delete each others + // paths - resulting in IOException and incomplete cleanup. + def hasRootAsShutdownDeleteDir(file: File): Boolean = { + val absolutePath = file.getAbsolutePath() + val retval = shutdownDeletePaths.synchronized { + shutdownDeletePaths.find { path => + !absolutePath.equals(path) && absolutePath.startsWith(path) + }.isDefined + } + if (retval) { + logInfo("path = " + file + ", already present as root for deletion.") + } + retval + } + /** Create a temporary directory inside the given parent directory */ def createTempDir(root: String = System.getProperty("java.io.tmpdir")): File = { var attempts = 0 @@ -76,8 +116,8 @@ private object Utils extends Logging { while (dir == null) { attempts += 1 if (attempts > maxAttempts) { - throw new IOException("Failed to create a temp directory after " + maxAttempts + - " attempts!") + throw new IOException("Failed to create a temp directory (under " + root + ") after " + + maxAttempts + " attempts!") } try { dir = new File(root, "spark-" + UUID.randomUUID.toString) @@ -86,13 +126,17 @@ private object Utils extends Logging { } } catch { case e: IOException => ; } } + + registerShutdownDeleteDir(dir) + // Add a shutdown hook to delete the temp dir when the JVM exits Runtime.getRuntime.addShutdownHook(new Thread("delete Spark temp dir " + dir) { override def run() { - Utils.deleteRecursively(dir) + // Attempt to delete if some patch which is parent of this is not already registered. + if (! hasRootAsShutdownDeleteDir(dir)) Utils.deleteRecursively(dir) } }) - return dir + dir } /** Copy all data from an InputStream to an OutputStream */ @@ -135,40 +179,35 @@ private object Utils extends Logging { Utils.copyStream(in, out, true) if (targetFile.exists && !Files.equal(tempFile, targetFile)) { tempFile.delete() - throw new SparkException("File " + targetFile + " exists and does not match contents of" + - " " + url) + throw new SparkException( + "File " + targetFile + " exists and does not match contents of" + " " + url) } else { Files.move(tempFile, targetFile) } case "file" | null => - val sourceFile = if (uri.isAbsolute) { - new File(uri) - } else { - new File(url) - } - if (targetFile.exists && !Files.equal(sourceFile, targetFile)) { - throw new SparkException("File " + targetFile + " exists and does not match contents of" + - " " + url) - } else { - // Remove the file if it already exists - targetFile.delete() - // Symlink the file locally. - if (uri.isAbsolute) { - // url is absolute, i.e. it starts with "file:///". Extract the source - // file's absolute path from the url. - val sourceFile = new File(uri) - logInfo("Symlinking " + sourceFile.getAbsolutePath + " to " + targetFile.getAbsolutePath) - FileUtil.symLink(sourceFile.getAbsolutePath, targetFile.getAbsolutePath) + // In the case of a local file, copy the local file to the target directory. + // Note the difference between uri vs url. + val sourceFile = if (uri.isAbsolute) new File(uri) else new File(url) + if (targetFile.exists) { + // If the target file already exists, warn the user if + if (!Files.equal(sourceFile, targetFile)) { + throw new SparkException( + "File " + targetFile + " exists and does not match contents of" + " " + url) } else { - // url is not absolute, i.e. itself is the path to the source file. - logInfo("Symlinking " + url + " to " + targetFile.getAbsolutePath) - FileUtil.symLink(url, targetFile.getAbsolutePath) + // Do nothing if the file contents are the same, i.e. this file has been copied + // previously. + logInfo(sourceFile.getAbsolutePath + " has been previously copied to " + + targetFile.getAbsolutePath) } + } else { + // The file does not exist in the target directory. Copy it there. + logInfo("Copying " + sourceFile.getAbsolutePath + " to " + targetFile.getAbsolutePath) + Files.copy(sourceFile, targetFile) } case _ => // Use the Hadoop filesystem library, which supports file://, hdfs://, s3://, and others val uri = new URI(url) - val conf = new Configuration() + val conf = SparkHadoopUtil.newConfiguration() val fs = FileSystem.get(uri, conf) val in = fs.open(new Path(uri)) val out = new FileOutputStream(tempFile) @@ -227,8 +266,10 @@ private object Utils extends Logging { /** * Get the local host's IP address in dotted-quad format (e.g. 1.2.3.4). + * Note, this is typically not used from within core spark. */ lazy val localIpAddress: String = findLocalIpAddress() + lazy val localIpAddressHostname: String = getAddressHostName(localIpAddress) private def findLocalIpAddress(): String = { val defaultIpOverride = System.getenv("SPARK_LOCAL_IP") @@ -266,6 +307,8 @@ private object Utils extends Logging { * hostname it reports to the master. */ def setCustomHostname(hostname: String) { + // DEBUG code + Utils.checkHost(hostname) customHostname = Some(hostname) } @@ -273,7 +316,91 @@ private object Utils extends Logging { * Get the local machine's hostname. */ def localHostName(): String = { - customHostname.getOrElse(InetAddress.getLocalHost.getHostName) + customHostname.getOrElse(localIpAddressHostname) + } + + def getAddressHostName(address: String): String = { + InetAddress.getByName(address).getHostName + } + + def localHostPort(): String = { + val retval = System.getProperty("spark.hostPort", null) + if (retval == null) { + logErrorWithStack("spark.hostPort not set but invoking localHostPort") + return localHostName() + } + + retval + } + +/* + // Used by DEBUG code : remove when all testing done + private val ipPattern = Pattern.compile("^[0-9]+(\\.[0-9]+)*$") + def checkHost(host: String, message: String = "") { + // Currently catches only ipv4 pattern, this is just a debugging tool - not rigourous ! + // if (host.matches("^[0-9]+(\\.[0-9]+)*$")) { + if (ipPattern.matcher(host).matches()) { + Utils.logErrorWithStack("Unexpected to have host " + host + " which matches IP pattern. Message " + message) + } + if (Utils.parseHostPort(host)._2 != 0){ + Utils.logErrorWithStack("Unexpected to have host " + host + " which has port in it. Message " + message) + } + } + + // Used by DEBUG code : remove when all testing done + def checkHostPort(hostPort: String, message: String = "") { + val (host, port) = Utils.parseHostPort(hostPort) + checkHost(host) + if (port <= 0){ + Utils.logErrorWithStack("Unexpected to have port " + port + " which is not valid in " + hostPort + ". Message " + message) + } + } + + // Used by DEBUG code : remove when all testing done + def logErrorWithStack(msg: String) { + try { throw new Exception } catch { case ex: Exception => { logError(msg, ex) } } + // temp code for debug + System.exit(-1) + } +*/ + + // Once testing is complete in various modes, replace with this ? + def checkHost(host: String, message: String = "") {} + def checkHostPort(hostPort: String, message: String = "") {} + + // Used by DEBUG code : remove when all testing done + def logErrorWithStack(msg: String) { + try { throw new Exception } catch { case ex: Exception => { logError(msg, ex) } } + } + + def getUserNameFromEnvironment(): String = { + SparkHadoopUtil.getUserNameFromEnvironment + } + + // Typically, this will be of order of number of nodes in cluster + // If not, we should change it to LRUCache or something. + private val hostPortParseResults = new ConcurrentHashMap[String, (String, Int)]() + + def parseHostPort(hostPort: String): (String, Int) = { + { + // Check cache first. + var cached = hostPortParseResults.get(hostPort) + if (cached != null) return cached + } + + val indx: Int = hostPort.lastIndexOf(':') + // This is potentially broken - when dealing with ipv6 addresses for example, sigh ... + // but then hadoop does not support ipv6 right now. + // For now, we assume that if port exists, then it is valid - not check if it is an int > 0 + if (-1 == indx) { + val retval = (hostPort, 0) + hostPortParseResults.put(hostPort, retval) + return retval + } + + val retval = (hostPort.substring(0, indx).trim(), hostPort.substring(indx + 1).trim().toInt) + hostPortParseResults.putIfAbsent(hostPort, retval) + hostPortParseResults.get(hostPort) } private[spark] val daemonThreadFactory: ThreadFactory = @@ -395,13 +522,14 @@ private object Utils extends Logging { execute(command, new File(".")) } - + private[spark] class CallSiteInfo(val lastSparkMethod: String, val firstUserFile: String, + val firstUserLine: Int, val firstUserClass: String) /** * When called inside a class in the spark package, returns the name of the user code class * (outside the spark package) that called into Spark, as well as which Spark method they called. * This is used, for example, to tell users where in their code each RDD got created. */ - def getSparkCallSite: String = { + def getCallSiteInfo: CallSiteInfo = { val trace = Thread.currentThread.getStackTrace().filter( el => (!el.getMethodName.contains("getStackTrace"))) @@ -413,6 +541,7 @@ private object Utils extends Logging { var firstUserFile = "<unknown>" var firstUserLine = 0 var finished = false + var firstUserClass = "<unknown>" for (el <- trace) { if (!finished) { @@ -427,13 +556,19 @@ private object Utils extends Logging { else { firstUserLine = el.getLineNumber firstUserFile = el.getFileName + firstUserClass = el.getClassName finished = true } } } - "%s at %s:%s".format(lastSparkMethod, firstUserFile, firstUserLine) + new CallSiteInfo(lastSparkMethod, firstUserFile, firstUserLine, firstUserClass) } + def formatSparkCallSite = { + val callSiteInfo = getCallSiteInfo + "%s at %s:%s".format(callSiteInfo.lastSparkMethod, callSiteInfo.firstUserFile, + callSiteInfo.firstUserLine) + } /** * Try to find a free port to bind to on the local host. This should ideally never be needed, * except that, unfortunately, some of the networking libraries we currently rely on (e.g. Spray) diff --git a/core/src/main/scala/spark/api/java/JavaDoubleRDD.scala b/core/src/main/scala/spark/api/java/JavaDoubleRDD.scala index da3cb2cd31..16692c0440 100644 --- a/core/src/main/scala/spark/api/java/JavaDoubleRDD.scala +++ b/core/src/main/scala/spark/api/java/JavaDoubleRDD.scala @@ -6,8 +6,8 @@ import spark.api.java.function.{Function => JFunction} import spark.util.StatCounter import spark.partial.{BoundedDouble, PartialResult} import spark.storage.StorageLevel - import java.lang.Double +import spark.Partitioner class JavaDoubleRDD(val srdd: RDD[scala.Double]) extends JavaRDDLike[Double, JavaDoubleRDD] { @@ -58,6 +58,33 @@ class JavaDoubleRDD(val srdd: RDD[scala.Double]) extends JavaRDDLike[Double, Jav def coalesce(numPartitions: Int): JavaDoubleRDD = fromRDD(srdd.coalesce(numPartitions)) /** + * Return a new RDD that is reduced into `numPartitions` partitions. + */ + def coalesce(numPartitions: Int, shuffle: Boolean): JavaDoubleRDD = + fromRDD(srdd.coalesce(numPartitions, shuffle)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + * + * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting + * RDD will be <= us. + */ + def subtract(other: JavaDoubleRDD): JavaDoubleRDD = + fromRDD(srdd.subtract(other)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: JavaDoubleRDD, numPartitions: Int): JavaDoubleRDD = + fromRDD(srdd.subtract(other, numPartitions)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: JavaDoubleRDD, p: Partitioner): JavaDoubleRDD = + fromRDD(srdd.subtract(other, p)) + + /** * Return a sampled subset of this RDD. */ def sample(withReplacement: Boolean, fraction: Double, seed: Int): JavaDoubleRDD = diff --git a/core/src/main/scala/spark/api/java/JavaPairRDD.scala b/core/src/main/scala/spark/api/java/JavaPairRDD.scala index df3af3817d..76051597b6 100644 --- a/core/src/main/scala/spark/api/java/JavaPairRDD.scala +++ b/core/src/main/scala/spark/api/java/JavaPairRDD.scala @@ -6,6 +6,7 @@ import java.util.Comparator import scala.Tuple2 import scala.collection.JavaConversions._ +import org.apache.hadoop.io.compress.CompressionCodec import org.apache.hadoop.mapred.JobConf import org.apache.hadoop.mapred.OutputFormat import org.apache.hadoop.mapreduce.{OutputFormat => NewOutputFormat} @@ -19,6 +20,7 @@ import spark.OrderedRDDFunctions import spark.storage.StorageLevel import spark.HashPartitioner import spark.Partitioner +import spark.Partitioner._ import spark.RDD import spark.SparkContext.rddToPairRDDFunctions @@ -59,13 +61,19 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif /** * Return a new RDD containing only the elements that satisfy a predicate. */ - def filter(f: Function[(K, V), java.lang.Boolean]): JavaPairRDD[K, V] = + def filter(f: JFunction[(K, V), java.lang.Boolean]): JavaPairRDD[K, V] = new JavaPairRDD[K, V](rdd.filter(x => f(x).booleanValue())) /** * Return a new RDD that is reduced into `numPartitions` partitions. */ - def coalesce(numPartitions: Int): JavaPairRDD[K, V] = new JavaPairRDD[K, V](rdd.coalesce(numPartitions)) + def coalesce(numPartitions: Int): JavaPairRDD[K, V] = fromRDD(rdd.coalesce(numPartitions)) + + /** + * Return a new RDD that is reduced into `numPartitions` partitions. + */ + def coalesce(numPartitions: Int, shuffle: Boolean): JavaPairRDD[K, V] = + fromRDD(rdd.coalesce(numPartitions, shuffle)) /** * Return a sampled subset of this RDD. @@ -102,7 +110,7 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif * In addition, users can control the partitioning of the output RDD, and whether to perform * map-side aggregation (if a mapper can produce multiple items with the same key). */ - def combineByKey[C](createCombiner: Function[V, C], + def combineByKey[C](createCombiner: JFunction[V, C], mergeValue: JFunction2[C, V, C], mergeCombiners: JFunction2[C, C, C], partitioner: Partitioner): JavaPairRDD[K, C] = { @@ -160,6 +168,30 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif rdd.countByKeyApprox(timeout, confidence).map(mapAsJavaMap) /** + * Merge the values for each key using an associative function and a neutral "zero value" which may + * be added to the result an arbitrary number of times, and must not change the result (e.g., Nil for + * list concatenation, 0 for addition, or 1 for multiplication.). + */ + def foldByKey(zeroValue: V, partitioner: Partitioner, func: JFunction2[V, V, V]): JavaPairRDD[K, V] = + fromRDD(rdd.foldByKey(zeroValue, partitioner)(func)) + + /** + * Merge the values for each key using an associative function and a neutral "zero value" which may + * be added to the result an arbitrary number of times, and must not change the result (e.g., Nil for + * list concatenation, 0 for addition, or 1 for multiplication.). + */ + def foldByKey(zeroValue: V, numPartitions: Int, func: JFunction2[V, V, V]): JavaPairRDD[K, V] = + fromRDD(rdd.foldByKey(zeroValue, numPartitions)(func)) + + /** + * Merge the values for each key using an associative function and a neutral "zero value" which may + * be added to the result an arbitrary number of times, and must not change the result (e.g., Nil for + * list concatenation, 0 for addition, or 1 for multiplication.). + */ + def foldByKey(zeroValue: V, func: JFunction2[V, V, V]): JavaPairRDD[K, V] = + fromRDD(rdd.foldByKey(zeroValue)(func)) + + /** * Merge the values for each key using an associative reduce function. This will also perform * the merging locally on each mapper before sending results to a reducer, similarly to a * "combiner" in MapReduce. Output will be hash-partitioned with numPartitions partitions. @@ -182,6 +214,27 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif fromRDD(groupByResultToJava(rdd.groupByKey(numPartitions))) /** + * Return an RDD with the elements from `this` that are not in `other`. + * + * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting + * RDD will be <= us. + */ + def subtract(other: JavaPairRDD[K, V]): JavaPairRDD[K, V] = + fromRDD(rdd.subtract(other)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: JavaPairRDD[K, V], numPartitions: Int): JavaPairRDD[K, V] = + fromRDD(rdd.subtract(other, numPartitions)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: JavaPairRDD[K, V], p: Partitioner): JavaPairRDD[K, V] = + fromRDD(rdd.subtract(other, p)) + + /** * Return a copy of the RDD partitioned using the specified partitioner. If `mapSideCombine` * is true, Spark will group values of the same key together on the map side before the * repartitioning, to only send each key over the network once. If a large number of @@ -220,30 +273,30 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif fromRDD(rdd.rightOuterJoin(other, partitioner)) /** - * Simplified version of combineByKey that hash-partitions the resulting RDD using the default - * parallelism level. + * Simplified version of combineByKey that hash-partitions the resulting RDD using the existing + * partitioner/parallelism level. */ def combineByKey[C](createCombiner: JFunction[V, C], mergeValue: JFunction2[C, V, C], mergeCombiners: JFunction2[C, C, C]): JavaPairRDD[K, C] = { implicit val cm: ClassManifest[C] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[C]] - fromRDD(combineByKey(createCombiner, mergeValue, mergeCombiners)) + fromRDD(combineByKey(createCombiner, mergeValue, mergeCombiners, defaultPartitioner(rdd))) } /** * Merge the values for each key using an associative reduce function. This will also perform * the merging locally on each mapper before sending results to a reducer, similarly to a - * "combiner" in MapReduce. Output will be hash-partitioned with the default parallelism level. + * "combiner" in MapReduce. Output will be hash-partitioned with the existing partitioner/ + * parallelism level. */ def reduceByKey(func: JFunction2[V, V, V]): JavaPairRDD[K, V] = { - val partitioner = rdd.defaultPartitioner(rdd) - fromRDD(reduceByKey(partitioner, func)) + fromRDD(reduceByKey(defaultPartitioner(rdd), func)) } /** * Group the values for each key in the RDD into a single sequence. Hash-partitions the - * resulting RDD with the default parallelism level. + * resulting RDD with the existing partitioner/parallelism level. */ def groupByKey(): JavaPairRDD[K, JList[V]] = fromRDD(groupByResultToJava(rdd.groupByKey())) @@ -268,7 +321,7 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif * Perform a left outer join of `this` and `other`. For each element (k, v) in `this`, the * resulting RDD will either contain all pairs (k, (v, Some(w))) for w in `other`, or the * pair (k, (v, None)) if no elements in `other` have key k. Hash-partitions the output - * using the default level of parallelism. + * using the existing partitioner/parallelism level. */ def leftOuterJoin[W](other: JavaPairRDD[K, W]): JavaPairRDD[K, (V, Option[W])] = fromRDD(rdd.leftOuterJoin(other)) @@ -286,7 +339,7 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif * Perform a right outer join of `this` and `other`. For each element (k, w) in `other`, the * resulting RDD will either contain all pairs (k, (Some(v), w)) for v in `this`, or the * pair (k, (None, w)) if no elements in `this` have key k. Hash-partitions the resulting - * RDD using the default parallelism level. + * RDD using the existing partitioner/parallelism level. */ def rightOuterJoin[W](other: JavaPairRDD[K, W]): JavaPairRDD[K, (Option[V], W)] = fromRDD(rdd.rightOuterJoin(other)) @@ -309,7 +362,7 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif * Pass each value in the key-value pair RDD through a map function without changing the keys; * this also retains the original RDD's partitioning. */ - def mapValues[U](f: Function[V, U]): JavaPairRDD[K, U] = { + def mapValues[U](f: JFunction[V, U]): JavaPairRDD[K, U] = { implicit val cm: ClassManifest[U] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[U]] fromRDD(rdd.mapValues(f)) @@ -407,6 +460,16 @@ class JavaPairRDD[K, V](val rdd: RDD[(K, V)])(implicit val kManifest: ClassManif rdd.saveAsHadoopFile(path, keyClass, valueClass, outputFormatClass) } + /** Output the RDD to any Hadoop-supported file system, compressing with the supplied codec. */ + def saveAsHadoopFile[F <: OutputFormat[_, _]]( + path: String, + keyClass: Class[_], + valueClass: Class[_], + outputFormatClass: Class[F], + codec: Class[_ <: CompressionCodec]) { + rdd.saveAsHadoopFile(path, keyClass, valueClass, outputFormatClass, codec) + } + /** Output the RDD to any Hadoop-supported file system. */ def saveAsNewAPIHadoopFile[F <: NewOutputFormat[_, _]]( path: String, diff --git a/core/src/main/scala/spark/api/java/JavaRDD.scala b/core/src/main/scala/spark/api/java/JavaRDD.scala index 3ccd6f055e..626b499454 100644 --- a/core/src/main/scala/spark/api/java/JavaRDD.scala +++ b/core/src/main/scala/spark/api/java/JavaRDD.scala @@ -14,12 +14,18 @@ JavaRDDLike[T, JavaRDD[T]] { /** Persist this RDD with the default storage level (`MEMORY_ONLY`). */ def cache(): JavaRDD[T] = wrapRDD(rdd.cache()) - /** + /** * Set this RDD's storage level to persist its values across operations after the first time - * it is computed. Can only be called once on each RDD. + * it is computed. This can only be used to assign a new storage level if the RDD does not + * have a storage level set yet.. */ def persist(newLevel: StorageLevel): JavaRDD[T] = wrapRDD(rdd.persist(newLevel)) + /** + * Mark the RDD as non-persistent, and remove all blocks for it from memory and disk. + */ + def unpersist(): JavaRDD[T] = wrapRDD(rdd.unpersist()) + // Transformations (return a new RDD) /** @@ -31,7 +37,7 @@ JavaRDDLike[T, JavaRDD[T]] { * Return a new RDD containing the distinct elements in this RDD. */ def distinct(numPartitions: Int): JavaRDD[T] = wrapRDD(rdd.distinct(numPartitions)) - + /** * Return a new RDD containing only the elements that satisfy a predicate. */ @@ -44,17 +50,42 @@ JavaRDDLike[T, JavaRDD[T]] { def coalesce(numPartitions: Int): JavaRDD[T] = rdd.coalesce(numPartitions) /** + * Return a new RDD that is reduced into `numPartitions` partitions. + */ + def coalesce(numPartitions: Int, shuffle: Boolean): JavaRDD[T] = + rdd.coalesce(numPartitions, shuffle) + + /** * Return a sampled subset of this RDD. */ def sample(withReplacement: Boolean, fraction: Double, seed: Int): JavaRDD[T] = wrapRDD(rdd.sample(withReplacement, fraction, seed)) - + /** * Return the union of this RDD and another one. Any identical elements will appear multiple * times (use `.distinct()` to eliminate them). */ def union(other: JavaRDD[T]): JavaRDD[T] = wrapRDD(rdd.union(other.rdd)) + /** + * Return an RDD with the elements from `this` that are not in `other`. + * + * Uses `this` partitioner/partition size, because even if `other` is huge, the resulting + * RDD will be <= us. + */ + def subtract(other: JavaRDD[T]): JavaRDD[T] = wrapRDD(rdd.subtract(other)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: JavaRDD[T], numPartitions: Int): JavaRDD[T] = + wrapRDD(rdd.subtract(other, numPartitions)) + + /** + * Return an RDD with the elements from `this` that are not in `other`. + */ + def subtract(other: JavaRDD[T], p: Partitioner): JavaRDD[T] = + wrapRDD(rdd.subtract(other, p)) } object JavaRDD { diff --git a/core/src/main/scala/spark/api/java/JavaRDDLike.scala b/core/src/main/scala/spark/api/java/JavaRDDLike.scala index 90b45cf875..b555f2030a 100644 --- a/core/src/main/scala/spark/api/java/JavaRDDLike.scala +++ b/core/src/main/scala/spark/api/java/JavaRDDLike.scala @@ -1,9 +1,10 @@ package spark.api.java -import java.util.{List => JList} +import java.util.{List => JList, Comparator} import scala.Tuple2 import scala.collection.JavaConversions._ +import org.apache.hadoop.io.compress.CompressionCodec import spark.{SparkContext, Partition, RDD, TaskContext} import spark.api.java.JavaPairRDD._ import spark.api.java.function.{Function2 => JFunction2, Function => JFunction, _} @@ -12,7 +13,7 @@ import spark.storage.StorageLevel import com.google.common.base.Optional -trait JavaRDDLike[T, This <: JavaRDDLike[T, This]] extends PairFlatMapWorkaround[T] { +trait JavaRDDLike[T, This <: JavaRDDLike[T, This]] extends Serializable { def wrapRDD(rdd: RDD[T]): This implicit val classManifest: ClassManifest[T] @@ -82,12 +83,13 @@ trait JavaRDDLike[T, This <: JavaRDDLike[T, This]] extends PairFlatMapWorkaround } /** - * Part of the workaround for SPARK-668; called in PairFlatMapWorkaround.java. + * Return a new RDD by first applying a function to all elements of this + * RDD, and then flattening the results. */ - private[spark] def doFlatMap[K, V](f: PairFlatMapFunction[T, K, V]): JavaPairRDD[K, V] = { + def flatMap[K2, V2](f: PairFlatMapFunction[T, K2, V2]): JavaPairRDD[K2, V2] = { import scala.collection.JavaConverters._ def fn = (x: T) => f.apply(x).asScala - def cm = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[Tuple2[K, V]]] + def cm = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[Tuple2[K2, V2]]] JavaPairRDD.fromRDD(rdd.flatMap(fn)(cm))(f.keyType(), f.valueType()) } @@ -110,8 +112,8 @@ trait JavaRDDLike[T, This <: JavaRDDLike[T, This]] extends PairFlatMapWorkaround /** * Return a new RDD by applying a function to each partition of this RDD. */ - def mapPartitions[K, V](f: PairFlatMapFunction[java.util.Iterator[T], K, V]): - JavaPairRDD[K, V] = { + def mapPartitions[K2, V2](f: PairFlatMapFunction[java.util.Iterator[T], K2, V2]): + JavaPairRDD[K2, V2] = { def fn = (x: Iterator[T]) => asScalaIterator(f.apply(asJavaIterator(x)).iterator()) JavaPairRDD.fromRDD(rdd.mapPartitions(fn))(f.keyType(), f.valueType()) } @@ -181,6 +183,21 @@ trait JavaRDDLike[T, This <: JavaRDDLike[T, This]] extends PairFlatMapWorkaround JavaPairRDD.fromRDD(rdd.zip(other.rdd)(other.classManifest))(classManifest, other.classManifest) } + /** + * Zip this RDD's partitions with one (or more) RDD(s) and return a new RDD by + * applying a function to the zipped partitions. Assumes that all the RDDs have the + * *same number of partitions*, but does *not* require them to have the same number + * of elements in each partition. + */ + def zipPartitions[U, V]( + f: FlatMapFunction2[java.util.Iterator[T], java.util.Iterator[U], V], + other: JavaRDDLike[U, _]): JavaRDD[V] = { + def fn = (x: Iterator[T], y: Iterator[U]) => asScalaIterator( + f.apply(asJavaIterator(x), asJavaIterator(y)).iterator()) + JavaRDD.fromRDD( + rdd.zipPartitions(fn, other.rdd)(other.classManifest, f.elementType()))(f.elementType()) + } + // Actions (launch a job to return a value to the user program) /** @@ -294,6 +311,13 @@ trait JavaRDDLike[T, This <: JavaRDDLike[T, This]] extends PairFlatMapWorkaround */ def saveAsTextFile(path: String) = rdd.saveAsTextFile(path) + + /** + * Save this RDD as a compressed text file, using string representations of elements. + */ + def saveAsTextFile(path: String, codec: Class[_ <: CompressionCodec]) = + rdd.saveAsTextFile(path, codec) + /** * Save this RDD as a SequenceFile of serialized objects. */ @@ -335,4 +359,29 @@ trait JavaRDDLike[T, This <: JavaRDDLike[T, This]] extends PairFlatMapWorkaround def toDebugString(): String = { rdd.toDebugString } + + /** + * Returns the top K elements from this RDD as defined by + * the specified Comparator[T]. + * @param num the number of top elements to return + * @param comp the comparator that defines the order + * @return an array of top elements + */ + def top(num: Int, comp: Comparator[T]): JList[T] = { + import scala.collection.JavaConversions._ + val topElems = rdd.top(num)(Ordering.comparatorToOrdering(comp)) + val arr: java.util.Collection[T] = topElems.toSeq + new java.util.ArrayList(arr) + } + + /** + * Returns the top K elements from this RDD using the + * natural ordering for T. + * @param num the number of top elements to return + * @return an array of top elements + */ + def top(num: Int): JList[T] = { + val comp = com.google.common.collect.Ordering.natural().asInstanceOf[Comparator[T]] + top(num, comp) + } } diff --git a/core/src/main/scala/spark/api/java/JavaSparkContext.scala b/core/src/main/scala/spark/api/java/JavaSparkContext.scala index f75fc27c7b..5f18b1e15b 100644 --- a/core/src/main/scala/spark/api/java/JavaSparkContext.scala +++ b/core/src/main/scala/spark/api/java/JavaSparkContext.scala @@ -31,8 +31,8 @@ class JavaSparkContext(val sc: SparkContext) extends JavaSparkContextVarargsWork * @param master Cluster URL to connect to (e.g. mesos://host:port, spark://host:port, local[4]). * @param appName A name for your application, to display on the cluster web UI * @param sparkHome The SPARK_HOME directory on the slave nodes - * @param jars Collection of JARs to send to the cluster. These can be paths on the local file - * system or HDFS, HTTP, HTTPS, or FTP URLs. + * @param jarFile JAR file to send to the cluster. This can be a path on the local file system + * or an HDFS, HTTP, HTTPS, or FTP URL. */ def this(master: String, appName: String, sparkHome: String, jarFile: String) = this(new SparkContext(master, appName, sparkHome, Seq(jarFile))) diff --git a/core/src/main/scala/spark/api/java/PairFlatMapWorkaround.java b/core/src/main/scala/spark/api/java/PairFlatMapWorkaround.java deleted file mode 100644 index 68b6fd6622..0000000000 --- a/core/src/main/scala/spark/api/java/PairFlatMapWorkaround.java +++ /dev/null @@ -1,20 +0,0 @@ -package spark.api.java; - -import spark.api.java.JavaPairRDD; -import spark.api.java.JavaRDDLike; -import spark.api.java.function.PairFlatMapFunction; - -import java.io.Serializable; - -/** - * Workaround for SPARK-668. - */ -class PairFlatMapWorkaround<T> implements Serializable { - /** - * Return a new RDD by first applying a function to all elements of this - * RDD, and then flattening the results. - */ - public <K, V> JavaPairRDD<K, V> flatMap(PairFlatMapFunction<T, K, V> f) { - return ((JavaRDDLike <T, ?>) this).doFlatMap(f); - } -} diff --git a/core/src/main/scala/spark/api/java/function/FlatMapFunction2.scala b/core/src/main/scala/spark/api/java/function/FlatMapFunction2.scala new file mode 100644 index 0000000000..6044043add --- /dev/null +++ b/core/src/main/scala/spark/api/java/function/FlatMapFunction2.scala @@ -0,0 +1,11 @@ +package spark.api.java.function + +/** + * A function that takes two inputs and returns zero or more output records. + */ +abstract class FlatMapFunction2[A, B, C] extends Function2[A, B, java.lang.Iterable[C]] { + @throws(classOf[Exception]) + def call(a: A, b:B) : java.lang.Iterable[C] + + def elementType() : ClassManifest[C] = ClassManifest.Any.asInstanceOf[ClassManifest[C]] +} diff --git a/core/src/main/scala/spark/api/python/PythonRDD.scala b/core/src/main/scala/spark/api/python/PythonRDD.scala index 8c73477384..63140cf37f 100644 --- a/core/src/main/scala/spark/api/python/PythonRDD.scala +++ b/core/src/main/scala/spark/api/python/PythonRDD.scala @@ -2,10 +2,9 @@ package spark.api.python import java.io._ import java.net._ -import java.util.{List => JList, ArrayList => JArrayList, Collections} +import java.util.{List => JList, ArrayList => JArrayList, Map => JMap, Collections} import scala.collection.JavaConversions._ -import scala.io.Source import spark.api.java.{JavaSparkContext, JavaPairRDD, JavaRDD} import spark.broadcast.Broadcast @@ -16,7 +15,7 @@ import spark.rdd.PipedRDD private[spark] class PythonRDD[T: ClassManifest]( parent: RDD[T], command: Seq[String], - envVars: java.util.Map[String, String], + envVars: JMap[String, String], preservePartitoning: Boolean, pythonExec: String, broadcastVars: JList[Broadcast[Array[Byte]]], @@ -25,7 +24,7 @@ private[spark] class PythonRDD[T: ClassManifest]( // Similar to Runtime.exec(), if we are given a single string, split it into words // using a standard StringTokenizer (i.e. by spaces) - def this(parent: RDD[T], command: String, envVars: java.util.Map[String, String], + def this(parent: RDD[T], command: String, envVars: JMap[String, String], preservePartitoning: Boolean, pythonExec: String, broadcastVars: JList[Broadcast[Array[Byte]]], accumulator: Accumulator[JList[Array[Byte]]]) = @@ -36,35 +35,18 @@ private[spark] class PythonRDD[T: ClassManifest]( override val partitioner = if (preservePartitoning) parent.partitioner else None - override def compute(split: Partition, context: TaskContext): Iterator[Array[Byte]] = { - val SPARK_HOME = new ProcessBuilder().environment().get("SPARK_HOME") - - val pb = new ProcessBuilder(Seq(pythonExec, SPARK_HOME + "/python/pyspark/worker.py")) - // Add the environmental variables to the process. - val currentEnvVars = pb.environment() - - for ((variable, value) <- envVars) { - currentEnvVars.put(variable, value) - } - val proc = pb.start() + override def compute(split: Partition, context: TaskContext): Iterator[Array[Byte]] = { + val startTime = System.currentTimeMillis val env = SparkEnv.get - - // Start a thread to print the process's stderr to ours - new Thread("stderr reader for " + command) { - override def run() { - for (line <- Source.fromInputStream(proc.getErrorStream).getLines) { - System.err.println(line) - } - } - }.start() + val worker = env.createPythonWorker(pythonExec, envVars.toMap) // Start a thread to feed the process input from our parent's iterator - new Thread("stdin writer for " + command) { + new Thread("stdin writer for " + pythonExec) { override def run() { SparkEnv.set(env) - val out = new PrintWriter(proc.getOutputStream) - val dOut = new DataOutputStream(proc.getOutputStream) + val out = new PrintWriter(worker.getOutputStream) + val dOut = new DataOutputStream(worker.getOutputStream) // Partition index dOut.writeInt(split.index) // sparkFilesDir @@ -88,16 +70,21 @@ private[spark] class PythonRDD[T: ClassManifest]( } dOut.flush() out.flush() - proc.getOutputStream.close() + worker.shutdownOutput() } }.start() // Return an iterator that read lines from the process's stdout - val stream = new DataInputStream(proc.getInputStream) + val stream = new DataInputStream(worker.getInputStream) return new Iterator[Array[Byte]] { def next(): Array[Byte] = { val obj = _nextObj - _nextObj = read() + if (hasNext) { + // FIXME: can deadlock if worker is waiting for us to + // respond to current message (currently irrelevant because + // output is shutdown before we read any input) + _nextObj = read() + } obj } @@ -108,6 +95,17 @@ private[spark] class PythonRDD[T: ClassManifest]( val obj = new Array[Byte](length) stream.readFully(obj) obj + case -3 => + // Timing data from worker + val bootTime = stream.readLong() + val initTime = stream.readLong() + val finishTime = stream.readLong() + val boot = bootTime - startTime + val init = initTime - bootTime + val finish = finishTime - initTime + val total = finishTime - startTime + logInfo("Times: total = %s, boot = %s, init = %s, finish = %s".format(total, boot, init, finish)) + read case -2 => // Signals that an exception has been thrown in python val exLength = stream.readInt() @@ -115,23 +113,21 @@ private[spark] class PythonRDD[T: ClassManifest]( stream.readFully(obj) throw new PythonException(new String(obj)) case -1 => - // We've finished the data section of the output, but we can still read some - // accumulator updates; let's do that, breaking when we get EOFException - while (true) { - val len2 = stream.readInt() + // We've finished the data section of the output, but we can still + // read some accumulator updates; let's do that, breaking when we + // get a negative length record. + var len2 = stream.readInt() + while (len2 >= 0) { val update = new Array[Byte](len2) stream.readFully(update) accumulator += Collections.singletonList(update) + len2 = stream.readInt() } new Array[Byte](0) } } catch { case eof: EOFException => { - val exitStatus = proc.waitFor() - if (exitStatus != 0) { - throw new Exception("Subprocess exited with status " + exitStatus) - } - new Array[Byte](0) + throw new SparkException("Python worker exited unexpectedly (crashed)", eof) } case e => throw e } @@ -159,7 +155,7 @@ private class PairwiseRDD(prev: RDD[Array[Byte]]) extends override def compute(split: Partition, context: TaskContext) = prev.iterator(split, context).grouped(2).map { case Seq(a, b) => (a, b) - case x => throw new Exception("PairwiseRDD: unexpected value: " + x) + case x => throw new SparkException("PairwiseRDD: unexpected value: " + x) } val asJavaPairRDD : JavaPairRDD[Array[Byte], Array[Byte]] = JavaPairRDD.fromRDD(this) } @@ -215,7 +211,7 @@ private[spark] object PythonRDD { dOut.write(s) dOut.writeByte(Pickle.STOP) } else { - throw new Exception("Unexpected RDD type") + throw new SparkException("Unexpected RDD type") } } @@ -277,6 +273,8 @@ private class BytesToString extends spark.api.java.function.Function[Array[Byte] */ class PythonAccumulatorParam(@transient serverHost: String, serverPort: Int) extends AccumulatorParam[JList[Array[Byte]]] { + + Utils.checkHost(serverHost, "Expected hostname") override def zero(value: JList[Array[Byte]]): JList[Array[Byte]] = new JArrayList diff --git a/core/src/main/scala/spark/api/python/PythonWorkerFactory.scala b/core/src/main/scala/spark/api/python/PythonWorkerFactory.scala new file mode 100644 index 0000000000..8844411d73 --- /dev/null +++ b/core/src/main/scala/spark/api/python/PythonWorkerFactory.scala @@ -0,0 +1,95 @@ +package spark.api.python + +import java.io.{DataInputStream, IOException} +import java.net.{Socket, SocketException, InetAddress} + +import scala.collection.JavaConversions._ + +import spark._ + +private[spark] class PythonWorkerFactory(pythonExec: String, envVars: Map[String, String]) + extends Logging { + var daemon: Process = null + val daemonHost = InetAddress.getByAddress(Array(127, 0, 0, 1)) + var daemonPort: Int = 0 + + def create(): Socket = { + synchronized { + // Start the daemon if it hasn't been started + startDaemon() + + // Attempt to connect, restart and retry once if it fails + try { + new Socket(daemonHost, daemonPort) + } catch { + case exc: SocketException => { + logWarning("Python daemon unexpectedly quit, attempting to restart") + stopDaemon() + startDaemon() + new Socket(daemonHost, daemonPort) + } + case e => throw e + } + } + } + + def stop() { + stopDaemon() + } + + private def startDaemon() { + synchronized { + // Is it already running? + if (daemon != null) { + return + } + + try { + // Create and start the daemon + val sparkHome = new ProcessBuilder().environment().get("SPARK_HOME") + val pb = new ProcessBuilder(Seq(pythonExec, sparkHome + "/python/pyspark/daemon.py")) + val workerEnv = pb.environment() + workerEnv.putAll(envVars) + daemon = pb.start() + daemonPort = new DataInputStream(daemon.getInputStream).readInt() + + // Redirect the stderr to ours + new Thread("stderr reader for " + pythonExec) { + override def run() { + scala.util.control.Exception.ignoring(classOf[IOException]) { + // FIXME HACK: We copy the stream on the level of bytes to + // attempt to dodge encoding problems. + val in = daemon.getErrorStream + var buf = new Array[Byte](1024) + var len = in.read(buf) + while (len != -1) { + System.err.write(buf, 0, len) + len = in.read(buf) + } + } + } + }.start() + } catch { + case e => { + stopDaemon() + throw e + } + } + + // Important: don't close daemon's stdin (daemon.getOutputStream) so it can correctly + // detect our disappearance. + } + } + + private def stopDaemon() { + synchronized { + // Request shutdown of existing daemon by sending SIGTERM + if (daemon != null) { + daemon.destroy() + } + + daemon = null + daemonPort = 0 + } + } +} diff --git a/core/src/main/scala/spark/deploy/ApplicationDescription.scala b/core/src/main/scala/spark/deploy/ApplicationDescription.scala index 6659e53b25..02193c7008 100644 --- a/core/src/main/scala/spark/deploy/ApplicationDescription.scala +++ b/core/src/main/scala/spark/deploy/ApplicationDescription.scala @@ -2,10 +2,11 @@ package spark.deploy private[spark] class ApplicationDescription( val name: String, - val cores: Int, + val maxCores: Int, /* Integer.MAX_VALUE denotes an unlimited number of cores */ val memoryPerSlave: Int, val command: Command, - val sparkHome: String) + val sparkHome: String, + val appUiUrl: String) extends Serializable { val user = System.getProperty("user.name", "<unknown>") diff --git a/core/src/main/scala/spark/deploy/DeployMessage.scala b/core/src/main/scala/spark/deploy/DeployMessage.scala index 3cbf4fdd98..51274acb1e 100644 --- a/core/src/main/scala/spark/deploy/DeployMessage.scala +++ b/core/src/main/scala/spark/deploy/DeployMessage.scala @@ -4,6 +4,7 @@ import spark.deploy.ExecutorState.ExecutorState import spark.deploy.master.{WorkerInfo, ApplicationInfo} import spark.deploy.worker.ExecutorRunner import scala.collection.immutable.List +import spark.Utils private[spark] sealed trait DeployMessage extends Serializable @@ -19,7 +20,10 @@ case class RegisterWorker( memory: Int, webUiPort: Int, publicAddress: String) - extends DeployMessage + extends DeployMessage { + Utils.checkHost(host, "Required hostname") + assert (port > 0) +} private[spark] case class ExecutorStateChanged( @@ -58,14 +62,16 @@ private[spark] case class RegisteredApplication(appId: String) extends DeployMessage private[spark] -case class ExecutorAdded(id: Int, workerId: String, host: String, cores: Int, memory: Int) +case class ExecutorAdded(id: Int, workerId: String, hostPort: String, cores: Int, memory: Int) { + Utils.checkHostPort(hostPort, "Required hostport") +} private[spark] case class ExecutorUpdated(id: Int, state: ExecutorState, message: Option[String], exitStatus: Option[Int]) private[spark] -case class appKilled(message: String) +case class ApplicationRemoved(message: String) // Internal message in Client @@ -81,6 +87,9 @@ private[spark] case class MasterState(host: String, port: Int, workers: Array[WorkerInfo], activeApps: Array[ApplicationInfo], completedApps: Array[ApplicationInfo]) { + Utils.checkHost(host, "Required hostname") + assert (port > 0) + def uri = "spark://" + host + ":" + port } @@ -92,4 +101,8 @@ private[spark] case object RequestWorkerState private[spark] case class WorkerState(host: String, port: Int, workerId: String, executors: List[ExecutorRunner], finishedExecutors: List[ExecutorRunner], masterUrl: String, cores: Int, memory: Int, - coresUsed: Int, memoryUsed: Int, masterWebUiUrl: String) + coresUsed: Int, memoryUsed: Int, masterWebUiUrl: String) { + + Utils.checkHost(host, "Required hostname") + assert (port > 0) +} diff --git a/core/src/main/scala/spark/deploy/JsonProtocol.scala b/core/src/main/scala/spark/deploy/JsonProtocol.scala index 38a6ebfc24..ea832101d2 100644 --- a/core/src/main/scala/spark/deploy/JsonProtocol.scala +++ b/core/src/main/scala/spark/deploy/JsonProtocol.scala @@ -12,6 +12,7 @@ private[spark] object JsonProtocol extends DefaultJsonProtocol { def write(obj: WorkerInfo) = JsObject( "id" -> JsString(obj.id), "host" -> JsString(obj.host), + "port" -> JsNumber(obj.port), "webuiaddress" -> JsString(obj.webUiAddress), "cores" -> JsNumber(obj.cores), "coresused" -> JsNumber(obj.coresUsed), @@ -25,7 +26,7 @@ private[spark] object JsonProtocol extends DefaultJsonProtocol { "starttime" -> JsNumber(obj.startTime), "id" -> JsString(obj.id), "name" -> JsString(obj.desc.name), - "cores" -> JsNumber(obj.desc.cores), + "cores" -> JsNumber(obj.desc.maxCores), "user" -> JsString(obj.desc.user), "memoryperslave" -> JsNumber(obj.desc.memoryPerSlave), "submitdate" -> JsString(obj.submitDate.toString)) @@ -34,7 +35,7 @@ private[spark] object JsonProtocol extends DefaultJsonProtocol { implicit object AppDescriptionJsonFormat extends RootJsonWriter[ApplicationDescription] { def write(obj: ApplicationDescription) = JsObject( "name" -> JsString(obj.name), - "cores" -> JsNumber(obj.cores), + "cores" -> JsNumber(obj.maxCores), "memoryperslave" -> JsNumber(obj.memoryPerSlave), "user" -> JsString(obj.user) ) diff --git a/core/src/main/scala/spark/deploy/LocalSparkCluster.scala b/core/src/main/scala/spark/deploy/LocalSparkCluster.scala index 22319a96ca..55bb61b0cc 100644 --- a/core/src/main/scala/spark/deploy/LocalSparkCluster.scala +++ b/core/src/main/scala/spark/deploy/LocalSparkCluster.scala @@ -18,7 +18,7 @@ import scala.collection.mutable.ArrayBuffer private[spark] class LocalSparkCluster(numWorkers: Int, coresPerWorker: Int, memoryPerWorker: Int) extends Logging { - private val localIpAddress = Utils.localIpAddress + private val localHostname = Utils.localHostName() private val masterActorSystems = ArrayBuffer[ActorSystem]() private val workerActorSystems = ArrayBuffer[ActorSystem]() @@ -26,13 +26,13 @@ class LocalSparkCluster(numWorkers: Int, coresPerWorker: Int, memoryPerWorker: I logInfo("Starting a local Spark cluster with " + numWorkers + " workers.") /* Start the Master */ - val (masterSystem, masterPort) = Master.startSystemAndActor(localIpAddress, 0, 0) + val (masterSystem, masterPort) = Master.startSystemAndActor(localHostname, 0, 0) masterActorSystems += masterSystem - val masterUrl = "spark://" + localIpAddress + ":" + masterPort + val masterUrl = "spark://" + localHostname + ":" + masterPort /* Start the Workers */ for (workerNum <- 1 to numWorkers) { - val (workerSystem, _) = Worker.startSystemAndActor(localIpAddress, 0, 0, coresPerWorker, + val (workerSystem, _) = Worker.startSystemAndActor(localHostname, 0, 0, coresPerWorker, memoryPerWorker, masterUrl, null, Some(workerNum)) workerActorSystems += workerSystem } diff --git a/core/src/main/scala/spark/deploy/client/Client.scala b/core/src/main/scala/spark/deploy/client/Client.scala index 1a95524cf9..4af44f9c16 100644 --- a/core/src/main/scala/spark/deploy/client/Client.scala +++ b/core/src/main/scala/spark/deploy/client/Client.scala @@ -3,6 +3,7 @@ package spark.deploy.client import spark.deploy._ import akka.actor._ import akka.pattern.ask +import akka.util.Duration import akka.util.duration._ import akka.pattern.AskTimeoutException import spark.{SparkException, Logging} @@ -54,10 +55,15 @@ private[spark] class Client( appId = appId_ listener.connected(appId) - case ExecutorAdded(id: Int, workerId: String, host: String, cores: Int, memory: Int) => + case ApplicationRemoved(message) => + logError("Master removed our application: %s; stopping client".format(message)) + markDisconnected() + context.stop(self) + + case ExecutorAdded(id: Int, workerId: String, hostPort: String, cores: Int, memory: Int) => val fullId = appId + "/" + id - logInfo("Executor added: %s on %s (%s) with %d cores".format(fullId, workerId, host, cores)) - listener.executorAdded(fullId, workerId, host, cores, memory) + logInfo("Executor added: %s on %s (%s) with %d cores".format(fullId, workerId, hostPort, cores)) + listener.executorAdded(fullId, workerId, hostPort, cores, memory) case ExecutorUpdated(id, state, message, exitStatus) => val fullId = appId + "/" + id @@ -107,7 +113,7 @@ private[spark] class Client( def stop() { if (actor != null) { try { - val timeout = 5.seconds + val timeout = Duration.create(System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds") val future = actor.ask(StopClient)(timeout) Await.result(future, timeout) } catch { diff --git a/core/src/main/scala/spark/deploy/client/ClientListener.scala b/core/src/main/scala/spark/deploy/client/ClientListener.scala index b7008321df..e8c4083f9d 100644 --- a/core/src/main/scala/spark/deploy/client/ClientListener.scala +++ b/core/src/main/scala/spark/deploy/client/ClientListener.scala @@ -12,7 +12,7 @@ private[spark] trait ClientListener { def disconnected(): Unit - def executorAdded(fullId: String, workerId: String, host: String, cores: Int, memory: Int): Unit + def executorAdded(fullId: String, workerId: String, hostPort: String, cores: Int, memory: Int): Unit def executorRemoved(fullId: String, message: String, exitStatus: Option[Int]): Unit } diff --git a/core/src/main/scala/spark/deploy/client/TestClient.scala b/core/src/main/scala/spark/deploy/client/TestClient.scala index dc004b59ca..f195082808 100644 --- a/core/src/main/scala/spark/deploy/client/TestClient.scala +++ b/core/src/main/scala/spark/deploy/client/TestClient.scala @@ -16,7 +16,7 @@ private[spark] object TestClient { System.exit(0) } - def executorAdded(id: String, workerId: String, host: String, cores: Int, memory: Int) {} + def executorAdded(id: String, workerId: String, hostPort: String, cores: Int, memory: Int) {} def executorRemoved(id: String, message: String, exitStatus: Option[Int]) {} } @@ -25,7 +25,7 @@ private[spark] object TestClient { val url = args(0) val (actorSystem, port) = AkkaUtils.createActorSystem("spark", Utils.localIpAddress, 0) val desc = new ApplicationDescription( - "TestClient", 1, 512, Command("spark.deploy.client.TestExecutor", Seq(), Map()), "dummy-spark-home") + "TestClient", 1, 512, Command("spark.deploy.client.TestExecutor", Seq(), Map()), "dummy-spark-home", "ignored") val listener = new TestListener val client = new Client(actorSystem, url, desc, listener) client.start() diff --git a/core/src/main/scala/spark/deploy/master/ApplicationInfo.scala b/core/src/main/scala/spark/deploy/master/ApplicationInfo.scala index 3591a94072..785c16e2be 100644 --- a/core/src/main/scala/spark/deploy/master/ApplicationInfo.scala +++ b/core/src/main/scala/spark/deploy/master/ApplicationInfo.scala @@ -10,7 +10,8 @@ private[spark] class ApplicationInfo( val id: String, val desc: ApplicationDescription, val submitDate: Date, - val driver: ActorRef) + val driver: ActorRef, + val appUiUrl: String) { var state = ApplicationState.WAITING var executors = new mutable.HashMap[Int, ExecutorInfo] @@ -37,7 +38,7 @@ private[spark] class ApplicationInfo( coresGranted -= exec.cores } - def coresLeft: Int = desc.cores - coresGranted + def coresLeft: Int = desc.maxCores - coresGranted private var _retryCount = 0 @@ -60,4 +61,5 @@ private[spark] class ApplicationInfo( System.currentTimeMillis() - startTime } } + } diff --git a/core/src/main/scala/spark/deploy/master/Master.scala b/core/src/main/scala/spark/deploy/master/Master.scala index 1cd68a2aa6..3e965e82ac 100644 --- a/core/src/main/scala/spark/deploy/master/Master.scala +++ b/core/src/main/scala/spark/deploy/master/Master.scala @@ -15,7 +15,7 @@ import spark.{Logging, SparkException, Utils} import spark.util.AkkaUtils -private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor with Logging { +private[spark] class Master(host: String, port: Int, webUiPort: Int) extends Actor with Logging { val DATE_FORMAT = new SimpleDateFormat("yyyyMMddHHmmss") // For application IDs val WORKER_TIMEOUT = System.getProperty("spark.worker.timeout", "60").toLong * 1000 @@ -33,18 +33,22 @@ private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor val waitingApps = new ArrayBuffer[ApplicationInfo] val completedApps = new ArrayBuffer[ApplicationInfo] + var firstApp: Option[ApplicationInfo] = None + + Utils.checkHost(host, "Expected hostname") + val masterPublicAddress = { val envVar = System.getenv("SPARK_PUBLIC_DNS") - if (envVar != null) envVar else ip + if (envVar != null) envVar else host } // As a temporary workaround before better ways of configuring memory, we allow users to set // a flag that will perform round-robin scheduling across the nodes (spreading out each app // among all the nodes) instead of trying to consolidate each app onto a small # of nodes. - val spreadOutApps = System.getProperty("spark.deploy.spreadOut", "false").toBoolean + val spreadOutApps = System.getProperty("spark.deploy.spreadOut", "true").toBoolean override def preStart() { - logInfo("Starting Spark master at spark://" + ip + ":" + port) + logInfo("Starting Spark master at spark://" + host + ":" + port) // Listen for remote client disconnection events, since they don't go through Akka's watch() context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent]) startWebUi() @@ -105,7 +109,7 @@ private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor } else { logError("Application %s with ID %s failed %d times, removing it".format( appInfo.desc.name, appInfo.id, appInfo.retryCount)) - removeApplication(appInfo) + removeApplication(appInfo, ApplicationState.FAILED) } } } @@ -127,23 +131,23 @@ private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor // The disconnected actor could've been either a worker or an app; remove whichever of // those we have an entry for in the corresponding actor hashmap actorToWorker.get(actor).foreach(removeWorker) - actorToApp.get(actor).foreach(removeApplication) + actorToApp.get(actor).foreach(finishApplication) } case RemoteClientDisconnected(transport, address) => { // The disconnected client could've been either a worker or an app; remove whichever it was addressToWorker.get(address).foreach(removeWorker) - addressToApp.get(address).foreach(removeApplication) + addressToApp.get(address).foreach(finishApplication) } case RemoteClientShutdown(transport, address) => { // The disconnected client could've been either a worker or an app; remove whichever it was addressToWorker.get(address).foreach(removeWorker) - addressToApp.get(address).foreach(removeApplication) + addressToApp.get(address).foreach(finishApplication) } case RequestMasterState => { - sender ! MasterState(ip, port, workers.toArray, apps.toArray, completedApps.toArray) + sender ! MasterState(host, port, workers.toArray, apps.toArray, completedApps.toArray) } } @@ -167,7 +171,7 @@ private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor // Try to spread out each app among all the nodes, until it has all its cores for (app <- waitingApps if app.coresLeft > 0) { val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE) - .filter(canUse(app, _)).sortBy(_.coresFree).reverse + .filter(canUse(app, _)).sortBy(_.coresFree).reverse val numUsable = usableWorkers.length val assigned = new Array[Int](numUsable) // Number of cores to give on each node var toAssign = math.min(app.coresLeft, usableWorkers.map(_.coresFree).sum) @@ -190,7 +194,7 @@ private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor } } else { // Pack each app into as few nodes as possible until we've assigned all its cores - for (worker <- workers if worker.coresFree > 0) { + for (worker <- workers if worker.coresFree > 0 && worker.state == WorkerState.ALIVE) { for (app <- waitingApps if app.coresLeft > 0) { if (canUse(app, worker)) { val coresToUse = math.min(worker.coresFree, app.coresLeft) @@ -209,13 +213,13 @@ private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor logInfo("Launching executor " + exec.fullId + " on worker " + worker.id) worker.addExecutor(exec) worker.actor ! LaunchExecutor(exec.application.id, exec.id, exec.application.desc, exec.cores, exec.memory, sparkHome) - exec.application.driver ! ExecutorAdded(exec.id, worker.id, worker.host, exec.cores, exec.memory) + exec.application.driver ! ExecutorAdded(exec.id, worker.id, worker.hostPort, exec.cores, exec.memory) } def addWorker(id: String, host: String, port: Int, cores: Int, memory: Int, webUiPort: Int, publicAddress: String): WorkerInfo = { // There may be one or more refs to dead workers on this same node (w/ different ID's), remove them. - workers.filter(w => (w.host == host) && (w.state == WorkerState.DEAD)).foreach(workers -= _) + workers.filter(w => (w.host == host && w.port == port) && (w.state == WorkerState.DEAD)).foreach(workers -= _) val worker = new WorkerInfo(id, host, port, cores, memory, sender, webUiPort, publicAddress) workers += worker idToWorker(worker.id) = worker @@ -240,28 +244,41 @@ private[spark] class Master(ip: String, port: Int, webUiPort: Int) extends Actor def addApplication(desc: ApplicationDescription, driver: ActorRef): ApplicationInfo = { val now = System.currentTimeMillis() val date = new Date(now) - val app = new ApplicationInfo(now, newApplicationId(date), desc, date, driver) + val app = new ApplicationInfo(now, newApplicationId(date), desc, date, driver, desc.appUiUrl) apps += app idToApp(app.id) = app actorToApp(driver) = app addressToApp(driver.path.address) = app + if (firstApp == None) { + firstApp = Some(app) + } + val workersAlive = workers.filter(_.state == WorkerState.ALIVE).toArray + if (workersAlive.size > 0 && !workersAlive.exists(_.memoryFree >= desc.memoryPerSlave)) { + logWarning("Could not find any workers with enough memory for " + firstApp.get.id) + } return app } - def removeApplication(app: ApplicationInfo) { + def finishApplication(app: ApplicationInfo) { + removeApplication(app, ApplicationState.FINISHED) + } + + def removeApplication(app: ApplicationInfo, state: ApplicationState.Value) { if (apps.contains(app)) { logInfo("Removing app " + app.id) apps -= app idToApp -= app.id actorToApp -= app.driver - addressToWorker -= app.driver.path.address + addressToApp -= app.driver.path.address completedApps += app // Remember it in our history waitingApps -= app for (exec <- app.executors.values) { exec.worker.removeExecutor(exec) exec.worker.actor ! KillExecutor(exec.application.id, exec.id) + exec.state = ExecutorState.KILLED } - app.markFinished(ApplicationState.FINISHED) // TODO: Mark it as FAILED if it failed + app.markFinished(state) + app.driver ! ApplicationRemoved(state.toString) schedule() } } @@ -293,7 +310,7 @@ private[spark] object Master { def main(argStrings: Array[String]) { val args = new MasterArguments(argStrings) - val (actorSystem, _) = startSystemAndActor(args.ip, args.port, args.webUiPort) + val (actorSystem, _) = startSystemAndActor(args.host, args.port, args.webUiPort) actorSystem.awaitTermination() } diff --git a/core/src/main/scala/spark/deploy/master/MasterArguments.scala b/core/src/main/scala/spark/deploy/master/MasterArguments.scala index 4ceab3fc03..3d28ecabb4 100644 --- a/core/src/main/scala/spark/deploy/master/MasterArguments.scala +++ b/core/src/main/scala/spark/deploy/master/MasterArguments.scala @@ -7,13 +7,13 @@ import spark.Utils * Command-line parser for the master. */ private[spark] class MasterArguments(args: Array[String]) { - var ip = Utils.localHostName() + var host = Utils.localHostName() var port = 7077 var webUiPort = 8080 // Check for settings in environment variables - if (System.getenv("SPARK_MASTER_IP") != null) { - ip = System.getenv("SPARK_MASTER_IP") + if (System.getenv("SPARK_MASTER_HOST") != null) { + host = System.getenv("SPARK_MASTER_HOST") } if (System.getenv("SPARK_MASTER_PORT") != null) { port = System.getenv("SPARK_MASTER_PORT").toInt @@ -26,7 +26,13 @@ private[spark] class MasterArguments(args: Array[String]) { def parse(args: List[String]): Unit = args match { case ("--ip" | "-i") :: value :: tail => - ip = value + Utils.checkHost(value, "ip no longer supported, please use hostname " + value) + host = value + parse(tail) + + case ("--host" | "-h") :: value :: tail => + Utils.checkHost(value, "Please use hostname " + value) + host = value parse(tail) case ("--port" | "-p") :: IntParam(value) :: tail => @@ -54,7 +60,8 @@ private[spark] class MasterArguments(args: Array[String]) { "Usage: Master [options]\n" + "\n" + "Options:\n" + - " -i IP, --ip IP IP address or DNS name to listen on\n" + + " -i HOST, --ip HOST Hostname to listen on (deprecated, please use --host or -h) \n" + + " -h HOST, --host HOST Hostname to listen on\n" + " -p PORT, --port PORT Port to listen on (default: 7077)\n" + " --webui-port PORT Port for web UI (default: 8080)") System.exit(exitCode) diff --git a/core/src/main/scala/spark/deploy/master/MasterWebUI.scala b/core/src/main/scala/spark/deploy/master/MasterWebUI.scala index 54faa375fb..a4e21c8130 100644 --- a/core/src/main/scala/spark/deploy/master/MasterWebUI.scala +++ b/core/src/main/scala/spark/deploy/master/MasterWebUI.scala @@ -3,7 +3,7 @@ package spark.deploy.master import akka.actor.{ActorRef, ActorSystem} import akka.dispatch.Await import akka.pattern.ask -import akka.util.Timeout +import akka.util.{Duration, Timeout} import akka.util.duration._ import cc.spray.Directives import cc.spray.directives._ @@ -22,7 +22,7 @@ class MasterWebUI(val actorSystem: ActorSystem, master: ActorRef) extends Direct val RESOURCE_DIR = "spark/deploy/master/webui" val STATIC_RESOURCE_DIR = "spark/deploy/static" - implicit val timeout = Timeout(10 seconds) + implicit val timeout = Timeout(Duration.create(System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds")) val handler = { get { diff --git a/core/src/main/scala/spark/deploy/master/WorkerInfo.scala b/core/src/main/scala/spark/deploy/master/WorkerInfo.scala index 23df1bb463..0c08c5f417 100644 --- a/core/src/main/scala/spark/deploy/master/WorkerInfo.scala +++ b/core/src/main/scala/spark/deploy/master/WorkerInfo.scala @@ -2,6 +2,7 @@ package spark.deploy.master import akka.actor.ActorRef import scala.collection.mutable +import spark.Utils private[spark] class WorkerInfo( val id: String, @@ -13,6 +14,9 @@ private[spark] class WorkerInfo( val webUiPort: Int, val publicAddress: String) { + Utils.checkHost(host, "Expected hostname") + assert (port > 0) + var executors = new mutable.HashMap[String, ExecutorInfo] // fullId => info var state: WorkerState.Value = WorkerState.ALIVE var coresUsed = 0 @@ -23,6 +27,11 @@ private[spark] class WorkerInfo( def coresFree: Int = cores - coresUsed def memoryFree: Int = memory - memoryUsed + def hostPort: String = { + assert (port > 0) + host + ":" + port + } + def addExecutor(exec: ExecutorInfo) { executors(exec.fullId) = exec coresUsed += exec.cores diff --git a/core/src/main/scala/spark/deploy/worker/ExecutorRunner.scala b/core/src/main/scala/spark/deploy/worker/ExecutorRunner.scala index 38216ce62f..4d31657d9e 100644 --- a/core/src/main/scala/spark/deploy/worker/ExecutorRunner.scala +++ b/core/src/main/scala/spark/deploy/worker/ExecutorRunner.scala @@ -21,11 +21,13 @@ private[spark] class ExecutorRunner( val memory: Int, val worker: ActorRef, val workerId: String, - val hostname: String, + val hostPort: String, val sparkHome: File, val workDir: File) extends Logging { + Utils.checkHostPort(hostPort, "Expected hostport") + val fullId = appId + "/" + execId var workerThread: Thread = null var process: Process = null @@ -68,7 +70,7 @@ private[spark] class ExecutorRunner( /** Replace variables such as {{EXECUTOR_ID}} and {{CORES}} in a command argument passed to us */ def substituteVariables(argument: String): String = argument match { case "{{EXECUTOR_ID}}" => execId.toString - case "{{HOSTNAME}}" => hostname + case "{{HOSTNAME}}" => Utils.parseHostPort(hostPort)._1 case "{{CORES}}" => cores.toString case other => other } diff --git a/core/src/main/scala/spark/deploy/worker/Worker.scala b/core/src/main/scala/spark/deploy/worker/Worker.scala index 2bbc931316..3dc2207170 100644 --- a/core/src/main/scala/spark/deploy/worker/Worker.scala +++ b/core/src/main/scala/spark/deploy/worker/Worker.scala @@ -16,7 +16,7 @@ import spark.deploy.master.Master import java.io.File private[spark] class Worker( - ip: String, + host: String, port: Int, webUiPort: Int, cores: Int, @@ -25,6 +25,9 @@ private[spark] class Worker( workDirPath: String = null) extends Actor with Logging { + Utils.checkHost(host, "Expected hostname") + assert (port > 0) + val DATE_FORMAT = new SimpleDateFormat("yyyyMMddHHmmss") // For worker and executor IDs // Send a heartbeat every (heartbeat timeout) / 4 milliseconds @@ -39,7 +42,7 @@ private[spark] class Worker( val finishedExecutors = new HashMap[String, ExecutorRunner] val publicAddress = { val envVar = System.getenv("SPARK_PUBLIC_DNS") - if (envVar != null) envVar else ip + if (envVar != null) envVar else host } var coresUsed = 0 @@ -51,10 +54,14 @@ private[spark] class Worker( def createWorkDir() { workDir = Option(workDirPath).map(new File(_)).getOrElse(new File(sparkHome, "work")) try { - if (!workDir.exists() && !workDir.mkdirs()) { + // This sporadically fails - not sure why ... !workDir.exists() && !workDir.mkdirs() + // So attempting to create and then check if directory was created or not. + workDir.mkdirs() + if ( !workDir.exists() || !workDir.isDirectory) { logError("Failed to create work directory " + workDir) System.exit(1) } + assert (workDir.isDirectory) } catch { case e: Exception => logError("Failed to create work directory " + workDir, e) @@ -64,7 +71,7 @@ private[spark] class Worker( override def preStart() { logInfo("Starting Spark worker %s:%d with %d cores, %s RAM".format( - ip, port, cores, Utils.memoryMegabytesToString(memory))) + host, port, cores, Utils.memoryMegabytesToString(memory))) sparkHome = new File(Option(System.getenv("SPARK_HOME")).getOrElse(".")) logInfo("Spark home: " + sparkHome) createWorkDir() @@ -74,20 +81,14 @@ private[spark] class Worker( def connectToMaster() { logInfo("Connecting to master " + masterUrl) - try { - master = context.actorFor(Master.toAkkaUrl(masterUrl)) - master ! RegisterWorker(workerId, ip, port, cores, memory, webUiPort, publicAddress) - context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent]) - context.watch(master) // Doesn't work with remote actors, but useful for testing - } catch { - case e: Exception => - logError("Failed to connect to master", e) - System.exit(1) - } + master = context.actorFor(Master.toAkkaUrl(masterUrl)) + master ! RegisterWorker(workerId, host, port, cores, memory, webUiPort, publicAddress) + context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent]) + context.watch(master) // Doesn't work with remote actors, but useful for testing } def startWebUi() { - val webUi = new WorkerWebUI(context.system, self) + val webUi = new WorkerWebUI(context.system, self, workDir) try { AkkaUtils.startSprayServer(context.system, "0.0.0.0", webUiPort, webUi.handler) } catch { @@ -112,7 +113,7 @@ private[spark] class Worker( case LaunchExecutor(appId, execId, appDesc, cores_, memory_, execSparkHome_) => logInfo("Asked to launch executor %s/%d for %s".format(appId, execId, appDesc.name)) val manager = new ExecutorRunner( - appId, execId, appDesc, cores_, memory_, self, workerId, ip, new File(execSparkHome_), workDir) + appId, execId, appDesc, cores_, memory_, self, workerId, host + ":" + port, new File(execSparkHome_), workDir) executors(appId + "/" + execId) = manager manager.start() coresUsed += cores_ @@ -147,7 +148,7 @@ private[spark] class Worker( masterDisconnected() case RequestWorkerState => { - sender ! WorkerState(ip, port, workerId, executors.values.toList, + sender ! WorkerState(host, port, workerId, executors.values.toList, finishedExecutors.values.toList, masterUrl, cores, memory, coresUsed, memoryUsed, masterWebUiUrl) } @@ -162,7 +163,7 @@ private[spark] class Worker( } def generateWorkerId(): String = { - "worker-%s-%s-%d".format(DATE_FORMAT.format(new Date), ip, port) + "worker-%s-%s-%d".format(DATE_FORMAT.format(new Date), host, port) } override def postStop() { @@ -173,7 +174,7 @@ private[spark] class Worker( private[spark] object Worker { def main(argStrings: Array[String]) { val args = new WorkerArguments(argStrings) - val (actorSystem, _) = startSystemAndActor(args.ip, args.port, args.webUiPort, args.cores, + val (actorSystem, _) = startSystemAndActor(args.host, args.port, args.webUiPort, args.cores, args.memory, args.master, args.workDir) actorSystem.awaitTermination() } diff --git a/core/src/main/scala/spark/deploy/worker/WorkerArguments.scala b/core/src/main/scala/spark/deploy/worker/WorkerArguments.scala index 08f02bad80..2b96611ee3 100644 --- a/core/src/main/scala/spark/deploy/worker/WorkerArguments.scala +++ b/core/src/main/scala/spark/deploy/worker/WorkerArguments.scala @@ -9,7 +9,7 @@ import java.lang.management.ManagementFactory * Command-line parser for the master. */ private[spark] class WorkerArguments(args: Array[String]) { - var ip = Utils.localHostName() + var host = Utils.localHostName() var port = 0 var webUiPort = 8081 var cores = inferDefaultCores() @@ -38,7 +38,13 @@ private[spark] class WorkerArguments(args: Array[String]) { def parse(args: List[String]): Unit = args match { case ("--ip" | "-i") :: value :: tail => - ip = value + Utils.checkHost(value, "ip no longer supported, please use hostname " + value) + host = value + parse(tail) + + case ("--host" | "-h") :: value :: tail => + Utils.checkHost(value, "Please use hostname " + value) + host = value parse(tail) case ("--port" | "-p") :: IntParam(value) :: tail => @@ -93,7 +99,8 @@ private[spark] class WorkerArguments(args: Array[String]) { " -c CORES, --cores CORES Number of cores to use\n" + " -m MEM, --memory MEM Amount of memory to use (e.g. 1000M, 2G)\n" + " -d DIR, --work-dir DIR Directory to run apps in (default: SPARK_HOME/work)\n" + - " -i IP, --ip IP IP address or DNS name to listen on\n" + + " -i HOST, --ip IP Hostname to listen on (deprecated, please use --host or -h)\n" + + " -h HOST, --host HOST Hostname to listen on\n" + " -p PORT, --port PORT Port to listen on (default: random)\n" + " --webui-port PORT Port for web UI (default: 8081)") System.exit(exitCode) diff --git a/core/src/main/scala/spark/deploy/worker/WorkerWebUI.scala b/core/src/main/scala/spark/deploy/worker/WorkerWebUI.scala index 135cc2e86c..3235c50d1b 100644 --- a/core/src/main/scala/spark/deploy/worker/WorkerWebUI.scala +++ b/core/src/main/scala/spark/deploy/worker/WorkerWebUI.scala @@ -3,7 +3,7 @@ package spark.deploy.worker import akka.actor.{ActorRef, ActorSystem} import akka.dispatch.Await import akka.pattern.ask -import akka.util.Timeout +import akka.util.{Duration, Timeout} import akka.util.duration._ import cc.spray.Directives import cc.spray.typeconversion.TwirlSupport._ @@ -12,16 +12,17 @@ import cc.spray.typeconversion.SprayJsonSupport._ import spark.deploy.{WorkerState, RequestWorkerState} import spark.deploy.JsonProtocol._ +import java.io.File /** * Web UI server for the standalone worker. */ private[spark] -class WorkerWebUI(val actorSystem: ActorSystem, worker: ActorRef) extends Directives { +class WorkerWebUI(val actorSystem: ActorSystem, worker: ActorRef, workDir: File) extends Directives { val RESOURCE_DIR = "spark/deploy/worker/webui" val STATIC_RESOURCE_DIR = "spark/deploy/static" - implicit val timeout = Timeout(10 seconds) + implicit val timeout = Timeout(Duration.create(System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds")) val handler = { get { @@ -43,7 +44,7 @@ class WorkerWebUI(val actorSystem: ActorSystem, worker: ActorRef) extends Direct path("log") { parameters("appId", "executorId", "logType") { (appId, executorId, logType) => respondWithMediaType(cc.spray.http.MediaTypes.`text/plain`) { - getFromFileName("work/" + appId + "/" + executorId + "/" + logType) + getFromFileName(workDir.getPath() + "/" + appId + "/" + executorId + "/" + logType) } } } ~ diff --git a/core/src/main/scala/spark/executor/Executor.scala b/core/src/main/scala/spark/executor/Executor.scala index 5de09030aa..2bf55ea9a9 100644 --- a/core/src/main/scala/spark/executor/Executor.scala +++ b/core/src/main/scala/spark/executor/Executor.scala @@ -16,66 +16,68 @@ import java.nio.ByteBuffer /** * The Mesos executor for Spark. */ -private[spark] class Executor extends Logging { - var urlClassLoader : ExecutorURLClassLoader = null - var threadPool: ExecutorService = null - var env: SparkEnv = null +private[spark] class Executor(executorId: String, slaveHostname: String, properties: Seq[(String, String)]) extends Logging { // Application dependencies (added through SparkContext) that we've fetched so far on this node. // Each map holds the master's timestamp for the version of that file or JAR we got. - val currentFiles: HashMap[String, Long] = new HashMap[String, Long]() - val currentJars: HashMap[String, Long] = new HashMap[String, Long]() + private val currentFiles: HashMap[String, Long] = new HashMap[String, Long]() + private val currentJars: HashMap[String, Long] = new HashMap[String, Long]() - val EMPTY_BYTE_BUFFER = ByteBuffer.wrap(new Array[Byte](0)) + private val EMPTY_BYTE_BUFFER = ByteBuffer.wrap(new Array[Byte](0)) initLogging() - def initialize(executorId: String, slaveHostname: String, properties: Seq[(String, String)]) { - // Make sure the local hostname we report matches the cluster scheduler's name for this host - Utils.setCustomHostname(slaveHostname) + // No ip or host:port - just hostname + Utils.checkHost(slaveHostname, "Expected executed slave to be a hostname") + // must not have port specified. + assert (0 == Utils.parseHostPort(slaveHostname)._2) - // Set spark.* system properties from executor arg - for ((key, value) <- properties) { - System.setProperty(key, value) - } + // Make sure the local hostname we report matches the cluster scheduler's name for this host + Utils.setCustomHostname(slaveHostname) + + // Set spark.* system properties from executor arg + for ((key, value) <- properties) { + System.setProperty(key, value) + } - // Create our ClassLoader and set it on this thread - urlClassLoader = createClassLoader() - Thread.currentThread.setContextClassLoader(urlClassLoader) - - // Make any thread terminations due to uncaught exceptions kill the entire - // executor process to avoid surprising stalls. - Thread.setDefaultUncaughtExceptionHandler( - new Thread.UncaughtExceptionHandler { - override def uncaughtException(thread: Thread, exception: Throwable) { - try { - logError("Uncaught exception in thread " + thread, exception) - - // We may have been called from a shutdown hook. If so, we must not call System.exit(). - // (If we do, we will deadlock.) - if (!Utils.inShutdown()) { - if (exception.isInstanceOf[OutOfMemoryError]) { - System.exit(ExecutorExitCode.OOM) - } else { - System.exit(ExecutorExitCode.UNCAUGHT_EXCEPTION) - } + // Create our ClassLoader and set it on this thread + private val urlClassLoader = createClassLoader() + private val replClassLoader = addReplClassLoaderIfNeeded(urlClassLoader) + Thread.currentThread.setContextClassLoader(replClassLoader) + + // Make any thread terminations due to uncaught exceptions kill the entire + // executor process to avoid surprising stalls. + Thread.setDefaultUncaughtExceptionHandler( + new Thread.UncaughtExceptionHandler { + override def uncaughtException(thread: Thread, exception: Throwable) { + try { + logError("Uncaught exception in thread " + thread, exception) + + // We may have been called from a shutdown hook. If so, we must not call System.exit(). + // (If we do, we will deadlock.) + if (!Utils.inShutdown()) { + if (exception.isInstanceOf[OutOfMemoryError]) { + System.exit(ExecutorExitCode.OOM) + } else { + System.exit(ExecutorExitCode.UNCAUGHT_EXCEPTION) } - } catch { - case oom: OutOfMemoryError => Runtime.getRuntime.halt(ExecutorExitCode.OOM) - case t: Throwable => Runtime.getRuntime.halt(ExecutorExitCode.UNCAUGHT_EXCEPTION_TWICE) } + } catch { + case oom: OutOfMemoryError => Runtime.getRuntime.halt(ExecutorExitCode.OOM) + case t: Throwable => Runtime.getRuntime.halt(ExecutorExitCode.UNCAUGHT_EXCEPTION_TWICE) } } - ) + } + ) - // Initialize Spark environment (using system properties read above) - env = SparkEnv.createFromSystemProperties(executorId, slaveHostname, 0, false, false) - SparkEnv.set(env) + // Initialize Spark environment (using system properties read above) + val env = SparkEnv.createFromSystemProperties(executorId, slaveHostname, 0, false, false) + SparkEnv.set(env) + private val akkaFrameSize = env.actorSystem.settings.config.getBytes("akka.remote.netty.message-frame-size") - // Start worker thread pool - threadPool = new ThreadPoolExecutor( - 1, 128, 600, TimeUnit.SECONDS, new SynchronousQueue[Runnable]) - } + // Start worker thread pool + val threadPool = new ThreadPoolExecutor( + 1, 128, 600, TimeUnit.SECONDS, new SynchronousQueue[Runnable]) def launchTask(context: ExecutorBackend, taskId: Long, serializedTask: ByteBuffer) { threadPool.execute(new TaskRunner(context, taskId, serializedTask)) @@ -85,8 +87,9 @@ private[spark] class Executor extends Logging { extends Runnable { override def run() { + val startTime = System.currentTimeMillis() SparkEnv.set(env) - Thread.currentThread.setContextClassLoader(urlClassLoader) + Thread.currentThread.setContextClassLoader(replClassLoader) val ser = SparkEnv.get.closureSerializer.newInstance() logInfo("Running task ID " + taskId) context.statusUpdate(taskId, TaskState.RUNNING, EMPTY_BYTE_BUFFER) @@ -98,11 +101,25 @@ private[spark] class Executor extends Logging { val task = ser.deserialize[Task[Any]](taskBytes, Thread.currentThread.getContextClassLoader) logInfo("Its generation is " + task.generation) env.mapOutputTracker.updateGeneration(task.generation) + val taskStart = System.currentTimeMillis() val value = task.run(taskId.toInt) + val taskFinish = System.currentTimeMillis() + task.metrics.foreach{ m => + m.hostname = Utils.localHostName + m.executorDeserializeTime = (taskStart - startTime).toInt + m.executorRunTime = (taskFinish - taskStart).toInt + } + //TODO I'd also like to track the time it takes to serialize the task results, but that is huge headache, b/c + // we need to serialize the task metrics first. If TaskMetrics had a custom serialized format, we could + // just change the relevants bytes in the byte buffer val accumUpdates = Accumulators.values - val result = new TaskResult(value, accumUpdates) + val result = new TaskResult(value, accumUpdates, task.metrics.getOrElse(null)) val serializedResult = ser.serialize(result) logInfo("Serialized size of result for " + taskId + " is " + serializedResult.limit) + if (serializedResult.limit >= (akkaFrameSize - 1024)) { + context.statusUpdate(taskId, TaskState.FAILED, ser.serialize(TaskResultTooBigFailure())) + return + } context.statusUpdate(taskId, TaskState.FINISHED, serializedResult) logInfo("Finished task ID " + taskId) } catch { @@ -112,7 +129,7 @@ private[spark] class Executor extends Logging { } case t: Throwable => { - val reason = ExceptionFailure(t) + val reason = ExceptionFailure(t.getClass.getName, t.toString, t.getStackTrace) context.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason)) // TODO: Should we exit the whole executor here? On the one hand, the failed task may @@ -137,26 +154,31 @@ private[spark] class Executor extends Logging { val urls = currentJars.keySet.map { uri => new File(uri.split("/").last).toURI.toURL }.toArray - loader = new URLClassLoader(urls, loader) + new ExecutorURLClassLoader(urls, loader) + } - // If the REPL is in use, add another ClassLoader that will read - // new classes defined by the REPL as the user types code + /** + * If the REPL is in use, add another ClassLoader that will read + * new classes defined by the REPL as the user types code + */ + private def addReplClassLoaderIfNeeded(parent: ClassLoader): ClassLoader = { val classUri = System.getProperty("spark.repl.class.uri") if (classUri != null) { logInfo("Using REPL class URI: " + classUri) - loader = { - try { - val klass = Class.forName("spark.repl.ExecutorClassLoader") - .asInstanceOf[Class[_ <: ClassLoader]] - val constructor = klass.getConstructor(classOf[String], classOf[ClassLoader]) - constructor.newInstance(classUri, loader) - } catch { - case _: ClassNotFoundException => loader - } + try { + val klass = Class.forName("spark.repl.ExecutorClassLoader") + .asInstanceOf[Class[_ <: ClassLoader]] + val constructor = klass.getConstructor(classOf[String], classOf[ClassLoader]) + return constructor.newInstance(classUri, parent) + } catch { + case _: ClassNotFoundException => + logError("Could not find spark.repl.ExecutorClassLoader on classpath!") + System.exit(1) + null } + } else { + return parent } - - return new ExecutorURLClassLoader(Array(), loader) } /** diff --git a/core/src/main/scala/spark/executor/MesosExecutorBackend.scala b/core/src/main/scala/spark/executor/MesosExecutorBackend.scala index 818d6d1dda..10f3531df0 100644 --- a/core/src/main/scala/spark/executor/MesosExecutorBackend.scala +++ b/core/src/main/scala/spark/executor/MesosExecutorBackend.scala @@ -8,11 +8,12 @@ import com.google.protobuf.ByteString import spark.{Utils, Logging} import spark.TaskState -private[spark] class MesosExecutorBackend(executor: Executor) +private[spark] class MesosExecutorBackend extends MesosExecutor with ExecutorBackend with Logging { + var executor: Executor = null var driver: ExecutorDriver = null override def statusUpdate(taskId: Long, state: TaskState, data: ByteBuffer) { @@ -32,16 +33,19 @@ private[spark] class MesosExecutorBackend(executor: Executor) logInfo("Registered with Mesos as executor ID " + executorInfo.getExecutorId.getValue) this.driver = driver val properties = Utils.deserialize[Array[(String, String)]](executorInfo.getData.toByteArray) - executor.initialize( + executor = new Executor( executorInfo.getExecutorId.getValue, slaveInfo.getHostname, - properties - ) + properties) } override def launchTask(d: ExecutorDriver, taskInfo: TaskInfo) { val taskId = taskInfo.getTaskId.getValue.toLong - executor.launchTask(this, taskId, taskInfo.getData.asReadOnlyByteBuffer) + if (executor == null) { + logError("Received launchTask but executor was null") + } else { + executor.launchTask(this, taskId, taskInfo.getData.asReadOnlyByteBuffer) + } } override def error(d: ExecutorDriver, message: String) { @@ -68,7 +72,7 @@ private[spark] object MesosExecutorBackend { def main(args: Array[String]) { MesosNativeLibrary.load() // Create a new Executor and start it running - val runner = new MesosExecutorBackend(new Executor) + val runner = new MesosExecutorBackend() new MesosExecutorDriver(runner).run() } } diff --git a/core/src/main/scala/spark/executor/StandaloneExecutorBackend.scala b/core/src/main/scala/spark/executor/StandaloneExecutorBackend.scala index 9a82c3054c..ebe2ac68d8 100644 --- a/core/src/main/scala/spark/executor/StandaloneExecutorBackend.scala +++ b/core/src/main/scala/spark/executor/StandaloneExecutorBackend.scala @@ -12,23 +12,27 @@ import spark.scheduler.cluster.RegisteredExecutor import spark.scheduler.cluster.LaunchTask import spark.scheduler.cluster.RegisterExecutorFailed import spark.scheduler.cluster.RegisterExecutor +import spark.Utils +import spark.deploy.SparkHadoopUtil private[spark] class StandaloneExecutorBackend( - executor: Executor, driverUrl: String, executorId: String, - hostname: String, + hostPort: String, cores: Int) extends Actor with ExecutorBackend with Logging { + Utils.checkHostPort(hostPort, "Expected hostport") + + var executor: Executor = null var driver: ActorRef = null override def preStart() { logInfo("Connecting to driver: " + driverUrl) driver = context.actorFor(driverUrl) - driver ! RegisterExecutor(executorId, hostname, cores) + driver ! RegisterExecutor(executorId, hostPort, cores) context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent]) context.watch(driver) // Doesn't work with remote actors, but useful for testing } @@ -36,7 +40,8 @@ private[spark] class StandaloneExecutorBackend( override def receive = { case RegisteredExecutor(sparkProperties) => logInfo("Successfully registered with driver") - executor.initialize(executorId, hostname, sparkProperties) + // Make this host instead of hostPort ? + executor = new Executor(executorId, Utils.parseHostPort(hostPort)._1, sparkProperties) case RegisterExecutorFailed(message) => logError("Slave registration failed: " + message) @@ -44,7 +49,12 @@ private[spark] class StandaloneExecutorBackend( case LaunchTask(taskDesc) => logInfo("Got assigned task " + taskDesc.taskId) - executor.launchTask(this, taskDesc.taskId, taskDesc.serializedTask) + if (executor == null) { + logError("Received launchTask but executor was null") + System.exit(1) + } else { + executor.launchTask(this, taskDesc.taskId, taskDesc.serializedTask) + } case Terminated(_) | RemoteClientDisconnected(_, _) | RemoteClientShutdown(_, _) => logError("Driver terminated or disconnected! Shutting down.") @@ -58,11 +68,30 @@ private[spark] class StandaloneExecutorBackend( private[spark] object StandaloneExecutorBackend { def run(driverUrl: String, executorId: String, hostname: String, cores: Int) { + SparkHadoopUtil.runAsUser(run0, Tuple4[Any, Any, Any, Any] (driverUrl, executorId, hostname, cores)) + } + + // This will be run 'as' the user + def run0(args: Product) { + assert(4 == args.productArity) + runImpl(args.productElement(0).asInstanceOf[String], + args.productElement(1).asInstanceOf[String], + args.productElement(2).asInstanceOf[String], + args.productElement(3).asInstanceOf[Int]) + } + + private def runImpl(driverUrl: String, executorId: String, hostname: String, cores: Int) { + // Debug code + Utils.checkHost(hostname) + // Create a new ActorSystem to run the backend, because we can't create a SparkEnv / Executor // before getting started with all our system properties, etc val (actorSystem, boundPort) = AkkaUtils.createActorSystem("sparkExecutor", hostname, 0) + // set it + val sparkHostPort = hostname + ":" + boundPort + System.setProperty("spark.hostPort", sparkHostPort) val actor = actorSystem.actorOf( - Props(new StandaloneExecutorBackend(new Executor, driverUrl, executorId, hostname, cores)), + Props(new StandaloneExecutorBackend(driverUrl, executorId, sparkHostPort, cores)), name = "Executor") actorSystem.awaitTermination() } diff --git a/core/src/main/scala/spark/executor/TaskMetrics.scala b/core/src/main/scala/spark/executor/TaskMetrics.scala new file mode 100644 index 0000000000..1dc13754f9 --- /dev/null +++ b/core/src/main/scala/spark/executor/TaskMetrics.scala @@ -0,0 +1,83 @@ +package spark.executor + +class TaskMetrics extends Serializable { + /** + * Host's name the task runs on + */ + var hostname: String = _ + + /** + * Time taken on the executor to deserialize this task + */ + var executorDeserializeTime: Int = _ + + /** + * Time the executor spends actually running the task (including fetching shuffle data) + */ + var executorRunTime:Int = _ + + /** + * The number of bytes this task transmitted back to the driver as the TaskResult + */ + var resultSize: Long = _ + + /** + * If this task reads from shuffle output, metrics on getting shuffle data will be collected here + */ + var shuffleReadMetrics: Option[ShuffleReadMetrics] = None + + /** + * If this task writes to shuffle output, metrics on the written shuffle data will be collected here + */ + var shuffleWriteMetrics: Option[ShuffleWriteMetrics] = None +} + +object TaskMetrics { + private[spark] def empty(): TaskMetrics = new TaskMetrics +} + + +class ShuffleReadMetrics extends Serializable { + /** + * Time when shuffle finishs + */ + var shuffleFinishTime: Long = _ + + /** + * Total number of blocks fetched in a shuffle (remote or local) + */ + var totalBlocksFetched: Int = _ + + /** + * Number of remote blocks fetched in a shuffle + */ + var remoteBlocksFetched: Int = _ + + /** + * Local blocks fetched in a shuffle + */ + var localBlocksFetched: Int = _ + + /** + * Total time that is spent blocked waiting for shuffle to fetch data + */ + var fetchWaitTime: Long = _ + + /** + * The total amount of time for all the shuffle fetches. This adds up time from overlapping + * shuffles, so can be longer than task time + */ + var remoteFetchTime: Long = _ + + /** + * Total number of remote bytes read from a shuffle + */ + var remoteBytesRead: Long = _ +} + +class ShuffleWriteMetrics extends Serializable { + /** + * Number of bytes written for a shuffle + */ + var shuffleBytesWritten: Long = _ +} diff --git a/core/src/main/scala/spark/network/BufferMessage.scala b/core/src/main/scala/spark/network/BufferMessage.scala new file mode 100644 index 0000000000..7b0e489a6c --- /dev/null +++ b/core/src/main/scala/spark/network/BufferMessage.scala @@ -0,0 +1,94 @@ +package spark.network + +import java.nio.ByteBuffer + +import scala.collection.mutable.ArrayBuffer + +import spark.storage.BlockManager + + +private[spark] +class BufferMessage(id_ : Int, val buffers: ArrayBuffer[ByteBuffer], var ackId: Int) + extends Message(Message.BUFFER_MESSAGE, id_) { + + val initialSize = currentSize() + var gotChunkForSendingOnce = false + + def size = initialSize + + def currentSize() = { + if (buffers == null || buffers.isEmpty) { + 0 + } else { + buffers.map(_.remaining).reduceLeft(_ + _) + } + } + + def getChunkForSending(maxChunkSize: Int): Option[MessageChunk] = { + if (maxChunkSize <= 0) { + throw new Exception("Max chunk size is " + maxChunkSize) + } + + if (size == 0 && gotChunkForSendingOnce == false) { + val newChunk = new MessageChunk( + new MessageChunkHeader(typ, id, 0, 0, ackId, senderAddress), null) + gotChunkForSendingOnce = true + return Some(newChunk) + } + + while(!buffers.isEmpty) { + val buffer = buffers(0) + if (buffer.remaining == 0) { + BlockManager.dispose(buffer) + buffers -= buffer + } else { + val newBuffer = if (buffer.remaining <= maxChunkSize) { + buffer.duplicate() + } else { + buffer.slice().limit(maxChunkSize).asInstanceOf[ByteBuffer] + } + buffer.position(buffer.position + newBuffer.remaining) + val newChunk = new MessageChunk(new MessageChunkHeader( + typ, id, size, newBuffer.remaining, ackId, senderAddress), newBuffer) + gotChunkForSendingOnce = true + return Some(newChunk) + } + } + None + } + + def getChunkForReceiving(chunkSize: Int): Option[MessageChunk] = { + // STRONG ASSUMPTION: BufferMessage created when receiving data has ONLY ONE data buffer + if (buffers.size > 1) { + throw new Exception("Attempting to get chunk from message with multiple data buffers") + } + val buffer = buffers(0) + if (buffer.remaining > 0) { + if (buffer.remaining < chunkSize) { + throw new Exception("Not enough space in data buffer for receiving chunk") + } + val newBuffer = buffer.slice().limit(chunkSize).asInstanceOf[ByteBuffer] + buffer.position(buffer.position + newBuffer.remaining) + val newChunk = new MessageChunk(new MessageChunkHeader( + typ, id, size, newBuffer.remaining, ackId, senderAddress), newBuffer) + return Some(newChunk) + } + None + } + + def flip() { + buffers.foreach(_.flip) + } + + def hasAckId() = (ackId != 0) + + def isCompletelyReceived() = !buffers(0).hasRemaining + + override def toString = { + if (hasAckId) { + "BufferAckMessage(aid = " + ackId + ", id = " + id + ", size = " + size + ")" + } else { + "BufferMessage(id = " + id + ", size = " + size + ")" + } + } +}
\ No newline at end of file diff --git a/core/src/main/scala/spark/network/Connection.scala b/core/src/main/scala/spark/network/Connection.scala index cd5b7d57f3..6e28f677a3 100644 --- a/core/src/main/scala/spark/network/Connection.scala +++ b/core/src/main/scala/spark/network/Connection.scala @@ -13,12 +13,13 @@ import java.net._ private[spark] abstract class Connection(val channel: SocketChannel, val selector: Selector, - val remoteConnectionManagerId: ConnectionManagerId) extends Logging { + val socketRemoteConnectionManagerId: ConnectionManagerId) + extends Logging { + def this(channel_ : SocketChannel, selector_ : Selector) = { this(channel_, selector_, - ConnectionManagerId.fromSocketAddress( - channel_.socket.getRemoteSocketAddress().asInstanceOf[InetSocketAddress] - )) + ConnectionManagerId.fromSocketAddress( + channel_.socket.getRemoteSocketAddress().asInstanceOf[InetSocketAddress])) } channel.configureBlocking(false) @@ -33,16 +34,47 @@ abstract class Connection(val channel: SocketChannel, val selector: Selector, val remoteAddress = getRemoteAddress() + // Read channels typically do not register for write and write does not for read + // Now, we do have write registering for read too (temporarily), but this is to detect + // channel close NOT to actually read/consume data on it ! + // How does this work if/when we move to SSL ? + + // What is the interest to register with selector for when we want this connection to be selected + def registerInterest() + + // What is the interest to register with selector for when we want this connection to + // be de-selected + // Traditionally, 0 - but in our case, for example, for close-detection on SendingConnection hack, + // it will be SelectionKey.OP_READ (until we fix it properly) + def unregisterInterest() + + // On receiving a read event, should we change the interest for this channel or not ? + // Will be true for ReceivingConnection, false for SendingConnection. + def changeInterestForRead(): Boolean + + // On receiving a write event, should we change the interest for this channel or not ? + // Will be false for ReceivingConnection, true for SendingConnection. + // Actually, for now, should not get triggered for ReceivingConnection + def changeInterestForWrite(): Boolean + + def getRemoteConnectionManagerId(): ConnectionManagerId = { + socketRemoteConnectionManagerId + } + def key() = channel.keyFor(selector) def getRemoteAddress() = channel.socket.getRemoteSocketAddress().asInstanceOf[InetSocketAddress] - def read() { - throw new UnsupportedOperationException("Cannot read on connection of type " + this.getClass.toString) + // Returns whether we have to register for further reads or not. + def read(): Boolean = { + throw new UnsupportedOperationException( + "Cannot read on connection of type " + this.getClass.toString) } - - def write() { - throw new UnsupportedOperationException("Cannot write on connection of type " + this.getClass.toString) + + // Returns whether we have to register for further writes or not. + def write(): Boolean = { + throw new UnsupportedOperationException( + "Cannot write on connection of type " + this.getClass.toString) } def close() { @@ -54,26 +86,32 @@ abstract class Connection(val channel: SocketChannel, val selector: Selector, callOnCloseCallback() } - def onClose(callback: Connection => Unit) {onCloseCallback = callback} + def onClose(callback: Connection => Unit) { + onCloseCallback = callback + } - def onException(callback: (Connection, Exception) => Unit) {onExceptionCallback = callback} + def onException(callback: (Connection, Exception) => Unit) { + onExceptionCallback = callback + } - def onKeyInterestChange(callback: (Connection, Int) => Unit) {onKeyInterestChangeCallback = callback} + def onKeyInterestChange(callback: (Connection, Int) => Unit) { + onKeyInterestChangeCallback = callback + } def callOnExceptionCallback(e: Exception) { if (onExceptionCallback != null) { onExceptionCallback(this, e) } else { - logError("Error in connection to " + remoteConnectionManagerId + + logError("Error in connection to " + getRemoteConnectionManagerId() + " and OnExceptionCallback not registered", e) } } - + def callOnCloseCallback() { if (onCloseCallback != null) { onCloseCallback(this) } else { - logWarning("Connection to " + remoteConnectionManagerId + + logWarning("Connection to " + getRemoteConnectionManagerId() + " closed and OnExceptionCallback not registered") } @@ -81,7 +119,7 @@ abstract class Connection(val channel: SocketChannel, val selector: Selector, def changeConnectionKeyInterest(ops: Int) { if (onKeyInterestChangeCallback != null) { - onKeyInterestChangeCallback(this, ops) + onKeyInterestChangeCallback(this, ops) } else { throw new Exception("OnKeyInterestChangeCallback not registered") } @@ -105,24 +143,25 @@ abstract class Connection(val channel: SocketChannel, val selector: Selector, print(" (" + position + ", " + length + ")") buffer.position(curPosition) } - } -private[spark] class SendingConnection(val address: InetSocketAddress, selector_ : Selector, - remoteId_ : ConnectionManagerId) -extends Connection(SocketChannel.open, selector_, remoteId_) { +private[spark] +class SendingConnection(val address: InetSocketAddress, selector_ : Selector, + remoteId_ : ConnectionManagerId) + extends Connection(SocketChannel.open, selector_, remoteId_) { class Outbox(fair: Int = 0) { val messages = new Queue[Message]() - val defaultChunkSize = 65536 //32768 //16384 + val defaultChunkSize = 65536 //32768 //16384 var nextMessageToBeUsed = 0 def addMessage(message: Message) { - messages.synchronized{ + messages.synchronized{ /*messages += message*/ messages.enqueue(message) - logDebug("Added [" + message + "] to outbox for sending to [" + remoteConnectionManagerId + "]") + logDebug("Added [" + message + "] to outbox for sending to " + + "[" + getRemoteConnectionManagerId() + "]") } } @@ -147,18 +186,18 @@ extends Connection(SocketChannel.open, selector_, remoteId_) { message.started = true message.startTime = System.currentTimeMillis } - return chunk + return chunk } else { - /*logInfo("Finished sending [" + message + "] to [" + remoteConnectionManagerId + "]")*/ + /*logInfo("Finished sending [" + message + "] to [" + getRemoteConnectionManagerId() + "]")*/ message.finishTime = System.currentTimeMillis - logDebug("Finished sending [" + message + "] to [" + remoteConnectionManagerId + + logDebug("Finished sending [" + message + "] to [" + getRemoteConnectionManagerId() + "] in " + message.timeTaken ) } } } None } - + private def getChunkRR(): Option[MessageChunk] = { messages.synchronized { while (!messages.isEmpty) { @@ -170,15 +209,17 @@ extends Connection(SocketChannel.open, selector_, remoteId_) { messages.enqueue(message) nextMessageToBeUsed = nextMessageToBeUsed + 1 if (!message.started) { - logDebug("Starting to send [" + message + "] to [" + remoteConnectionManagerId + "]") + logDebug( + "Starting to send [" + message + "] to [" + getRemoteConnectionManagerId() + "]") message.started = true message.startTime = System.currentTimeMillis } - logTrace("Sending chunk from [" + message+ "] to [" + remoteConnectionManagerId + "]") - return chunk + logTrace( + "Sending chunk from [" + message+ "] to [" + getRemoteConnectionManagerId() + "]") + return chunk } else { message.finishTime = System.currentTimeMillis - logDebug("Finished sending [" + message + "] to [" + remoteConnectionManagerId + + logDebug("Finished sending [" + message + "] to [" + getRemoteConnectionManagerId() + "] in " + message.timeTaken ) } } @@ -186,27 +227,40 @@ extends Connection(SocketChannel.open, selector_, remoteId_) { None } } - - val outbox = new Outbox(1) + + private val outbox = new Outbox(1) val currentBuffers = new ArrayBuffer[ByteBuffer]() /*channel.socket.setSendBufferSize(256 * 1024)*/ - override def getRemoteAddress() = address + override def getRemoteAddress() = address + + val DEFAULT_INTEREST = SelectionKey.OP_READ + + override def registerInterest() { + // Registering read too - does not really help in most cases, but for some + // it does - so let us keep it for now. + changeConnectionKeyInterest(SelectionKey.OP_WRITE | DEFAULT_INTEREST) + } + + override def unregisterInterest() { + changeConnectionKeyInterest(DEFAULT_INTEREST) + } def send(message: Message) { outbox.synchronized { outbox.addMessage(message) if (channel.isConnected) { - changeConnectionKeyInterest(SelectionKey.OP_WRITE) + registerInterest() } } } + // MUST be called within the selector loop def connect() { try{ - channel.connect(address) channel.register(selector, SelectionKey.OP_CONNECT) + channel.connect(address) logInfo("Initiating connection to [" + address + "]") } catch { case e: Exception => { @@ -216,37 +270,52 @@ extends Connection(SocketChannel.open, selector_, remoteId_) { } } - def finishConnect() { + def finishConnect(force: Boolean): Boolean = { try { - channel.finishConnect - changeConnectionKeyInterest(SelectionKey.OP_WRITE) + // Typically, this should finish immediately since it was triggered by a connect + // selection - though need not necessarily always complete successfully. + val connected = channel.finishConnect + if (!force && !connected) { + logInfo( + "finish connect failed [" + address + "], " + outbox.messages.size + " messages pending") + return false + } + + // Fallback to previous behavior - assume finishConnect completed + // This will happen only when finishConnect failed for some repeated number of times + // (10 or so) + // Is highly unlikely unless there was an unclean close of socket, etc + registerInterest() logInfo("Connected to [" + address + "], " + outbox.messages.size + " messages pending") + return true } catch { case e: Exception => { logWarning("Error finishing connection to " + address, e) callOnExceptionCallback(e) + // ignore + return true } } } - override def write() { - try{ - while(true) { + override def write(): Boolean = { + try { + while (true) { if (currentBuffers.size == 0) { outbox.synchronized { outbox.getChunk() match { case Some(chunk) => { - currentBuffers ++= chunk.buffers + currentBuffers ++= chunk.buffers } case None => { - changeConnectionKeyInterest(0) + // changeConnectionKeyInterest(0) /*key.interestOps(0)*/ - return + return false } } } } - + if (currentBuffers.size > 0) { val buffer = currentBuffers(0) val remainingBytes = buffer.remaining @@ -255,52 +324,109 @@ extends Connection(SocketChannel.open, selector_, remoteId_) { currentBuffers -= buffer } if (writtenBytes < remainingBytes) { - return + // re-register for write. + return true } } } } catch { - case e: Exception => { - logWarning("Error writing in connection to " + remoteConnectionManagerId, e) + case e: Exception => { + logWarning("Error writing in connection to " + getRemoteConnectionManagerId(), e) callOnExceptionCallback(e) close() + return false } } + // should not happen - to keep scala compiler happy + return true } + + // This is a hack to determine if remote socket was closed or not. + // SendingConnection DOES NOT expect to receive any data - if it does, it is an error + // For a bunch of cases, read will return -1 in case remote socket is closed : hence we + // register for reads to determine that. + override def read(): Boolean = { + // We don't expect the other side to send anything; so, we just read to detect an error or EOF. + try { + val length = channel.read(ByteBuffer.allocate(1)) + if (length == -1) { // EOF + close() + } else if (length > 0) { + logWarning( + "Unexpected data read from SendingConnection to " + getRemoteConnectionManagerId()) + } + } catch { + case e: Exception => + logError("Exception while reading SendingConnection to " + getRemoteConnectionManagerId(), e) + callOnExceptionCallback(e) + close() + } + + false + } + + override def changeInterestForRead(): Boolean = false + + override def changeInterestForWrite(): Boolean = true } -private[spark] class ReceivingConnection(channel_ : SocketChannel, selector_ : Selector) -extends Connection(channel_, selector_) { - +// Must be created within selector loop - else deadlock +private[spark] class ReceivingConnection(channel_ : SocketChannel, selector_ : Selector) + extends Connection(channel_, selector_) { + class Inbox() { val messages = new HashMap[Int, BufferMessage]() - + def getChunk(header: MessageChunkHeader): Option[MessageChunk] = { - + def createNewMessage: BufferMessage = { val newMessage = Message.create(header).asInstanceOf[BufferMessage] newMessage.started = true newMessage.startTime = System.currentTimeMillis - logDebug("Starting to receive [" + newMessage + "] from [" + remoteConnectionManagerId + "]") + logDebug( + "Starting to receive [" + newMessage + "] from [" + getRemoteConnectionManagerId() + "]") messages += ((newMessage.id, newMessage)) newMessage } - + val message = messages.getOrElseUpdate(header.id, createNewMessage) - logTrace("Receiving chunk of [" + message + "] from [" + remoteConnectionManagerId + "]") + logTrace( + "Receiving chunk of [" + message + "] from [" + getRemoteConnectionManagerId() + "]") message.getChunkForReceiving(header.chunkSize) } - + def getMessageForChunk(chunk: MessageChunk): Option[BufferMessage] = { - messages.get(chunk.header.id) + messages.get(chunk.header.id) } def removeMessage(message: Message) { messages -= message.id } } - + + @volatile private var inferredRemoteManagerId: ConnectionManagerId = null + + override def getRemoteConnectionManagerId(): ConnectionManagerId = { + val currId = inferredRemoteManagerId + if (currId != null) currId else super.getRemoteConnectionManagerId() + } + + // The reciever's remote address is the local socket on remote side : which is NOT + // the connection manager id of the receiver. + // We infer that from the messages we receive on the receiver socket. + private def processConnectionManagerId(header: MessageChunkHeader) { + val currId = inferredRemoteManagerId + if (header.address == null || currId != null) return + + val managerId = ConnectionManagerId.fromSocketAddress(header.address) + + if (managerId != null) { + inferredRemoteManagerId = managerId + } + } + + val inbox = new Inbox() val headerBuffer: ByteBuffer = ByteBuffer.allocate(MessageChunkHeader.HEADER_SIZE) var onReceiveCallback: (Connection , Message) => Unit = null @@ -308,24 +434,29 @@ extends Connection(channel_, selector_) { channel.register(selector, SelectionKey.OP_READ) - override def read() { + override def read(): Boolean = { try { while (true) { if (currentChunk == null) { val headerBytesRead = channel.read(headerBuffer) if (headerBytesRead == -1) { close() - return + return false } if (headerBuffer.remaining > 0) { - return + // re-register for read event ... + return true } headerBuffer.flip if (headerBuffer.remaining != MessageChunkHeader.HEADER_SIZE) { - throw new Exception("Unexpected number of bytes (" + headerBuffer.remaining + ") in the header") + throw new Exception( + "Unexpected number of bytes (" + headerBuffer.remaining + ") in the header") } val header = MessageChunkHeader.create(headerBuffer) headerBuffer.clear() + + processConnectionManagerId(header) + header.typ match { case Message.BUFFER_MESSAGE => { if (header.totalSize == 0) { @@ -333,7 +464,8 @@ extends Connection(channel_, selector_) { onReceiveCallback(this, Message.create(header)) } currentChunk = null - return + // re-register for read event ... + return true } else { currentChunk = inbox.getChunk(header).orNull } @@ -341,26 +473,28 @@ extends Connection(channel_, selector_) { case _ => throw new Exception("Message of unknown type received") } } - + if (currentChunk == null) throw new Exception("No message chunk to receive data") - + val bytesRead = channel.read(currentChunk.buffer) if (bytesRead == 0) { - return + // re-register for read event ... + return true } else if (bytesRead == -1) { close() - return + return false } /*logDebug("Read " + bytesRead + " bytes for the buffer")*/ - + if (currentChunk.buffer.remaining == 0) { /*println("Filled buffer at " + System.currentTimeMillis)*/ val bufferMessage = inbox.getMessageForChunk(currentChunk).get if (bufferMessage.isCompletelyReceived) { bufferMessage.flip bufferMessage.finishTime = System.currentTimeMillis - logDebug("Finished receiving [" + bufferMessage + "] from [" + remoteConnectionManagerId + "] in " + bufferMessage.timeTaken) + logDebug("Finished receiving [" + bufferMessage + "] from " + + "[" + getRemoteConnectionManagerId() + "] in " + bufferMessage.timeTaken) if (onReceiveCallback != null) { onReceiveCallback(this, bufferMessage) } @@ -370,13 +504,32 @@ extends Connection(channel_, selector_) { } } } catch { - case e: Exception => { - logWarning("Error reading from connection to " + remoteConnectionManagerId, e) + case e: Exception => { + logWarning("Error reading from connection to " + getRemoteConnectionManagerId(), e) callOnExceptionCallback(e) close() + return false } } + // should not happen - to keep scala compiler happy + return true } - + def onReceive(callback: (Connection, Message) => Unit) {onReceiveCallback = callback} + + override def changeInterestForRead(): Boolean = true + + override def changeInterestForWrite(): Boolean = { + throw new IllegalStateException("Unexpected invocation right now") + } + + override def registerInterest() { + // Registering read too - does not really help in most cases, but for some + // it does - so let us keep it for now. + changeConnectionKeyInterest(SelectionKey.OP_READ) + } + + override def unregisterInterest() { + changeConnectionKeyInterest(0) + } } diff --git a/core/src/main/scala/spark/network/ConnectionManager.scala b/core/src/main/scala/spark/network/ConnectionManager.scala index b6ec664d7e..624a094856 100644 --- a/core/src/main/scala/spark/network/ConnectionManager.scala +++ b/core/src/main/scala/spark/network/ConnectionManager.scala @@ -6,28 +6,19 @@ import java.nio._ import java.nio.channels._ import java.nio.channels.spi._ import java.net._ -import java.util.concurrent.Executors +import java.util.concurrent.{LinkedBlockingDeque, TimeUnit, ThreadPoolExecutor} +import scala.collection.mutable.HashSet import scala.collection.mutable.HashMap import scala.collection.mutable.SynchronizedMap import scala.collection.mutable.SynchronizedQueue -import scala.collection.mutable.Queue import scala.collection.mutable.ArrayBuffer import akka.dispatch.{Await, Promise, ExecutionContext, Future} import akka.util.Duration import akka.util.duration._ -private[spark] case class ConnectionManagerId(host: String, port: Int) { - def toSocketAddress() = new InetSocketAddress(host, port) -} -private[spark] object ConnectionManagerId { - def fromSocketAddress(socketAddress: InetSocketAddress): ConnectionManagerId = { - new ConnectionManagerId(socketAddress.getHostName(), socketAddress.getPort()) - } -} - private[spark] class ConnectionManager(port: Int) extends Logging { class MessageStatus( @@ -41,73 +32,263 @@ private[spark] class ConnectionManager(port: Int) extends Logging { def markDone() { completionHandler(this) } } - - val selector = SelectorProvider.provider.openSelector() - val handleMessageExecutor = Executors.newFixedThreadPool(System.getProperty("spark.core.connection.handler.threads","20").toInt) - val serverChannel = ServerSocketChannel.open() - val connectionsByKey = new HashMap[SelectionKey, Connection] with SynchronizedMap[SelectionKey, Connection] - val connectionsById = new HashMap[ConnectionManagerId, SendingConnection] with SynchronizedMap[ConnectionManagerId, SendingConnection] - val messageStatuses = new HashMap[Int, MessageStatus] - val connectionRequests = new HashMap[ConnectionManagerId, SendingConnection] with SynchronizedMap[ConnectionManagerId, SendingConnection] - val keyInterestChangeRequests = new SynchronizedQueue[(SelectionKey, Int)] - val sendMessageRequests = new Queue[(Message, SendingConnection)] + + private val selector = SelectorProvider.provider.openSelector() + + private val handleMessageExecutor = new ThreadPoolExecutor( + System.getProperty("spark.core.connection.handler.threads.min","20").toInt, + System.getProperty("spark.core.connection.handler.threads.max","60").toInt, + System.getProperty("spark.core.connection.handler.threads.keepalive","60").toInt, TimeUnit.SECONDS, + new LinkedBlockingDeque[Runnable]()) + + private val handleReadWriteExecutor = new ThreadPoolExecutor( + System.getProperty("spark.core.connection.io.threads.min","4").toInt, + System.getProperty("spark.core.connection.io.threads.max","32").toInt, + System.getProperty("spark.core.connection.io.threads.keepalive","60").toInt, TimeUnit.SECONDS, + new LinkedBlockingDeque[Runnable]()) + + // Use a different, yet smaller, thread pool - infrequently used with very short lived tasks : which should be executed asap + private val handleConnectExecutor = new ThreadPoolExecutor( + System.getProperty("spark.core.connection.connect.threads.min","1").toInt, + System.getProperty("spark.core.connection.connect.threads.max","8").toInt, + System.getProperty("spark.core.connection.connect.threads.keepalive","60").toInt, TimeUnit.SECONDS, + new LinkedBlockingDeque[Runnable]()) + + private val serverChannel = ServerSocketChannel.open() + private val connectionsByKey = new HashMap[SelectionKey, Connection] with SynchronizedMap[SelectionKey, Connection] + private val connectionsById = new HashMap[ConnectionManagerId, SendingConnection] with SynchronizedMap[ConnectionManagerId, SendingConnection] + private val messageStatuses = new HashMap[Int, MessageStatus] + private val keyInterestChangeRequests = new SynchronizedQueue[(SelectionKey, Int)] + private val registerRequests = new SynchronizedQueue[SendingConnection] implicit val futureExecContext = ExecutionContext.fromExecutor(Utils.newDaemonCachedThreadPool()) - var onReceiveCallback: (BufferMessage, ConnectionManagerId) => Option[Message]= null + private var onReceiveCallback: (BufferMessage, ConnectionManagerId) => Option[Message]= null serverChannel.configureBlocking(false) serverChannel.socket.setReuseAddress(true) - serverChannel.socket.setReceiveBufferSize(256 * 1024) + serverChannel.socket.setReceiveBufferSize(256 * 1024) serverChannel.socket.bind(new InetSocketAddress(port)) serverChannel.register(selector, SelectionKey.OP_ACCEPT) val id = new ConnectionManagerId(Utils.localHostName, serverChannel.socket.getLocalPort) logInfo("Bound socket to port " + serverChannel.socket.getLocalPort() + " with id = " + id) - - val selectorThread = new Thread("connection-manager-thread") { + + private val selectorThread = new Thread("connection-manager-thread") { override def run() = ConnectionManager.this.run() } selectorThread.setDaemon(true) selectorThread.start() - private def run() { - try { - while(!selectorThread.isInterrupted) { - for ((connectionManagerId, sendingConnection) <- connectionRequests) { - sendingConnection.connect() - addConnection(sendingConnection) - connectionRequests -= connectionManagerId + private val writeRunnableStarted: HashSet[SelectionKey] = new HashSet[SelectionKey]() + + private def triggerWrite(key: SelectionKey) { + val conn = connectionsByKey.getOrElse(key, null) + if (conn == null) return + + writeRunnableStarted.synchronized { + // So that we do not trigger more write events while processing this one. + // The write method will re-register when done. + if (conn.changeInterestForWrite()) conn.unregisterInterest() + if (writeRunnableStarted.contains(key)) { + // key.interestOps(key.interestOps() & ~ SelectionKey.OP_WRITE) + return + } + + writeRunnableStarted += key + } + handleReadWriteExecutor.execute(new Runnable { + override def run() { + var register: Boolean = false + try { + register = conn.write() + } finally { + writeRunnableStarted.synchronized { + writeRunnableStarted -= key + if (register && conn.changeInterestForWrite()) { + conn.registerInterest() + } + } } - sendMessageRequests.synchronized { - while (!sendMessageRequests.isEmpty) { - val (message, connection) = sendMessageRequests.dequeue - connection.send(message) + } + } ) + } + + private val readRunnableStarted: HashSet[SelectionKey] = new HashSet[SelectionKey]() + + private def triggerRead(key: SelectionKey) { + val conn = connectionsByKey.getOrElse(key, null) + if (conn == null) return + + readRunnableStarted.synchronized { + // So that we do not trigger more read events while processing this one. + // The read method will re-register when done. + if (conn.changeInterestForRead())conn.unregisterInterest() + if (readRunnableStarted.contains(key)) { + return + } + + readRunnableStarted += key + } + handleReadWriteExecutor.execute(new Runnable { + override def run() { + var register: Boolean = false + try { + register = conn.read() + } finally { + readRunnableStarted.synchronized { + readRunnableStarted -= key + if (register && conn.changeInterestForRead()) { + conn.registerInterest() + } } } + } + } ) + } + + private def triggerConnect(key: SelectionKey) { + val conn = connectionsByKey.getOrElse(key, null).asInstanceOf[SendingConnection] + if (conn == null) return + + // prevent other events from being triggered + // Since we are still trying to connect, we do not need to do the additional steps in triggerWrite + conn.changeConnectionKeyInterest(0) + + handleConnectExecutor.execute(new Runnable { + override def run() { + + var tries: Int = 10 + while (tries >= 0) { + if (conn.finishConnect(false)) return + // Sleep ? + Thread.sleep(1) + tries -= 1 + } + + // fallback to previous behavior : we should not really come here since this method was + // triggered since channel became connectable : but at times, the first finishConnect need not + // succeed : hence the loop to retry a few 'times'. + conn.finishConnect(true) + } + } ) + } + + // MUST be called within selector loop - else deadlock. + private def triggerForceCloseByException(key: SelectionKey, e: Exception) { + try { + key.interestOps(0) + } catch { + // ignore exceptions + case e: Exception => logDebug("Ignoring exception", e) + } + + val conn = connectionsByKey.getOrElse(key, null) + if (conn == null) return + + // Pushing to connect threadpool + handleConnectExecutor.execute(new Runnable { + override def run() { + try { + conn.callOnExceptionCallback(e) + } catch { + // ignore exceptions + case e: Exception => logDebug("Ignoring exception", e) + } + try { + conn.close() + } catch { + // ignore exceptions + case e: Exception => logDebug("Ignoring exception", e) + } + } + }) + } + - while (!keyInterestChangeRequests.isEmpty) { + def run() { + try { + while(!selectorThread.isInterrupted) { + while (! registerRequests.isEmpty) { + val conn: SendingConnection = registerRequests.dequeue + addListeners(conn) + conn.connect() + addConnection(conn) + } + + while(!keyInterestChangeRequests.isEmpty) { val (key, ops) = keyInterestChangeRequests.dequeue - val connection = connectionsByKey(key) - val lastOps = key.interestOps() - key.interestOps(ops) - - def intToOpStr(op: Int): String = { - val opStrs = ArrayBuffer[String]() - if ((op & SelectionKey.OP_READ) != 0) opStrs += "READ" - if ((op & SelectionKey.OP_WRITE) != 0) opStrs += "WRITE" - if ((op & SelectionKey.OP_CONNECT) != 0) opStrs += "CONNECT" - if ((op & SelectionKey.OP_ACCEPT) != 0) opStrs += "ACCEPT" - if (opStrs.size > 0) opStrs.reduceLeft(_ + " | " + _) else " " + + try { + if (key.isValid) { + val connection = connectionsByKey.getOrElse(key, null) + if (connection != null) { + val lastOps = key.interestOps() + key.interestOps(ops) + + // hot loop - prevent materialization of string if trace not enabled. + if (isTraceEnabled()) { + def intToOpStr(op: Int): String = { + val opStrs = ArrayBuffer[String]() + if ((op & SelectionKey.OP_READ) != 0) opStrs += "READ" + if ((op & SelectionKey.OP_WRITE) != 0) opStrs += "WRITE" + if ((op & SelectionKey.OP_CONNECT) != 0) opStrs += "CONNECT" + if ((op & SelectionKey.OP_ACCEPT) != 0) opStrs += "ACCEPT" + if (opStrs.size > 0) opStrs.reduceLeft(_ + " | " + _) else " " + } + + logTrace("Changed key for connection to [" + connection.getRemoteConnectionManagerId() + + "] changed from [" + intToOpStr(lastOps) + "] to [" + intToOpStr(ops) + "]") + } + } + } else { + logInfo("Key not valid ? " + key) + throw new CancelledKeyException() + } + } catch { + case e: CancelledKeyException => { + logInfo("key already cancelled ? " + key, e) + triggerForceCloseByException(key, e) + } + case e: Exception => { + logError("Exception processing key " + key, e) + triggerForceCloseByException(key, e) + } } - - logTrace("Changed key for connection to [" + connection.remoteConnectionManagerId + - "] changed from [" + intToOpStr(lastOps) + "] to [" + intToOpStr(ops) + "]") - } - val selectedKeysCount = selector.select() + val selectedKeysCount = + try { + selector.select() + } catch { + // Explicitly only dealing with CancelledKeyException here since other exceptions should be dealt with differently. + case e: CancelledKeyException => { + // Some keys within the selectors list are invalid/closed. clear them. + val allKeys = selector.keys().iterator() + + while (allKeys.hasNext()) { + val key = allKeys.next() + try { + if (! key.isValid) { + logInfo("Key not valid ? " + key) + throw new CancelledKeyException() + } + } catch { + case e: CancelledKeyException => { + logInfo("key already cancelled ? " + key, e) + triggerForceCloseByException(key, e) + } + case e: Exception => { + logError("Exception processing key " + key, e) + triggerForceCloseByException(key, e) + } + } + } + } + 0 + } + if (selectedKeysCount == 0) { logDebug("Selector selected " + selectedKeysCount + " of " + selector.keys.size + " keys") } @@ -115,20 +296,40 @@ private[spark] class ConnectionManager(port: Int) extends Logging { logInfo("Selector thread was interrupted!") return } - - val selectedKeys = selector.selectedKeys().iterator() - while (selectedKeys.hasNext()) { - val key = selectedKeys.next - selectedKeys.remove() - if (key.isValid) { - if (key.isAcceptable) { - acceptConnection(key) - } else if (key.isConnectable) { - connectionsByKey(key).asInstanceOf[SendingConnection].finishConnect() - } else if (key.isReadable) { - connectionsByKey(key).read() - } else if (key.isWritable) { - connectionsByKey(key).write() + + if (0 != selectedKeysCount) { + val selectedKeys = selector.selectedKeys().iterator() + while (selectedKeys.hasNext()) { + val key = selectedKeys.next + selectedKeys.remove() + try { + if (key.isValid) { + if (key.isAcceptable) { + acceptConnection(key) + } else + if (key.isConnectable) { + triggerConnect(key) + } else + if (key.isReadable) { + triggerRead(key) + } else + if (key.isWritable) { + triggerWrite(key) + } + } else { + logInfo("Key not valid ? " + key) + throw new CancelledKeyException() + } + } catch { + // weird, but we saw this happening - even though key.isValid was true, key.isAcceptable would throw CancelledKeyException. + case e: CancelledKeyException => { + logInfo("key already cancelled ? " + key, e) + triggerForceCloseByException(key, e) + } + case e: Exception => { + logError("Exception processing key " + key, e) + triggerForceCloseByException(key, e) + } } } } @@ -137,97 +338,119 @@ private[spark] class ConnectionManager(port: Int) extends Logging { case e: Exception => logError("Error in select loop", e) } } - - private def acceptConnection(key: SelectionKey) { + + def acceptConnection(key: SelectionKey) { val serverChannel = key.channel.asInstanceOf[ServerSocketChannel] - val newChannel = serverChannel.accept() - val newConnection = new ReceivingConnection(newChannel, selector) - newConnection.onReceive(receiveMessage) - newConnection.onClose(removeConnection) - addConnection(newConnection) - logInfo("Accepted connection from [" + newConnection.remoteAddress.getAddress + "]") - } - private def addConnection(connection: Connection) { - connectionsByKey += ((connection.key, connection)) - if (connection.isInstanceOf[SendingConnection]) { - val sendingConnection = connection.asInstanceOf[SendingConnection] - connectionsById += ((sendingConnection.remoteConnectionManagerId, sendingConnection)) + var newChannel = serverChannel.accept() + + // accept them all in a tight loop. non blocking accept with no processing, should be fine + while (newChannel != null) { + try { + val newConnection = new ReceivingConnection(newChannel, selector) + newConnection.onReceive(receiveMessage) + addListeners(newConnection) + addConnection(newConnection) + logInfo("Accepted connection from [" + newConnection.remoteAddress.getAddress + "]") + } catch { + // might happen in case of issues with registering with selector + case e: Exception => logError("Error in accept loop", e) + } + + newChannel = serverChannel.accept() } + } + + private def addListeners(connection: Connection) { connection.onKeyInterestChange(changeConnectionKeyInterest) connection.onException(handleConnectionError) connection.onClose(removeConnection) } - private def removeConnection(connection: Connection) { + def addConnection(connection: Connection) { + connectionsByKey += ((connection.key, connection)) + } + + def removeConnection(connection: Connection) { connectionsByKey -= connection.key - if (connection.isInstanceOf[SendingConnection]) { - val sendingConnection = connection.asInstanceOf[SendingConnection] - val sendingConnectionManagerId = sendingConnection.remoteConnectionManagerId - logInfo("Removing SendingConnection to " + sendingConnectionManagerId) - - connectionsById -= sendingConnectionManagerId - - messageStatuses.synchronized { - messageStatuses - .values.filter(_.connectionManagerId == sendingConnectionManagerId).foreach(status => { - logInfo("Notifying " + status) - status.synchronized { - status.attempted = true - status.acked = false - status.markDone() - } + + try { + if (connection.isInstanceOf[SendingConnection]) { + val sendingConnection = connection.asInstanceOf[SendingConnection] + val sendingConnectionManagerId = sendingConnection.getRemoteConnectionManagerId() + logInfo("Removing SendingConnection to " + sendingConnectionManagerId) + + connectionsById -= sendingConnectionManagerId + + messageStatuses.synchronized { + messageStatuses + .values.filter(_.connectionManagerId == sendingConnectionManagerId).foreach(status => { + logInfo("Notifying " + status) + status.synchronized { + status.attempted = true + status.acked = false + status.markDone() + } + }) + + messageStatuses.retain((i, status) => { + status.connectionManagerId != sendingConnectionManagerId }) + } + } else if (connection.isInstanceOf[ReceivingConnection]) { + val receivingConnection = connection.asInstanceOf[ReceivingConnection] + val remoteConnectionManagerId = receivingConnection.getRemoteConnectionManagerId() + logInfo("Removing ReceivingConnection to " + remoteConnectionManagerId) + + val sendingConnectionOpt = connectionsById.get(remoteConnectionManagerId) + if (! sendingConnectionOpt.isDefined) { + logError("Corresponding SendingConnectionManagerId not found") + return + } - messageStatuses.retain((i, status) => { - status.connectionManagerId != sendingConnectionManagerId - }) - } - } else if (connection.isInstanceOf[ReceivingConnection]) { - val receivingConnection = connection.asInstanceOf[ReceivingConnection] - val remoteConnectionManagerId = receivingConnection.remoteConnectionManagerId - logInfo("Removing ReceivingConnection to " + remoteConnectionManagerId) - - val sendingConnectionManagerId = connectionsById.keys.find(_.host == remoteConnectionManagerId.host).orNull - if (sendingConnectionManagerId == null) { - logError("Corresponding SendingConnectionManagerId not found") - return - } - logInfo("Corresponding SendingConnectionManagerId is " + sendingConnectionManagerId) - - val sendingConnection = connectionsById(sendingConnectionManagerId) - sendingConnection.close() - connectionsById -= sendingConnectionManagerId - - messageStatuses.synchronized { - for (s <- messageStatuses.values if s.connectionManagerId == sendingConnectionManagerId) { - logInfo("Notifying " + s) - s.synchronized { - s.attempted = true - s.acked = false - s.markDone() + val sendingConnection = sendingConnectionOpt.get + connectionsById -= remoteConnectionManagerId + sendingConnection.close() + + val sendingConnectionManagerId = sendingConnection.getRemoteConnectionManagerId() + + assert (sendingConnectionManagerId == remoteConnectionManagerId) + + messageStatuses.synchronized { + for (s <- messageStatuses.values if s.connectionManagerId == sendingConnectionManagerId) { + logInfo("Notifying " + s) + s.synchronized { + s.attempted = true + s.acked = false + s.markDone() + } } - } - messageStatuses.retain((i, status) => { - status.connectionManagerId != sendingConnectionManagerId - }) + messageStatuses.retain((i, status) => { + status.connectionManagerId != sendingConnectionManagerId + }) + } } + } finally { + // So that the selection keys can be removed. + wakeupSelector() } } - private def handleConnectionError(connection: Connection, e: Exception) { - logInfo("Handling connection error on connection to " + connection.remoteConnectionManagerId) + def handleConnectionError(connection: Connection, e: Exception) { + logInfo("Handling connection error on connection to " + connection.getRemoteConnectionManagerId()) removeConnection(connection) } - private def changeConnectionKeyInterest(connection: Connection, ops: Int) { - keyInterestChangeRequests += ((connection.key, ops)) + def changeConnectionKeyInterest(connection: Connection, ops: Int) { + keyInterestChangeRequests += ((connection.key, ops)) + // so that registerations happen ! + wakeupSelector() } - private def receiveMessage(connection: Connection, message: Message) { + def receiveMessage(connection: Connection, message: Message) { val connectionManagerId = ConnectionManagerId.fromSocketAddress(message.senderAddress) - logDebug("Received [" + message + "] from [" + connectionManagerId + "]") + logDebug("Received [" + message + "] from [" + connectionManagerId + "]") val runnable = new Runnable() { val creationTime = System.currentTimeMillis def run() { @@ -247,11 +470,11 @@ private[spark] class ConnectionManager(port: Int) extends Logging { if (bufferMessage.hasAckId) { val sentMessageStatus = messageStatuses.synchronized { messageStatuses.get(bufferMessage.ackId) match { - case Some(status) => { - messageStatuses -= bufferMessage.ackId + case Some(status) => { + messageStatuses -= bufferMessage.ackId status } - case None => { + case None => { throw new Exception("Could not find reference for received ack message " + message.id) null } @@ -271,7 +494,7 @@ private[spark] class ConnectionManager(port: Int) extends Logging { logDebug("Not calling back as callback is null") None } - + if (ackMessage.isDefined) { if (!ackMessage.get.isInstanceOf[BufferMessage]) { logDebug("Response to " + bufferMessage + " is not a buffer message, it is of type " + ackMessage.get.getClass()) @@ -281,7 +504,7 @@ private[spark] class ConnectionManager(port: Int) extends Logging { } } - sendMessage(connectionManagerId, ackMessage.getOrElse { + sendMessage(connectionManagerId, ackMessage.getOrElse { Message.createBufferMessage(bufferMessage.id) }) } @@ -293,18 +516,22 @@ private[spark] class ConnectionManager(port: Int) extends Logging { private def sendMessage(connectionManagerId: ConnectionManagerId, message: Message) { def startNewConnection(): SendingConnection = { val inetSocketAddress = new InetSocketAddress(connectionManagerId.host, connectionManagerId.port) - val newConnection = connectionRequests.getOrElseUpdate(connectionManagerId, - new SendingConnection(inetSocketAddress, selector, connectionManagerId)) - newConnection + val newConnection = new SendingConnection(inetSocketAddress, selector, connectionManagerId) + registerRequests.enqueue(newConnection) + + newConnection } - val lookupKey = ConnectionManagerId.fromSocketAddress(connectionManagerId.toSocketAddress) - val connection = connectionsById.getOrElse(lookupKey, startNewConnection()) + // I removed the lookupKey stuff as part of merge ... should I re-add it ? We did not find it useful in our test-env ... + // If we do re-add it, we should consistently use it everywhere I guess ? + val connection = connectionsById.getOrElseUpdate(connectionManagerId, startNewConnection()) message.senderAddress = id.toSocketAddress() logDebug("Sending [" + message + "] to [" + connectionManagerId + "]") - /*connection.send(message)*/ - sendMessageRequests.synchronized { - sendMessageRequests += ((message, connection)) - } + connection.send(message) + + wakeupSelector() + } + + private def wakeupSelector() { selector.wakeup() } @@ -337,6 +564,8 @@ private[spark] class ConnectionManager(port: Int) extends Logging { logWarning("All connections not cleaned up") } handleMessageExecutor.shutdown() + handleReadWriteExecutor.shutdown() + handleConnectExecutor.shutdown() logInfo("ConnectionManager stopped") } } @@ -346,17 +575,17 @@ private[spark] object ConnectionManager { def main(args: Array[String]) { val manager = new ConnectionManager(9999) - manager.onReceiveMessage((msg: Message, id: ConnectionManagerId) => { + manager.onReceiveMessage((msg: Message, id: ConnectionManagerId) => { println("Received [" + msg + "] from [" + id + "]") None }) - + /*testSequentialSending(manager)*/ /*System.gc()*/ /*testParallelSending(manager)*/ /*System.gc()*/ - + /*testParallelDecreasingSending(manager)*/ /*System.gc()*/ @@ -368,9 +597,9 @@ private[spark] object ConnectionManager { println("--------------------------") println("Sequential Sending") println("--------------------------") - val size = 10 * 1024 * 1024 + val size = 10 * 1024 * 1024 val count = 10 - + val buffer = ByteBuffer.allocate(size).put(Array.tabulate[Byte](size)(x => x.toByte)) buffer.flip @@ -386,7 +615,7 @@ private[spark] object ConnectionManager { println("--------------------------") println("Parallel Sending") println("--------------------------") - val size = 10 * 1024 * 1024 + val size = 10 * 1024 * 1024 val count = 10 val buffer = ByteBuffer.allocate(size).put(Array.tabulate[Byte](size)(x => x.toByte)) @@ -401,12 +630,12 @@ private[spark] object ConnectionManager { if (!g.isDefined) println("Failed") }) val finishTime = System.currentTimeMillis - + val mb = size * count / 1024.0 / 1024.0 val ms = finishTime - startTime val tput = mb * 1000.0 / ms println("--------------------------") - println("Started at " + startTime + ", finished at " + finishTime) + println("Started at " + startTime + ", finished at " + finishTime) println("Sent " + count + " messages of size " + size + " in " + ms + " ms (" + tput + " MB/s)") println("--------------------------") println() @@ -416,7 +645,7 @@ private[spark] object ConnectionManager { println("--------------------------") println("Parallel Decreasing Sending") println("--------------------------") - val size = 10 * 1024 * 1024 + val size = 10 * 1024 * 1024 val count = 10 val buffers = Array.tabulate(count)(i => ByteBuffer.allocate(size * (i + 1)).put(Array.tabulate[Byte](size * (i + 1))(x => x.toByte))) buffers.foreach(_.flip) @@ -431,7 +660,7 @@ private[spark] object ConnectionManager { if (!g.isDefined) println("Failed") }) val finishTime = System.currentTimeMillis - + val ms = finishTime - startTime val tput = mb * 1000.0 / ms println("--------------------------") @@ -445,7 +674,7 @@ private[spark] object ConnectionManager { println("--------------------------") println("Continuous Sending") println("--------------------------") - val size = 10 * 1024 * 1024 + val size = 10 * 1024 * 1024 val count = 10 val buffer = ByteBuffer.allocate(size).put(Array.tabulate[Byte](size)(x => x.toByte)) diff --git a/core/src/main/scala/spark/network/ConnectionManagerId.scala b/core/src/main/scala/spark/network/ConnectionManagerId.scala new file mode 100644 index 0000000000..b554e84251 --- /dev/null +++ b/core/src/main/scala/spark/network/ConnectionManagerId.scala @@ -0,0 +1,21 @@ +package spark.network + +import java.net.InetSocketAddress + +import spark.Utils + + +private[spark] case class ConnectionManagerId(host: String, port: Int) { + // DEBUG code + Utils.checkHost(host) + assert (port > 0) + + def toSocketAddress() = new InetSocketAddress(host, port) +} + + +private[spark] object ConnectionManagerId { + def fromSocketAddress(socketAddress: InetSocketAddress): ConnectionManagerId = { + new ConnectionManagerId(socketAddress.getHostName(), socketAddress.getPort()) + } +} diff --git a/core/src/main/scala/spark/network/Message.scala b/core/src/main/scala/spark/network/Message.scala index 525751b5bf..d4f03610eb 100644 --- a/core/src/main/scala/spark/network/Message.scala +++ b/core/src/main/scala/spark/network/Message.scala @@ -1,55 +1,10 @@ package spark.network -import spark._ - -import scala.collection.mutable.ArrayBuffer - import java.nio.ByteBuffer -import java.net.InetAddress import java.net.InetSocketAddress -import storage.BlockManager - -private[spark] class MessageChunkHeader( - val typ: Long, - val id: Int, - val totalSize: Int, - val chunkSize: Int, - val other: Int, - val address: InetSocketAddress) { - lazy val buffer = { - val ip = address.getAddress.getAddress() - val port = address.getPort() - ByteBuffer. - allocate(MessageChunkHeader.HEADER_SIZE). - putLong(typ). - putInt(id). - putInt(totalSize). - putInt(chunkSize). - putInt(other). - putInt(ip.size). - put(ip). - putInt(port). - position(MessageChunkHeader.HEADER_SIZE). - flip.asInstanceOf[ByteBuffer] - } - - override def toString = "" + this.getClass.getSimpleName + ":" + id + " of type " + typ + - " and sizes " + totalSize + " / " + chunkSize + " bytes" -} -private[spark] class MessageChunk(val header: MessageChunkHeader, val buffer: ByteBuffer) { - val size = if (buffer == null) 0 else buffer.remaining - lazy val buffers = { - val ab = new ArrayBuffer[ByteBuffer]() - ab += header.buffer - if (buffer != null) { - ab += buffer - } - ab - } +import scala.collection.mutable.ArrayBuffer - override def toString = "" + this.getClass.getSimpleName + " (id = " + header.id + ", size = " + size + ")" -} private[spark] abstract class Message(val typ: Long, val id: Int) { var senderAddress: InetSocketAddress = null @@ -58,120 +13,16 @@ private[spark] abstract class Message(val typ: Long, val id: Int) { var finishTime = -1L def size: Int - + def getChunkForSending(maxChunkSize: Int): Option[MessageChunk] - + def getChunkForReceiving(chunkSize: Int): Option[MessageChunk] - + def timeTaken(): String = (finishTime - startTime).toString + " ms" override def toString = this.getClass.getSimpleName + "(id = " + id + ", size = " + size + ")" } -private[spark] class BufferMessage(id_ : Int, val buffers: ArrayBuffer[ByteBuffer], var ackId: Int) -extends Message(Message.BUFFER_MESSAGE, id_) { - - val initialSize = currentSize() - var gotChunkForSendingOnce = false - - def size = initialSize - - def currentSize() = { - if (buffers == null || buffers.isEmpty) { - 0 - } else { - buffers.map(_.remaining).reduceLeft(_ + _) - } - } - - def getChunkForSending(maxChunkSize: Int): Option[MessageChunk] = { - if (maxChunkSize <= 0) { - throw new Exception("Max chunk size is " + maxChunkSize) - } - - if (size == 0 && gotChunkForSendingOnce == false) { - val newChunk = new MessageChunk(new MessageChunkHeader(typ, id, 0, 0, ackId, senderAddress), null) - gotChunkForSendingOnce = true - return Some(newChunk) - } - - while(!buffers.isEmpty) { - val buffer = buffers(0) - if (buffer.remaining == 0) { - BlockManager.dispose(buffer) - buffers -= buffer - } else { - val newBuffer = if (buffer.remaining <= maxChunkSize) { - buffer.duplicate() - } else { - buffer.slice().limit(maxChunkSize).asInstanceOf[ByteBuffer] - } - buffer.position(buffer.position + newBuffer.remaining) - val newChunk = new MessageChunk(new MessageChunkHeader( - typ, id, size, newBuffer.remaining, ackId, senderAddress), newBuffer) - gotChunkForSendingOnce = true - return Some(newChunk) - } - } - None - } - - def getChunkForReceiving(chunkSize: Int): Option[MessageChunk] = { - // STRONG ASSUMPTION: BufferMessage created when receiving data has ONLY ONE data buffer - if (buffers.size > 1) { - throw new Exception("Attempting to get chunk from message with multiple data buffers") - } - val buffer = buffers(0) - if (buffer.remaining > 0) { - if (buffer.remaining < chunkSize) { - throw new Exception("Not enough space in data buffer for receiving chunk") - } - val newBuffer = buffer.slice().limit(chunkSize).asInstanceOf[ByteBuffer] - buffer.position(buffer.position + newBuffer.remaining) - val newChunk = new MessageChunk(new MessageChunkHeader( - typ, id, size, newBuffer.remaining, ackId, senderAddress), newBuffer) - return Some(newChunk) - } - None - } - - def flip() { - buffers.foreach(_.flip) - } - - def hasAckId() = (ackId != 0) - - def isCompletelyReceived() = !buffers(0).hasRemaining - - override def toString = { - if (hasAckId) { - "BufferAckMessage(aid = " + ackId + ", id = " + id + ", size = " + size + ")" - } else { - "BufferMessage(id = " + id + ", size = " + size + ")" - } - } -} - -private[spark] object MessageChunkHeader { - val HEADER_SIZE = 40 - - def create(buffer: ByteBuffer): MessageChunkHeader = { - if (buffer.remaining != HEADER_SIZE) { - throw new IllegalArgumentException("Cannot convert buffer data to Message") - } - val typ = buffer.getLong() - val id = buffer.getInt() - val totalSize = buffer.getInt() - val chunkSize = buffer.getInt() - val other = buffer.getInt() - val ipSize = buffer.getInt() - val ipBytes = new Array[Byte](ipSize) - buffer.get(ipBytes) - val ip = InetAddress.getByAddress(ipBytes) - val port = buffer.getInt() - new MessageChunkHeader(typ, id, totalSize, chunkSize, other, new InetSocketAddress(ip, port)) - } -} private[spark] object Message { val BUFFER_MESSAGE = 1111111111L @@ -180,14 +31,16 @@ private[spark] object Message { def getNewId() = synchronized { lastId += 1 - if (lastId == 0) lastId += 1 + if (lastId == 0) { + lastId += 1 + } lastId } def createBufferMessage(dataBuffers: Seq[ByteBuffer], ackId: Int): BufferMessage = { if (dataBuffers == null) { return new BufferMessage(getNewId(), new ArrayBuffer[ByteBuffer], ackId) - } + } if (dataBuffers.exists(_ == null)) { throw new Exception("Attempting to create buffer message with null buffer") } @@ -196,7 +49,7 @@ private[spark] object Message { def createBufferMessage(dataBuffers: Seq[ByteBuffer]): BufferMessage = createBufferMessage(dataBuffers, 0) - + def createBufferMessage(dataBuffer: ByteBuffer, ackId: Int): BufferMessage = { if (dataBuffer == null) { return createBufferMessage(Array(ByteBuffer.allocate(0)), ackId) @@ -204,15 +57,18 @@ private[spark] object Message { return createBufferMessage(Array(dataBuffer), ackId) } } - - def createBufferMessage(dataBuffer: ByteBuffer): BufferMessage = + + def createBufferMessage(dataBuffer: ByteBuffer): BufferMessage = createBufferMessage(dataBuffer, 0) - - def createBufferMessage(ackId: Int): BufferMessage = createBufferMessage(new Array[ByteBuffer](0), ackId) + + def createBufferMessage(ackId: Int): BufferMessage = { + createBufferMessage(new Array[ByteBuffer](0), ackId) + } def create(header: MessageChunkHeader): Message = { val newMessage: Message = header.typ match { - case BUFFER_MESSAGE => new BufferMessage(header.id, ArrayBuffer(ByteBuffer.allocate(header.totalSize)), header.other) + case BUFFER_MESSAGE => new BufferMessage(header.id, + ArrayBuffer(ByteBuffer.allocate(header.totalSize)), header.other) } newMessage.senderAddress = header.address newMessage diff --git a/core/src/main/scala/spark/network/MessageChunk.scala b/core/src/main/scala/spark/network/MessageChunk.scala new file mode 100644 index 0000000000..aaf9204d0e --- /dev/null +++ b/core/src/main/scala/spark/network/MessageChunk.scala @@ -0,0 +1,25 @@ +package spark.network + +import java.nio.ByteBuffer + +import scala.collection.mutable.ArrayBuffer + + +private[network] +class MessageChunk(val header: MessageChunkHeader, val buffer: ByteBuffer) { + + val size = if (buffer == null) 0 else buffer.remaining + + lazy val buffers = { + val ab = new ArrayBuffer[ByteBuffer]() + ab += header.buffer + if (buffer != null) { + ab += buffer + } + ab + } + + override def toString = { + "" + this.getClass.getSimpleName + " (id = " + header.id + ", size = " + size + ")" + } +} diff --git a/core/src/main/scala/spark/network/MessageChunkHeader.scala b/core/src/main/scala/spark/network/MessageChunkHeader.scala new file mode 100644 index 0000000000..3693d509d6 --- /dev/null +++ b/core/src/main/scala/spark/network/MessageChunkHeader.scala @@ -0,0 +1,58 @@ +package spark.network + +import java.net.InetAddress +import java.net.InetSocketAddress +import java.nio.ByteBuffer + + +private[spark] class MessageChunkHeader( + val typ: Long, + val id: Int, + val totalSize: Int, + val chunkSize: Int, + val other: Int, + val address: InetSocketAddress) { + lazy val buffer = { + // No need to change this, at 'use' time, we do a reverse lookup of the hostname. + // Refer to network.Connection + val ip = address.getAddress.getAddress() + val port = address.getPort() + ByteBuffer. + allocate(MessageChunkHeader.HEADER_SIZE). + putLong(typ). + putInt(id). + putInt(totalSize). + putInt(chunkSize). + putInt(other). + putInt(ip.size). + put(ip). + putInt(port). + position(MessageChunkHeader.HEADER_SIZE). + flip.asInstanceOf[ByteBuffer] + } + + override def toString = "" + this.getClass.getSimpleName + ":" + id + " of type " + typ + + " and sizes " + totalSize + " / " + chunkSize + " bytes" +} + + +private[spark] object MessageChunkHeader { + val HEADER_SIZE = 40 + + def create(buffer: ByteBuffer): MessageChunkHeader = { + if (buffer.remaining != HEADER_SIZE) { + throw new IllegalArgumentException("Cannot convert buffer data to Message") + } + val typ = buffer.getLong() + val id = buffer.getInt() + val totalSize = buffer.getInt() + val chunkSize = buffer.getInt() + val other = buffer.getInt() + val ipSize = buffer.getInt() + val ipBytes = new Array[Byte](ipSize) + buffer.get(ipBytes) + val ip = InetAddress.getByAddress(ipBytes) + val port = buffer.getInt() + new MessageChunkHeader(typ, id, totalSize, chunkSize, other, new InetSocketAddress(ip, port)) + } +} diff --git a/core/src/main/scala/spark/network/netty/FileHeader.scala b/core/src/main/scala/spark/network/netty/FileHeader.scala new file mode 100644 index 0000000000..aed4254234 --- /dev/null +++ b/core/src/main/scala/spark/network/netty/FileHeader.scala @@ -0,0 +1,57 @@ +package spark.network.netty + +import io.netty.buffer._ + +import spark.Logging + +private[spark] class FileHeader ( + val fileLen: Int, + val blockId: String) extends Logging { + + lazy val buffer = { + val buf = Unpooled.buffer() + buf.capacity(FileHeader.HEADER_SIZE) + buf.writeInt(fileLen) + buf.writeInt(blockId.length) + blockId.foreach((x: Char) => buf.writeByte(x)) + //padding the rest of header + if (FileHeader.HEADER_SIZE - buf.readableBytes > 0 ) { + buf.writeZero(FileHeader.HEADER_SIZE - buf.readableBytes) + } else { + throw new Exception("too long header " + buf.readableBytes) + logInfo("too long header") + } + buf + } + +} + +private[spark] object FileHeader { + + val HEADER_SIZE = 40 + + def getFileLenOffset = 0 + def getFileLenSize = Integer.SIZE/8 + + def create(buf: ByteBuf): FileHeader = { + val length = buf.readInt + val idLength = buf.readInt + val idBuilder = new StringBuilder(idLength) + for (i <- 1 to idLength) { + idBuilder += buf.readByte().asInstanceOf[Char] + } + val blockId = idBuilder.toString() + new FileHeader(length, blockId) + } + + + def main (args:Array[String]){ + + val header = new FileHeader(25,"block_0"); + val buf = header.buffer; + val newheader = FileHeader.create(buf); + System.out.println("id="+newheader.blockId+",size="+newheader.fileLen) + + } +} + diff --git a/core/src/main/scala/spark/network/netty/ShuffleCopier.scala b/core/src/main/scala/spark/network/netty/ShuffleCopier.scala new file mode 100644 index 0000000000..8d5194a737 --- /dev/null +++ b/core/src/main/scala/spark/network/netty/ShuffleCopier.scala @@ -0,0 +1,101 @@ +package spark.network.netty + +import java.util.concurrent.Executors + +import io.netty.buffer.ByteBuf +import io.netty.channel.ChannelHandlerContext +import io.netty.util.CharsetUtil + +import spark.Logging +import spark.network.ConnectionManagerId + +import scala.collection.JavaConverters._ + + +private[spark] class ShuffleCopier extends Logging { + + def getBlock(host: String, port: Int, blockId: String, + resultCollectCallback: (String, Long, ByteBuf) => Unit) { + + val handler = new ShuffleCopier.ShuffleClientHandler(resultCollectCallback) + val connectTimeout = System.getProperty("spark.shuffle.netty.connect.timeout", "60000").toInt + val fc = new FileClient(handler, connectTimeout) + + try { + fc.init() + fc.connect(host, port) + fc.sendRequest(blockId) + fc.waitForClose() + fc.close() + } catch { + // Handle any socket-related exceptions in FileClient + case e: Exception => { + logError("Shuffle copy of block " + blockId + " from " + host + ":" + port + " failed", e) + handler.handleError(blockId) + } + } + } + + def getBlock(cmId: ConnectionManagerId, blockId: String, + resultCollectCallback: (String, Long, ByteBuf) => Unit) { + getBlock(cmId.host, cmId.port, blockId, resultCollectCallback) + } + + def getBlocks(cmId: ConnectionManagerId, + blocks: Seq[(String, Long)], + resultCollectCallback: (String, Long, ByteBuf) => Unit) { + + for ((blockId, size) <- blocks) { + getBlock(cmId, blockId, resultCollectCallback) + } + } +} + + +private[spark] object ShuffleCopier extends Logging { + + private class ShuffleClientHandler(resultCollectCallBack: (String, Long, ByteBuf) => Unit) + extends FileClientHandler with Logging { + + override def handle(ctx: ChannelHandlerContext, in: ByteBuf, header: FileHeader) { + logDebug("Received Block: " + header.blockId + " (" + header.fileLen + "B)"); + resultCollectCallBack(header.blockId, header.fileLen.toLong, in.readBytes(header.fileLen)) + } + + override def handleError(blockId: String) { + if (!isComplete) { + resultCollectCallBack(blockId, -1, null) + } + } + } + + def echoResultCollectCallBack(blockId: String, size: Long, content: ByteBuf) { + if (size != -1) { + logInfo("File: " + blockId + " content is : \" " + content.toString(CharsetUtil.UTF_8) + "\"") + } + } + + def main(args: Array[String]) { + if (args.length < 3) { + System.err.println("Usage: ShuffleCopier <host> <port> <shuffle_block_id> <threads>") + System.exit(1) + } + val host = args(0) + val port = args(1).toInt + val file = args(2) + val threads = if (args.length > 3) args(3).toInt else 10 + + val copiers = Executors.newFixedThreadPool(80) + val tasks = (for (i <- Range(0, threads)) yield { + Executors.callable(new Runnable() { + def run() { + val copier = new ShuffleCopier() + copier.getBlock(host, port, file, echoResultCollectCallBack) + } + }) + }).asJava + copiers.invokeAll(tasks) + copiers.shutdown + System.exit(0) + } +} diff --git a/core/src/main/scala/spark/network/netty/ShuffleSender.scala b/core/src/main/scala/spark/network/netty/ShuffleSender.scala new file mode 100644 index 0000000000..d6fa4b1e80 --- /dev/null +++ b/core/src/main/scala/spark/network/netty/ShuffleSender.scala @@ -0,0 +1,53 @@ +package spark.network.netty + +import java.io.File + +import spark.Logging + + +private[spark] class ShuffleSender(portIn: Int, val pResolver: PathResolver) extends Logging { + + val server = new FileServer(pResolver, portIn) + server.start() + + def stop() { + server.stop() + } + + def port: Int = server.getPort() +} + + +/** + * An application for testing the shuffle sender as a standalone program. + */ +private[spark] object ShuffleSender { + + def main(args: Array[String]) { + if (args.length < 3) { + System.err.println( + "Usage: ShuffleSender <port> <subDirsPerLocalDir> <list of shuffle_block_directories>") + System.exit(1) + } + + val port = args(0).toInt + val subDirsPerLocalDir = args(1).toInt + val localDirs = args.drop(2).map(new File(_)) + + val pResovler = new PathResolver { + override def getAbsolutePath(blockId: String): String = { + if (!blockId.startsWith("shuffle_")) { + throw new Exception("Block " + blockId + " is not a shuffle block") + } + // Figure out which local directory it hashes to, and which subdirectory in that + val hash = math.abs(blockId.hashCode) + val dirId = hash % localDirs.length + val subDirId = (hash / localDirs.length) % subDirsPerLocalDir + val subDir = new File(localDirs(dirId), "%02x".format(subDirId)) + val file = new File(subDir, blockId) + return file.getAbsolutePath + } + } + val sender = new ShuffleSender(port, pResovler) + } +} diff --git a/core/src/main/scala/spark/rdd/BlockRDD.scala b/core/src/main/scala/spark/rdd/BlockRDD.scala index 7348c4f15b..719d4bf03e 100644 --- a/core/src/main/scala/spark/rdd/BlockRDD.scala +++ b/core/src/main/scala/spark/rdd/BlockRDD.scala @@ -1,7 +1,7 @@ package spark.rdd -import scala.collection.mutable.HashMap import spark.{RDD, SparkContext, SparkEnv, Partition, TaskContext} +import spark.storage.BlockManager private[spark] class BlockRDDPartition(val blockId: String, idx: Int) extends Partition { val index = idx @@ -11,12 +11,7 @@ private[spark] class BlockRDD[T: ClassManifest](sc: SparkContext, @transient blockIds: Array[String]) extends RDD[T](sc, Nil) { - @transient lazy val locations_ = { - val blockManager = SparkEnv.get.blockManager - /*val locations = blockIds.map(id => blockManager.getLocations(id))*/ - val locations = blockManager.getLocations(blockIds) - HashMap(blockIds.zip(locations):_*) - } + @transient lazy val locations_ = BlockManager.blockIdsToExecutorLocations(blockIds, SparkEnv.get) override def getPartitions: Array[Partition] = (0 until blockIds.size).map(i => { new BlockRDDPartition(blockIds(i), i).asInstanceOf[Partition] diff --git a/core/src/main/scala/spark/rdd/CheckpointRDD.scala b/core/src/main/scala/spark/rdd/CheckpointRDD.scala index 36bfb0355e..377b1bdbe0 100644 --- a/core/src/main/scala/spark/rdd/CheckpointRDD.scala +++ b/core/src/main/scala/spark/rdd/CheckpointRDD.scala @@ -8,6 +8,7 @@ import org.apache.hadoop.util.ReflectionUtils import org.apache.hadoop.fs.Path import java.io.{File, IOException, EOFException} import java.text.NumberFormat +import spark.deploy.SparkHadoopUtil private[spark] class CheckpointRDDPartition(val index: Int) extends Partition {} @@ -21,13 +22,20 @@ class CheckpointRDD[T: ClassManifest](sc: SparkContext, val checkpointPath: Stri @transient val fs = new Path(checkpointPath).getFileSystem(sc.hadoopConfiguration) override def getPartitions: Array[Partition] = { - val dirContents = fs.listStatus(new Path(checkpointPath)) - val splitFiles = dirContents.map(_.getPath.toString).filter(_.contains("part-")).sorted - val numPartitions = splitFiles.size - if (!splitFiles(0).endsWith(CheckpointRDD.splitIdToFile(0)) || - !splitFiles(numPartitions-1).endsWith(CheckpointRDD.splitIdToFile(numPartitions-1))) { - throw new SparkException("Invalid checkpoint directory: " + checkpointPath) - } + val cpath = new Path(checkpointPath) + val numPartitions = + // listStatus can throw exception if path does not exist. + if (fs.exists(cpath)) { + val dirContents = fs.listStatus(cpath) + val partitionFiles = dirContents.map(_.getPath.toString).filter(_.contains("part-")).sorted + val numPart = partitionFiles.size + if (numPart > 0 && (! partitionFiles(0).endsWith(CheckpointRDD.splitIdToFile(0)) || + ! partitionFiles(numPart-1).endsWith(CheckpointRDD.splitIdToFile(numPart-1)))) { + throw new SparkException("Invalid checkpoint directory: " + checkpointPath) + } + numPart + } else 0 + Array.tabulate(numPartitions)(i => new CheckpointRDDPartition(i)) } @@ -35,7 +43,7 @@ class CheckpointRDD[T: ClassManifest](sc: SparkContext, val checkpointPath: Stri checkpointData.get.cpFile = Some(checkpointPath) override def getPreferredLocations(split: Partition): Seq[String] = { - val status = fs.getFileStatus(new Path(checkpointPath)) + val status = fs.getFileStatus(new Path(checkpointPath, CheckpointRDD.splitIdToFile(split.index))) val locations = fs.getFileBlockLocations(status, 0, status.getLen) locations.headOption.toList.flatMap(_.getHosts).filter(_ != "localhost") } @@ -58,7 +66,7 @@ private[spark] object CheckpointRDD extends Logging { def writeToFile[T](path: String, blockSize: Int = -1)(ctx: TaskContext, iterator: Iterator[T]) { val outputDir = new Path(path) - val fs = outputDir.getFileSystem(new Configuration()) + val fs = outputDir.getFileSystem(SparkHadoopUtil.newConfiguration()) val finalOutputName = splitIdToFile(ctx.splitId) val finalOutputPath = new Path(outputDir, finalOutputName) @@ -83,6 +91,7 @@ private[spark] object CheckpointRDD extends Logging { if (!fs.rename(tempOutputPath, finalOutputPath)) { if (!fs.exists(finalOutputPath)) { + logInfo("Deleting tempOutputPath " + tempOutputPath) fs.delete(tempOutputPath, false) throw new IOException("Checkpoint failed: failed to save output of task: " + ctx.attemptId + " and final output path does not exist") @@ -95,7 +104,7 @@ private[spark] object CheckpointRDD extends Logging { } def readFromFile[T](path: Path, context: TaskContext): Iterator[T] = { - val fs = path.getFileSystem(new Configuration()) + val fs = path.getFileSystem(SparkHadoopUtil.newConfiguration()) val bufferSize = System.getProperty("spark.buffer.size", "65536").toInt val fileInputStream = fs.open(path, bufferSize) val serializer = SparkEnv.get.serializer.newInstance() @@ -117,11 +126,11 @@ private[spark] object CheckpointRDD extends Logging { val sc = new SparkContext(cluster, "CheckpointRDD Test") val rdd = sc.makeRDD(1 to 10, 10).flatMap(x => 1 to 10000) val path = new Path(hdfsPath, "temp") - val fs = path.getFileSystem(new Configuration()) + val fs = path.getFileSystem(SparkHadoopUtil.newConfiguration()) sc.runJob(rdd, CheckpointRDD.writeToFile(path.toString, 1024) _) val cpRDD = new CheckpointRDD[Int](sc, path.toString) assert(cpRDD.partitions.length == rdd.partitions.length, "Number of partitions is not the same") assert(cpRDD.collect.toList == rdd.collect.toList, "Data of partitions not the same") - fs.delete(path) + fs.delete(path, true) } } diff --git a/core/src/main/scala/spark/rdd/CoGroupedRDD.scala b/core/src/main/scala/spark/rdd/CoGroupedRDD.scala index 5200fb6b65..8966f9f86e 100644 --- a/core/src/main/scala/spark/rdd/CoGroupedRDD.scala +++ b/core/src/main/scala/spark/rdd/CoGroupedRDD.scala @@ -2,10 +2,11 @@ package spark.rdd import java.io.{ObjectOutputStream, IOException} import java.util.{HashMap => JHashMap} + import scala.collection.JavaConversions import scala.collection.mutable.ArrayBuffer -import spark.{Aggregator, Logging, Partitioner, RDD, SparkEnv, Partition, TaskContext} +import spark.{Aggregator, Partition, Partitioner, RDD, SparkEnv, TaskContext} import spark.{Dependency, OneToOneDependency, ShuffleDependency} @@ -28,7 +29,8 @@ private[spark] case class NarrowCoGroupSplitDep( private[spark] case class ShuffleCoGroupSplitDep(shuffleId: Int) extends CoGroupSplitDep private[spark] -class CoGroupPartition(idx: Int, val deps: Seq[CoGroupSplitDep]) extends Partition with Serializable { +class CoGroupPartition(idx: Int, val deps: Array[CoGroupSplitDep]) + extends Partition with Serializable { override val index: Int = idx override def hashCode(): Int = idx } @@ -40,7 +42,24 @@ private[spark] class CoGroupAggregator { (b1, b2) => b1 ++ b2 }) with Serializable -class CoGroupedRDD[K](@transient var rdds: Seq[RDD[(K, _)]], part: Partitioner) + +/** + * A RDD that cogroups its parents. For each key k in parent RDDs, the resulting RDD contains a + * tuple with the list of values for that key. + * + * @param rdds parent RDDs. + * @param part partitioner used to partition the shuffle output. + * @param mapSideCombine flag indicating whether to merge values before shuffle step. If the flag + * is on, Spark does an extra pass over the data on the map side to merge + * all values belonging to the same key together. This can reduce the amount + * of data shuffled if and only if the number of distinct keys is very small, + * and the ratio of key size to value size is also very small. + */ +class CoGroupedRDD[K]( + @transient var rdds: Seq[RDD[(K, _)]], + part: Partitioner, + val mapSideCombine: Boolean = false, + val serializerClass: String = null) extends RDD[(K, Seq[Seq[_]])](rdds.head.context, Nil) { private val aggr = new CoGroupAggregator @@ -52,8 +71,12 @@ class CoGroupedRDD[K](@transient var rdds: Seq[RDD[(K, _)]], part: Partitioner) new OneToOneDependency(rdd) } else { logInfo("Adding shuffle dependency with " + rdd) - val mapSideCombinedRDD = rdd.mapPartitions(aggr.combineValuesByKey(_), true) - new ShuffleDependency[Any, ArrayBuffer[Any]](mapSideCombinedRDD, part) + if (mapSideCombine) { + val mapSideCombinedRDD = rdd.mapPartitions(aggr.combineValuesByKey(_), true) + new ShuffleDependency[Any, ArrayBuffer[Any]](mapSideCombinedRDD, part, serializerClass) + } else { + new ShuffleDependency[Any, Any](rdd.asInstanceOf[RDD[(Any, Any)]], part, serializerClass) + } } } } @@ -70,7 +93,7 @@ class CoGroupedRDD[K](@transient var rdds: Seq[RDD[(K, _)]], part: Partitioner) case _ => new NarrowCoGroupSplitDep(rdd, i, rdd.partitions(i)) } - }.toList) + }.toArray) } array } @@ -82,6 +105,7 @@ class CoGroupedRDD[K](@transient var rdds: Seq[RDD[(K, _)]], part: Partitioner) val numRdds = split.deps.size // e.g. for `(k, a) cogroup (k, b)`, K -> Seq(ArrayBuffer as, ArrayBuffer bs) val map = new JHashMap[K, Seq[ArrayBuffer[Any]]] + def getSeq(k: K): Seq[ArrayBuffer[Any]] = { val seq = map.get(k) if (seq != null) { @@ -92,6 +116,8 @@ class CoGroupedRDD[K](@transient var rdds: Seq[RDD[(K, _)]], part: Partitioner) seq } } + + val ser = SparkEnv.get.serializerManager.get(serializerClass) for ((dep, depNum) <- split.deps.zipWithIndex) dep match { case NarrowCoGroupSplitDep(rdd, _, itsSplit) => { // Read them from the parent @@ -102,8 +128,16 @@ class CoGroupedRDD[K](@transient var rdds: Seq[RDD[(K, _)]], part: Partitioner) case ShuffleCoGroupSplitDep(shuffleId) => { // Read map outputs of shuffle val fetcher = SparkEnv.get.shuffleFetcher - for ((k, vs) <- fetcher.fetch[K, Seq[Any]](shuffleId, split.index)) { - getSeq(k)(depNum) ++= vs + if (mapSideCombine) { + // With map side combine on, for each key, the shuffle fetcher returns a list of values. + fetcher.fetch[K, Seq[Any]](shuffleId, split.index, context.taskMetrics, ser).foreach { + case (key, values) => getSeq(key)(depNum) ++= values + } + } else { + // With map side combine off, for each key the shuffle fetcher returns a single value. + fetcher.fetch[K, Any](shuffleId, split.index, context.taskMetrics, ser).foreach { + case (key, value) => getSeq(key)(depNum) += value + } } } } diff --git a/core/src/main/scala/spark/rdd/CoalescedRDD.scala b/core/src/main/scala/spark/rdd/CoalescedRDD.scala index 0d16cf6e85..6d862c0c28 100644 --- a/core/src/main/scala/spark/rdd/CoalescedRDD.scala +++ b/core/src/main/scala/spark/rdd/CoalescedRDD.scala @@ -37,8 +37,8 @@ class CoalescedRDD[T: ClassManifest]( prevSplits.map(_.index).map{idx => new CoalescedRDDPartition(idx, prev, Array(idx)) } } else { (0 until maxPartitions).map { i => - val rangeStart = (i * prevSplits.length) / maxPartitions - val rangeEnd = ((i + 1) * prevSplits.length) / maxPartitions + val rangeStart = ((i.toLong * prevSplits.length) / maxPartitions).toInt + val rangeEnd = (((i.toLong + 1) * prevSplits.length) / maxPartitions).toInt new CoalescedRDDPartition(i, prev, (rangeStart until rangeEnd).toArray) }.toArray } diff --git a/core/src/main/scala/spark/rdd/EmptyRDD.scala b/core/src/main/scala/spark/rdd/EmptyRDD.scala new file mode 100644 index 0000000000..e4dd3a7fa7 --- /dev/null +++ b/core/src/main/scala/spark/rdd/EmptyRDD.scala @@ -0,0 +1,16 @@ +package spark.rdd + +import spark.{RDD, SparkContext, SparkEnv, Partition, TaskContext} + + +/** + * An RDD that is empty, i.e. has no element in it. + */ +class EmptyRDD[T: ClassManifest](sc: SparkContext) extends RDD[T](sc, Nil) { + + override def getPartitions: Array[Partition] = Array.empty + + override def compute(split: Partition, context: TaskContext): Iterator[T] = { + throw new UnsupportedOperationException("empty RDD") + } +} diff --git a/core/src/main/scala/spark/rdd/HadoopRDD.scala b/core/src/main/scala/spark/rdd/HadoopRDD.scala index 8139a2a40c..cbf5512e24 100644 --- a/core/src/main/scala/spark/rdd/HadoopRDD.scala +++ b/core/src/main/scala/spark/rdd/HadoopRDD.scala @@ -15,7 +15,9 @@ import org.apache.hadoop.mapred.RecordReader import org.apache.hadoop.mapred.Reporter import org.apache.hadoop.util.ReflectionUtils -import spark.{Dependency, RDD, SerializableWritable, SparkContext, Partition, TaskContext} +import spark.{Dependency, Logging, Partition, RDD, SerializableWritable, SparkContext, TaskContext} +import spark.util.NextIterator +import org.apache.hadoop.conf.Configurable /** @@ -42,13 +44,16 @@ class HadoopRDD[K, V]( keyClass: Class[K], valueClass: Class[V], minSplits: Int) - extends RDD[(K, V)](sc, Nil) { + extends RDD[(K, V)](sc, Nil) with Logging { // A Hadoop JobConf can be about 10 KB, which is pretty big, so broadcast it private val confBroadcast = sc.broadcast(new SerializableWritable(conf)) override def getPartitions: Array[Partition] = { val inputFormat = createInputFormat(conf) + if (inputFormat.isInstanceOf[Configurable]) { + inputFormat.asInstanceOf[Configurable].setConf(conf) + } val inputSplits = inputFormat.getSplits(conf, minSplits) val array = new Array[Partition](inputSplits.size) for (i <- 0 until inputSplits.size) { @@ -62,48 +67,40 @@ class HadoopRDD[K, V]( .asInstanceOf[InputFormat[K, V]] } - override def compute(theSplit: Partition, context: TaskContext) = new Iterator[(K, V)] { + override def compute(theSplit: Partition, context: TaskContext) = new NextIterator[(K, V)] { val split = theSplit.asInstanceOf[HadoopPartition] var reader: RecordReader[K, V] = null val conf = confBroadcast.value.value val fmt = createInputFormat(conf) + if (fmt.isInstanceOf[Configurable]) { + fmt.asInstanceOf[Configurable].setConf(conf) + } reader = fmt.getRecordReader(split.inputSplit.value, conf, Reporter.NULL) // Register an on-task-completion callback to close the input stream. - context.addOnCompleteCallback(() => reader.close()) + context.addOnCompleteCallback{ () => closeIfNeeded() } val key: K = reader.createKey() val value: V = reader.createValue() - var gotNext = false - var finished = false - - override def hasNext: Boolean = { - if (!gotNext) { - try { - finished = !reader.next(key, value) - } catch { - case eof: EOFException => - finished = true - } - gotNext = true - } - if (finished) { - reader.close() - } - !finished - } - override def next: (K, V) = { - if (!gotNext) { + override def getNext() = { + try { finished = !reader.next(key, value) + } catch { + case eof: EOFException => + finished = true } - if (finished) { - throw new NoSuchElementException("End of stream") - } - gotNext = false (key, value) } + + override def close() { + try { + reader.close() + } catch { + case e: Exception => logWarning("Exception in RecordReader.close()", e) + } + } } override def getPreferredLocations(split: Partition): Seq[String] = { diff --git a/core/src/main/scala/spark/rdd/JdbcRDD.scala b/core/src/main/scala/spark/rdd/JdbcRDD.scala new file mode 100644 index 0000000000..a50f407737 --- /dev/null +++ b/core/src/main/scala/spark/rdd/JdbcRDD.scala @@ -0,0 +1,103 @@ +package spark.rdd + +import java.sql.{Connection, ResultSet} + +import spark.{Logging, Partition, RDD, SparkContext, TaskContext} +import spark.util.NextIterator + +private[spark] class JdbcPartition(idx: Int, val lower: Long, val upper: Long) extends Partition { + override def index = idx +} + +/** + * An RDD that executes an SQL query on a JDBC connection and reads results. + * For usage example, see test case JdbcRDDSuite. + * + * @param getConnection a function that returns an open Connection. + * The RDD takes care of closing the connection. + * @param sql the text of the query. + * The query must contain two ? placeholders for parameters used to partition the results. + * E.g. "select title, author from books where ? <= id and id <= ?" + * @param lowerBound the minimum value of the first placeholder + * @param upperBound the maximum value of the second placeholder + * The lower and upper bounds are inclusive. + * @param numPartitions the number of partitions. + * Given a lowerBound of 1, an upperBound of 20, and a numPartitions of 2, + * the query would be executed twice, once with (1, 10) and once with (11, 20) + * @param mapRow a function from a ResultSet to a single row of the desired result type(s). + * This should only call getInt, getString, etc; the RDD takes care of calling next. + * The default maps a ResultSet to an array of Object. + */ +class JdbcRDD[T: ClassManifest]( + sc: SparkContext, + getConnection: () => Connection, + sql: String, + lowerBound: Long, + upperBound: Long, + numPartitions: Int, + mapRow: (ResultSet) => T = JdbcRDD.resultSetToObjectArray _) + extends RDD[T](sc, Nil) with Logging { + + override def getPartitions: Array[Partition] = { + // bounds are inclusive, hence the + 1 here and - 1 on end + val length = 1 + upperBound - lowerBound + (0 until numPartitions).map(i => { + val start = lowerBound + ((i * length) / numPartitions).toLong + val end = lowerBound + (((i + 1) * length) / numPartitions).toLong - 1 + new JdbcPartition(i, start, end) + }).toArray + } + + override def compute(thePart: Partition, context: TaskContext) = new NextIterator[T] { + context.addOnCompleteCallback{ () => closeIfNeeded() } + val part = thePart.asInstanceOf[JdbcPartition] + val conn = getConnection() + val stmt = conn.prepareStatement(sql, ResultSet.TYPE_FORWARD_ONLY, ResultSet.CONCUR_READ_ONLY) + + // setFetchSize(Integer.MIN_VALUE) is a mysql driver specific way to force streaming results, + // rather than pulling entire resultset into memory. + // see http://dev.mysql.com/doc/refman/5.0/en/connector-j-reference-implementation-notes.html + if (conn.getMetaData.getURL.matches("jdbc:mysql:.*")) { + stmt.setFetchSize(Integer.MIN_VALUE) + logInfo("statement fetch size set to: " + stmt.getFetchSize + " to force MySQL streaming ") + } + + stmt.setLong(1, part.lower) + stmt.setLong(2, part.upper) + val rs = stmt.executeQuery() + + override def getNext: T = { + if (rs.next()) { + mapRow(rs) + } else { + finished = true + null.asInstanceOf[T] + } + } + + override def close() { + try { + if (null != rs && ! rs.isClosed()) rs.close() + } catch { + case e: Exception => logWarning("Exception closing resultset", e) + } + try { + if (null != stmt && ! stmt.isClosed()) stmt.close() + } catch { + case e: Exception => logWarning("Exception closing statement", e) + } + try { + if (null != conn && ! stmt.isClosed()) conn.close() + logInfo("closed connection") + } catch { + case e: Exception => logWarning("Exception closing connection", e) + } + } + } +} + +object JdbcRDD { + def resultSetToObjectArray(rs: ResultSet) = { + Array.tabulate[Object](rs.getMetaData.getColumnCount)(i => rs.getObject(i + 1)) + } +} diff --git a/core/src/main/scala/spark/rdd/NewHadoopRDD.scala b/core/src/main/scala/spark/rdd/NewHadoopRDD.scala index ebd4c3f0e2..901d01ef30 100644 --- a/core/src/main/scala/spark/rdd/NewHadoopRDD.scala +++ b/core/src/main/scala/spark/rdd/NewHadoopRDD.scala @@ -3,11 +3,11 @@ package spark.rdd import java.text.SimpleDateFormat import java.util.Date -import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.conf.{Configurable, Configuration} import org.apache.hadoop.io.Writable import org.apache.hadoop.mapreduce._ -import spark.{Dependency, RDD, SerializableWritable, SparkContext, Partition, TaskContext} +import spark.{Dependency, Logging, Partition, RDD, SerializableWritable, SparkContext, TaskContext} private[spark] @@ -26,7 +26,8 @@ class NewHadoopRDD[K, V]( valueClass: Class[V], @transient conf: Configuration) extends RDD[(K, V)](sc, Nil) - with HadoopMapReduceUtil { + with HadoopMapReduceUtil + with Logging { // A Hadoop Configuration can be about 10 KB, which is pretty big, so broadcast it private val confBroadcast = sc.broadcast(new SerializableWritable(conf)) @@ -41,6 +42,9 @@ class NewHadoopRDD[K, V]( override def getPartitions: Array[Partition] = { val inputFormat = inputFormatClass.newInstance + if (inputFormat.isInstanceOf[Configurable]) { + inputFormat.asInstanceOf[Configurable].setConf(conf) + } val jobContext = newJobContext(conf, jobId) val rawSplits = inputFormat.getSplits(jobContext).toArray val result = new Array[Partition](rawSplits.size) @@ -53,15 +57,18 @@ class NewHadoopRDD[K, V]( override def compute(theSplit: Partition, context: TaskContext) = new Iterator[(K, V)] { val split = theSplit.asInstanceOf[NewHadoopPartition] val conf = confBroadcast.value.value - val attemptId = new TaskAttemptID(jobtrackerId, id, true, split.index, 0) + val attemptId = newTaskAttemptID(jobtrackerId, id, true, split.index, 0) val hadoopAttemptContext = newTaskAttemptContext(conf, attemptId) val format = inputFormatClass.newInstance + if (format.isInstanceOf[Configurable]) { + format.asInstanceOf[Configurable].setConf(conf) + } val reader = format.createRecordReader( split.serializableHadoopSplit.value, hadoopAttemptContext) reader.initialize(split.serializableHadoopSplit.value, hadoopAttemptContext) // Register an on-task-completion callback to close the input stream. - context.addOnCompleteCallback(() => reader.close()) + context.addOnCompleteCallback(() => close()) var havePair = false var finished = false @@ -81,6 +88,14 @@ class NewHadoopRDD[K, V]( havePair = false return (reader.getCurrentKey, reader.getCurrentValue) } + + private def close() { + try { + reader.close() + } catch { + case e: Exception => logWarning("Exception in RecordReader.close()", e) + } + } } override def getPreferredLocations(split: Partition): Seq[String] = { diff --git a/core/src/main/scala/spark/rdd/PipedRDD.scala b/core/src/main/scala/spark/rdd/PipedRDD.scala index 962a1b21ad..c0baf43d43 100644 --- a/core/src/main/scala/spark/rdd/PipedRDD.scala +++ b/core/src/main/scala/spark/rdd/PipedRDD.scala @@ -9,6 +9,7 @@ import scala.collection.mutable.ArrayBuffer import scala.io.Source import spark.{RDD, SparkEnv, Partition, TaskContext} +import spark.broadcast.Broadcast /** @@ -18,14 +19,21 @@ import spark.{RDD, SparkEnv, Partition, TaskContext} class PipedRDD[T: ClassManifest]( prev: RDD[T], command: Seq[String], - envVars: Map[String, String]) + envVars: Map[String, String], + printPipeContext: (String => Unit) => Unit, + printRDDElement: (T, String => Unit) => Unit) extends RDD[String](prev) { - def this(prev: RDD[T], command: Seq[String]) = this(prev, command, Map()) - // Similar to Runtime.exec(), if we are given a single string, split it into words // using a standard StringTokenizer (i.e. by spaces) - def this(prev: RDD[T], command: String) = this(prev, PipedRDD.tokenize(command)) + def this( + prev: RDD[T], + command: String, + envVars: Map[String, String] = Map(), + printPipeContext: (String => Unit) => Unit = null, + printRDDElement: (T, String => Unit) => Unit = null) = + this(prev, PipedRDD.tokenize(command), envVars, printPipeContext, printRDDElement) + override def getPartitions: Array[Partition] = firstParent[T].partitions @@ -52,8 +60,17 @@ class PipedRDD[T: ClassManifest]( override def run() { SparkEnv.set(env) val out = new PrintWriter(proc.getOutputStream) + + // input the pipe context firstly + if (printPipeContext != null) { + printPipeContext(out.println(_)) + } for (elem <- firstParent[T].iterator(split, context)) { - out.println(elem) + if (printRDDElement != null) { + printRDDElement(elem, out.println(_)) + } else { + out.println(elem) + } } out.close() } diff --git a/core/src/main/scala/spark/rdd/ShuffledRDD.scala b/core/src/main/scala/spark/rdd/ShuffledRDD.scala index c2f118305f..c7d1926b83 100644 --- a/core/src/main/scala/spark/rdd/ShuffledRDD.scala +++ b/core/src/main/scala/spark/rdd/ShuffledRDD.scala @@ -3,6 +3,7 @@ package spark.rdd import spark.{Partitioner, RDD, SparkEnv, ShuffleDependency, Partition, TaskContext} import spark.SparkContext._ + private[spark] class ShuffledRDDPartition(val idx: Int) extends Partition { override val index = idx override def hashCode(): Int = idx @@ -12,13 +13,15 @@ private[spark] class ShuffledRDDPartition(val idx: Int) extends Partition { * The resulting RDD from a shuffle (e.g. repartitioning of data). * @param prev the parent RDD. * @param part the partitioner used to partition the RDD + * @param serializerClass class name of the serializer to use. * @tparam K the key class. * @tparam V the value class. */ class ShuffledRDD[K, V]( - prev: RDD[(K, V)], - part: Partitioner) - extends RDD[(K, V)](prev.context, List(new ShuffleDependency(prev, part))) { + @transient prev: RDD[(K, V)], + part: Partitioner, + serializerClass: String = null) + extends RDD[(K, V)](prev.context, List(new ShuffleDependency(prev, part, serializerClass))) { override val partitioner = Some(part) @@ -28,6 +31,7 @@ class ShuffledRDD[K, V]( override def compute(split: Partition, context: TaskContext): Iterator[(K, V)] = { val shuffledId = dependencies.head.asInstanceOf[ShuffleDependency[K, V]].shuffleId - SparkEnv.get.shuffleFetcher.fetch[K, V](shuffledId, split.index) + SparkEnv.get.shuffleFetcher.fetch[K, V](shuffledId, split.index, context.taskMetrics, + SparkEnv.get.serializerManager.get(serializerClass)) } } diff --git a/core/src/main/scala/spark/rdd/SubtractedRDD.scala b/core/src/main/scala/spark/rdd/SubtractedRDD.scala new file mode 100644 index 0000000000..8a9efc5da2 --- /dev/null +++ b/core/src/main/scala/spark/rdd/SubtractedRDD.scala @@ -0,0 +1,108 @@ +package spark.rdd + +import java.util.{HashMap => JHashMap} +import scala.collection.JavaConversions._ +import scala.collection.mutable.ArrayBuffer +import spark.RDD +import spark.Partitioner +import spark.Dependency +import spark.TaskContext +import spark.Partition +import spark.SparkEnv +import spark.ShuffleDependency +import spark.OneToOneDependency + + +/** + * An optimized version of cogroup for set difference/subtraction. + * + * It is possible to implement this operation with just `cogroup`, but + * that is less efficient because all of the entries from `rdd2`, for + * both matching and non-matching values in `rdd1`, are kept in the + * JHashMap until the end. + * + * With this implementation, only the entries from `rdd1` are kept in-memory, + * and the entries from `rdd2` are essentially streamed, as we only need to + * touch each once to decide if the value needs to be removed. + * + * This is particularly helpful when `rdd1` is much smaller than `rdd2`, as + * you can use `rdd1`'s partitioner/partition size and not worry about running + * out of memory because of the size of `rdd2`. + */ +private[spark] class SubtractedRDD[K: ClassManifest, V: ClassManifest, W: ClassManifest]( + @transient var rdd1: RDD[(K, V)], + @transient var rdd2: RDD[(K, W)], + part: Partitioner, + val serializerClass: String = null) + extends RDD[(K, V)](rdd1.context, Nil) { + + override def getDependencies: Seq[Dependency[_]] = { + Seq(rdd1, rdd2).map { rdd => + if (rdd.partitioner == Some(part)) { + logInfo("Adding one-to-one dependency with " + rdd) + new OneToOneDependency(rdd) + } else { + logInfo("Adding shuffle dependency with " + rdd) + new ShuffleDependency(rdd.asInstanceOf[RDD[(K, Any)]], part, serializerClass) + } + } + } + + override def getPartitions: Array[Partition] = { + val array = new Array[Partition](part.numPartitions) + for (i <- 0 until array.size) { + // Each CoGroupPartition will depend on rdd1 and rdd2 + array(i) = new CoGroupPartition(i, Seq(rdd1, rdd2).zipWithIndex.map { case (rdd, j) => + dependencies(j) match { + case s: ShuffleDependency[_, _] => + new ShuffleCoGroupSplitDep(s.shuffleId) + case _ => + new NarrowCoGroupSplitDep(rdd, i, rdd.partitions(i)) + } + }.toArray) + } + array + } + + override val partitioner = Some(part) + + override def compute(p: Partition, context: TaskContext): Iterator[(K, V)] = { + val partition = p.asInstanceOf[CoGroupPartition] + val serializer = SparkEnv.get.serializerManager.get(serializerClass) + val map = new JHashMap[K, ArrayBuffer[V]] + def getSeq(k: K): ArrayBuffer[V] = { + val seq = map.get(k) + if (seq != null) { + seq + } else { + val seq = new ArrayBuffer[V]() + map.put(k, seq) + seq + } + } + def integrate(dep: CoGroupSplitDep, op: ((K, V)) => Unit) = dep match { + case NarrowCoGroupSplitDep(rdd, _, itsSplit) => { + for (t <- rdd.iterator(itsSplit, context)) + op(t.asInstanceOf[(K, V)]) + } + case ShuffleCoGroupSplitDep(shuffleId) => { + val iter = SparkEnv.get.shuffleFetcher.fetch(shuffleId, partition.index, + context.taskMetrics, serializer) + for (t <- iter) + op(t.asInstanceOf[(K, V)]) + } + } + // the first dep is rdd1; add all values to the map + integrate(partition.deps(0), t => getSeq(t._1) += t._2) + // the second dep is rdd2; remove all of its keys + integrate(partition.deps(1), t => map.remove(t._1)) + map.iterator.map { t => t._2.iterator.map { (t._1, _) } }.flatten + } + + override def clearDependencies() { + super.clearDependencies() + rdd1 = null + rdd2 = null + } + +} diff --git a/core/src/main/scala/spark/rdd/ZippedPartitionsRDD.scala b/core/src/main/scala/spark/rdd/ZippedPartitionsRDD.scala new file mode 100644 index 0000000000..b234428ab2 --- /dev/null +++ b/core/src/main/scala/spark/rdd/ZippedPartitionsRDD.scala @@ -0,0 +1,138 @@ +package spark.rdd + +import spark.{Utils, OneToOneDependency, RDD, SparkContext, Partition, TaskContext} +import java.io.{ObjectOutputStream, IOException} + +private[spark] class ZippedPartitionsPartition( + idx: Int, + @transient rdds: Seq[RDD[_]]) + extends Partition { + + override val index: Int = idx + var partitionValues = rdds.map(rdd => rdd.partitions(idx)) + def partitions = partitionValues + + @throws(classOf[IOException]) + private def writeObject(oos: ObjectOutputStream) { + // Update the reference to parent split at the time of task serialization + partitionValues = rdds.map(rdd => rdd.partitions(idx)) + oos.defaultWriteObject() + } +} + +abstract class ZippedPartitionsBaseRDD[V: ClassManifest]( + sc: SparkContext, + var rdds: Seq[RDD[_]]) + extends RDD[V](sc, rdds.map(x => new OneToOneDependency(x))) { + + override def getPartitions: Array[Partition] = { + val sizes = rdds.map(x => x.partitions.size) + if (!sizes.forall(x => x == sizes(0))) { + throw new IllegalArgumentException("Can't zip RDDs with unequal numbers of partitions") + } + val array = new Array[Partition](sizes(0)) + for (i <- 0 until sizes(0)) { + array(i) = new ZippedPartitionsPartition(i, rdds) + } + array + } + + override def getPreferredLocations(s: Partition): Seq[String] = { + // Note that as number of rdd's increase and/or number of slaves in cluster increase, the computed preferredLocations below + // become diminishingly small : so we might need to look at alternate strategies to alleviate this. + // If there are no (or very small number of preferred locations), we will end up transferred the blocks to 'any' node in the + // cluster - paying with n/w and cache cost. + // Maybe pick a node which figures max amount of time ? + // Choose node which is hosting 'larger' of some subset of blocks ? + // Look at rack locality to ensure chosen host is atleast rack local to both hosting node ?, etc (would be good to defer this if possible) + val splits = s.asInstanceOf[ZippedPartitionsPartition].partitions + val rddSplitZip = rdds.zip(splits) + + // exact match. + val exactMatchPreferredLocations = rddSplitZip.map(x => x._1.preferredLocations(x._2)) + val exactMatchLocations = exactMatchPreferredLocations.reduce((x, y) => x.intersect(y)) + + // Remove exact match and then do host local match. + val exactMatchHosts = exactMatchLocations.map(Utils.parseHostPort(_)._1) + val matchPreferredHosts = exactMatchPreferredLocations.map(locs => locs.map(Utils.parseHostPort(_)._1)) + .reduce((x, y) => x.intersect(y)) + val otherNodeLocalLocations = matchPreferredHosts.filter { s => !exactMatchHosts.contains(s) } + + otherNodeLocalLocations ++ exactMatchLocations + } + + override def clearDependencies() { + super.clearDependencies() + rdds = null + } +} + +class ZippedPartitionsRDD2[A: ClassManifest, B: ClassManifest, V: ClassManifest]( + sc: SparkContext, + f: (Iterator[A], Iterator[B]) => Iterator[V], + var rdd1: RDD[A], + var rdd2: RDD[B]) + extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2)) { + + override def compute(s: Partition, context: TaskContext): Iterator[V] = { + val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions + f(rdd1.iterator(partitions(0), context), rdd2.iterator(partitions(1), context)) + } + + override def clearDependencies() { + super.clearDependencies() + rdd1 = null + rdd2 = null + } +} + +class ZippedPartitionsRDD3 + [A: ClassManifest, B: ClassManifest, C: ClassManifest, V: ClassManifest]( + sc: SparkContext, + f: (Iterator[A], Iterator[B], Iterator[C]) => Iterator[V], + var rdd1: RDD[A], + var rdd2: RDD[B], + var rdd3: RDD[C]) + extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2, rdd3)) { + + override def compute(s: Partition, context: TaskContext): Iterator[V] = { + val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions + f(rdd1.iterator(partitions(0), context), + rdd2.iterator(partitions(1), context), + rdd3.iterator(partitions(2), context)) + } + + override def clearDependencies() { + super.clearDependencies() + rdd1 = null + rdd2 = null + rdd3 = null + } +} + +class ZippedPartitionsRDD4 + [A: ClassManifest, B: ClassManifest, C: ClassManifest, D:ClassManifest, V: ClassManifest]( + sc: SparkContext, + f: (Iterator[A], Iterator[B], Iterator[C], Iterator[D]) => Iterator[V], + var rdd1: RDD[A], + var rdd2: RDD[B], + var rdd3: RDD[C], + var rdd4: RDD[D]) + extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2, rdd3, rdd4)) { + + override def compute(s: Partition, context: TaskContext): Iterator[V] = { + val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions + f(rdd1.iterator(partitions(0), context), + rdd2.iterator(partitions(1), context), + rdd3.iterator(partitions(2), context), + rdd4.iterator(partitions(3), context)) + } + + override def clearDependencies() { + super.clearDependencies() + rdd1 = null + rdd2 = null + rdd3 = null + rdd4 = null + } +} diff --git a/core/src/main/scala/spark/rdd/ZippedRDD.scala b/core/src/main/scala/spark/rdd/ZippedRDD.scala index e80ec17aa5..f728e93d24 100644 --- a/core/src/main/scala/spark/rdd/ZippedRDD.scala +++ b/core/src/main/scala/spark/rdd/ZippedRDD.scala @@ -1,6 +1,6 @@ package spark.rdd -import spark.{OneToOneDependency, RDD, SparkContext, Partition, TaskContext} +import spark.{Utils, OneToOneDependency, RDD, SparkContext, Partition, TaskContext} import java.io.{ObjectOutputStream, IOException} @@ -10,17 +10,17 @@ private[spark] class ZippedPartition[T: ClassManifest, U: ClassManifest]( @transient rdd2: RDD[U] ) extends Partition { - var split1 = rdd1.partitions(idx) - var split2 = rdd1.partitions(idx) + var partition1 = rdd1.partitions(idx) + var partition2 = rdd2.partitions(idx) override val index: Int = idx - def splits = (split1, split2) + def partitions = (partition1, partition2) @throws(classOf[IOException]) private def writeObject(oos: ObjectOutputStream) { - // Update the reference to parent split at the time of task serialization - split1 = rdd1.partitions(idx) - split2 = rdd2.partitions(idx) + // Update the reference to parent partition at the time of task serialization + partition1 = rdd1.partitions(idx) + partition2 = rdd2.partitions(idx) oos.defaultWriteObject() } } @@ -43,13 +43,32 @@ class ZippedRDD[T: ClassManifest, U: ClassManifest]( } override def compute(s: Partition, context: TaskContext): Iterator[(T, U)] = { - val (split1, split2) = s.asInstanceOf[ZippedPartition[T, U]].splits - rdd1.iterator(split1, context).zip(rdd2.iterator(split2, context)) + val (partition1, partition2) = s.asInstanceOf[ZippedPartition[T, U]].partitions + rdd1.iterator(partition1, context).zip(rdd2.iterator(partition2, context)) } override def getPreferredLocations(s: Partition): Seq[String] = { - val (split1, split2) = s.asInstanceOf[ZippedPartition[T, U]].splits - rdd1.preferredLocations(split1).intersect(rdd2.preferredLocations(split2)) + // Note that as number of slaves in cluster increase, the computed preferredLocations can become small : so we might need + // to look at alternate strategies to alleviate this. (If there are no (or very small number of preferred locations), we + // will end up transferred the blocks to 'any' node in the cluster - paying with n/w and cache cost. + // Maybe pick one or the other ? (so that atleast one block is local ?). + // Choose node which is hosting 'larger' of the blocks ? + // Look at rack locality to ensure chosen host is atleast rack local to both hosting node ?, etc (would be good to defer this if possible) + val (partition1, partition2) = s.asInstanceOf[ZippedPartition[T, U]].partitions + val pref1 = rdd1.preferredLocations(partition1) + val pref2 = rdd2.preferredLocations(partition2) + + // exact match - instance local and host local. + val exactMatchLocations = pref1.intersect(pref2) + + // remove locations which are already handled via exactMatchLocations, and intersect where both partitions are node local. + val otherNodeLocalPref1 = pref1.filter(loc => ! exactMatchLocations.contains(loc)).map(loc => Utils.parseHostPort(loc)._1) + val otherNodeLocalPref2 = pref2.filter(loc => ! exactMatchLocations.contains(loc)).map(loc => Utils.parseHostPort(loc)._1) + val otherNodeLocalLocations = otherNodeLocalPref1.intersect(otherNodeLocalPref2) + + + // Can have mix of instance local (hostPort) and node local (host) locations as preference ! + exactMatchLocations ++ otherNodeLocalLocations } override def clearDependencies() { diff --git a/core/src/main/scala/spark/scheduler/ActiveJob.scala b/core/src/main/scala/spark/scheduler/ActiveJob.scala index 5a4e9a582d..105eaecb22 100644 --- a/core/src/main/scala/spark/scheduler/ActiveJob.scala +++ b/core/src/main/scala/spark/scheduler/ActiveJob.scala @@ -2,6 +2,8 @@ package spark.scheduler import spark.TaskContext +import java.util.Properties + /** * Tracks information about an active job in the DAGScheduler. */ @@ -11,7 +13,8 @@ private[spark] class ActiveJob( val func: (TaskContext, Iterator[_]) => _, val partitions: Array[Int], val callSite: String, - val listener: JobListener) { + val listener: JobListener, + val properties: Properties) { val numPartitions = partitions.length val finished = Array.fill[Boolean](numPartitions)(false) diff --git a/core/src/main/scala/spark/scheduler/DAGScheduler.scala b/core/src/main/scala/spark/scheduler/DAGScheduler.scala index bf0837c066..f7d60be5db 100644 --- a/core/src/main/scala/spark/scheduler/DAGScheduler.scala +++ b/core/src/main/scala/spark/scheduler/DAGScheduler.scala @@ -1,20 +1,20 @@ package spark.scheduler -import java.net.URI +import cluster.TaskInfo import java.util.concurrent.atomic.AtomicInteger -import java.util.concurrent.Future import java.util.concurrent.LinkedBlockingQueue import java.util.concurrent.TimeUnit +import java.util.Properties -import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet, Queue, Map} +import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet, Map} import spark._ +import spark.executor.TaskMetrics import spark.partial.ApproximateActionListener import spark.partial.ApproximateEvaluator import spark.partial.PartialResult -import spark.storage.BlockManagerMaster -import spark.storage.BlockManagerId -import util.{MetadataCleaner, TimeStampedHashMap} +import spark.storage.{BlockManager, BlockManagerMaster} +import spark.util.{MetadataCleaner, TimeStampedHashMap} /** * A Scheduler subclass that implements stage-oriented scheduling. It computes a DAG of stages for @@ -40,8 +40,10 @@ class DAGScheduler( task: Task[_], reason: TaskEndReason, result: Any, - accumUpdates: Map[Long, Any]) { - eventQueue.put(CompletionEvent(task, reason, result, accumUpdates)) + accumUpdates: Map[Long, Any], + taskInfo: TaskInfo, + taskMetrics: TaskMetrics) { + eventQueue.put(CompletionEvent(task, reason, result, accumUpdates, taskInfo, taskMetrics)) } // Called by TaskScheduler when an executor fails. @@ -49,6 +51,11 @@ class DAGScheduler( eventQueue.put(ExecutorLost(execId)) } + // Called by TaskScheduler when a host is added + override def executorGained(execId: String, hostPort: String) { + eventQueue.put(ExecutorGained(execId, hostPort)) + } + // Called by TaskScheduler to cancel an entire TaskSet due to repeated failures. override def taskSetFailed(taskSet: TaskSet, reason: String) { eventQueue.put(TaskSetFailed(taskSet, reason)) @@ -73,6 +80,10 @@ class DAGScheduler( val shuffleToMapStage = new TimeStampedHashMap[Int, Stage] + private[spark] val stageToInfos = new TimeStampedHashMap[Stage, StageInfo] + + private[spark] val sparkListeners = ArrayBuffer[SparkListener]() + var cacheLocs = new HashMap[Int, Array[List[String]]] // For tracking failed nodes, we use the MapOutputTracker's generation number, which is @@ -83,6 +94,8 @@ class DAGScheduler( // stray messages to detect. val failedGeneration = new HashMap[String, Long] + val idToActiveJob = new HashMap[Int, ActiveJob] + val waiting = new HashSet[Stage] // Stages we need to run whose parents aren't done val running = new HashSet[Stage] // Stages we are running right now val failed = new HashSet[Stage] // Stages that must be resubmitted due to fetch failures @@ -107,9 +120,8 @@ class DAGScheduler( private def getCacheLocs(rdd: RDD[_]): Array[List[String]] = { if (!cacheLocs.contains(rdd.id)) { val blockIds = rdd.partitions.indices.map(index=> "rdd_%d_%d".format(rdd.id, index)).toArray - cacheLocs(rdd.id) = blockManagerMaster.getLocations(blockIds).map { - locations => locations.map(_.ip).toList - }.toArray + val locs = BlockManager.blockIdsToExecutorLocations(blockIds, env, blockManagerMaster) + cacheLocs(rdd.id) = blockIds.map(locs.getOrElse(_, Nil)) } cacheLocs(rdd.id) } @@ -148,6 +160,7 @@ class DAGScheduler( val id = nextStageId.getAndIncrement() val stage = new Stage(id, rdd, shuffleDep, getParentStages(rdd, priority), priority) idToStage(id) = stage + stageToInfos(stage) = StageInfo(stage) stage } @@ -215,13 +228,14 @@ class DAGScheduler( partitions: Seq[Int], callSite: String, allowLocal: Boolean, - resultHandler: (Int, U) => Unit) + resultHandler: (Int, U) => Unit, + properties: Properties = null) : (JobSubmitted, JobWaiter[U]) = { assert(partitions.size > 0) val waiter = new JobWaiter(partitions.size, resultHandler) val func2 = func.asInstanceOf[(TaskContext, Iterator[_]) => _] - val toSubmit = JobSubmitted(finalRdd, func2, partitions.toArray, allowLocal, callSite, waiter) + val toSubmit = JobSubmitted(finalRdd, func2, partitions.toArray, allowLocal, callSite, waiter, properties) return (toSubmit, waiter) } @@ -231,13 +245,14 @@ class DAGScheduler( partitions: Seq[Int], callSite: String, allowLocal: Boolean, - resultHandler: (Int, U) => Unit) + resultHandler: (Int, U) => Unit, + properties: Properties = null) { if (partitions.size == 0) { return } val (toSubmit, waiter) = prepareJob( - finalRdd, func, partitions, callSite, allowLocal, resultHandler) + finalRdd, func, partitions, callSite, allowLocal, resultHandler, properties) eventQueue.put(toSubmit) waiter.awaitResult() match { case JobSucceeded => {} @@ -252,13 +267,14 @@ class DAGScheduler( func: (TaskContext, Iterator[T]) => U, evaluator: ApproximateEvaluator[U, R], callSite: String, - timeout: Long) + timeout: Long, + properties: Properties = null) : PartialResult[R] = { val listener = new ApproximateActionListener(rdd, func, evaluator, timeout) val func2 = func.asInstanceOf[(TaskContext, Iterator[_]) => _] val partitions = (0 until rdd.partitions.size).toArray - eventQueue.put(JobSubmitted(rdd, func2, partitions, false, callSite, listener)) + eventQueue.put(JobSubmitted(rdd, func2, partitions, false, callSite, listener, properties)) return listener.awaitResult() // Will throw an exception if the job fails } @@ -268,10 +284,10 @@ class DAGScheduler( */ private[scheduler] def processEvent(event: DAGSchedulerEvent): Boolean = { event match { - case JobSubmitted(finalRDD, func, partitions, allowLocal, callSite, listener) => + case JobSubmitted(finalRDD, func, partitions, allowLocal, callSite, listener, properties) => val runId = nextRunId.getAndIncrement() val finalStage = newStage(finalRDD, None, runId) - val job = new ActiveJob(runId, finalStage, func, partitions, callSite, listener) + val job = new ActiveJob(runId, finalStage, func, partitions, callSite, listener, properties) clearCacheLocs() logInfo("Got job " + job.runId + " (" + callSite + ") with " + partitions.length + " output partitions (allowLocal=" + allowLocal + ")") @@ -282,15 +298,22 @@ class DAGScheduler( // Compute very short actions like first() or take() with no parent stages locally. runLocally(job) } else { + sparkListeners.foreach(_.onJobStart(SparkListenerJobStart(job, properties))) + idToActiveJob(runId) = job activeJobs += job resultStageToJob(finalStage) = job submitStage(finalStage) } + case ExecutorGained(execId, hostPort) => + handleExecutorGained(execId, hostPort) + case ExecutorLost(execId) => handleExecutorLost(execId) case completion: CompletionEvent => + sparkListeners.foreach(_.onTaskEnd(SparkListenerTaskEnd(completion.task, + completion.reason, completion.taskInfo, completion.taskMetrics))) handleTaskCompletion(completion) case TaskSetFailed(taskSet, reason) => @@ -301,6 +324,7 @@ class DAGScheduler( for (job <- activeJobs) { val error = new SparkException("Job cancelled because SparkContext was shut down") job.listener.jobFailed(error) + sparkListeners.foreach(_.onJobEnd(SparkListenerJobEnd(job, JobFailed(error)))) } return true } @@ -320,7 +344,7 @@ class DAGScheduler( submitStage(stage) } } - + /** * Check for waiting or failed stages which are now eligible for resubmission. * Ordinarily run on every iteration of the event loop. @@ -379,29 +403,34 @@ class DAGScheduler( * We run the operation in a separate thread just in case it takes a bunch of time, so that we * don't block the DAGScheduler event loop or other concurrent jobs. */ - private def runLocally(job: ActiveJob) { + protected def runLocally(job: ActiveJob) { logInfo("Computing the requested partition locally") new Thread("Local computation of job " + job.runId) { override def run() { - try { - SparkEnv.set(env) - val rdd = job.finalStage.rdd - val split = rdd.partitions(job.partitions(0)) - val taskContext = new TaskContext(job.finalStage.id, job.partitions(0), 0) - try { - val result = job.func(taskContext, rdd.iterator(split, taskContext)) - job.listener.taskSucceeded(0, result) - } finally { - taskContext.executeOnCompleteCallbacks() - } - } catch { - case e: Exception => - job.listener.jobFailed(e) - } + runLocallyWithinThread(job) } }.start() } + // Broken out for easier testing in DAGSchedulerSuite. + protected def runLocallyWithinThread(job: ActiveJob) { + try { + SparkEnv.set(env) + val rdd = job.finalStage.rdd + val split = rdd.partitions(job.partitions(0)) + val taskContext = new TaskContext(job.finalStage.id, job.partitions(0), 0) + try { + val result = job.func(taskContext, rdd.iterator(split, taskContext)) + job.listener.taskSucceeded(0, result) + } finally { + taskContext.executeOnCompleteCallbacks() + } + } catch { + case e: Exception => + job.listener.jobFailed(e) + } + } + /** Submits stage, but first recursively submits any missing parents. */ private def submitStage(stage: Stage) { logDebug("submitStage(" + stage + ")") @@ -443,11 +472,13 @@ class DAGScheduler( } } if (tasks.size > 0) { + sparkListeners.foreach(_.onStageSubmitted(SparkListenerStageSubmitted(stage, tasks.size))) logInfo("Submitting " + tasks.size + " missing tasks from " + stage + " (" + stage.rdd + ")") myPending ++= tasks logDebug("New pending tasks: " + myPending) + val properties = idToActiveJob(stage.priority).properties taskSched.submitTasks( - new TaskSet(tasks.toArray, stage.id, stage.newAttemptId(), stage.priority)) + new TaskSet(tasks.toArray, stage.id, stage.newAttemptId(), stage.priority, properties)) if (!stage.submissionTime.isDefined) { stage.submissionTime = Some(System.currentTimeMillis()) } @@ -472,6 +503,8 @@ class DAGScheduler( case _ => "Unkown" } logInfo("%s (%s) finished in %s s".format(stage, stage.origin, serviceTime)) + val stageComp = StageCompleted(stageToInfos(stage)) + sparkListeners.foreach{_.onStageCompleted(stageComp)} running -= stage } event.reason match { @@ -481,6 +514,7 @@ class DAGScheduler( Accumulators.add(event.accumUpdates) // TODO: do this only if task wasn't resubmitted } pendingTasks(stage) -= task + stageToInfos(stage).taskInfos += event.taskInfo -> event.taskMetrics task match { case rt: ResultTask[_, _] => resultStageToJob.get(stage) match { @@ -493,6 +527,7 @@ class DAGScheduler( activeJobs -= job resultStageToJob -= stage markStageAsFinished(stage) + sparkListeners.foreach(_.onJobEnd(SparkListenerJobEnd(job, JobSucceeded))) } job.listener.taskSucceeded(rt.outputId, event.result) } @@ -501,7 +536,6 @@ class DAGScheduler( } case smt: ShuffleMapTask => - val stage = idToStage(smt.stageId) val status = event.result.asInstanceOf[MapStatus] val execId = status.location.executorId logDebug("ShuffleMapTask finished on " + execId) @@ -617,6 +651,14 @@ class DAGScheduler( "(generation " + currentGeneration + ")") } } + + private def handleExecutorGained(execId: String, hostPort: String) { + // remove from failedGeneration(execId) ? + if (failedGeneration.contains(execId)) { + logInfo("Host gained which was in lost list earlier: " + hostPort) + failedGeneration -= execId + } + } /** * Aborts all jobs depending on a particular Stage. This is called in response to a task set @@ -626,7 +668,9 @@ class DAGScheduler( val dependentStages = resultStageToJob.keys.filter(x => stageDependsOn(x, failedStage)).toSeq for (resultStage <- dependentStages) { val job = resultStageToJob(resultStage) - job.listener.jobFailed(new SparkException("Job failed: " + reason)) + val error = new SparkException("Job failed: " + reason) + job.listener.jobFailed(error) + sparkListeners.foreach(_.onJobEnd(SparkListenerJobEnd(job, JobFailed(error)))) activeJobs -= job resultStageToJob -= resultStage } @@ -699,7 +743,7 @@ class DAGScheduler( sizeBefore = shuffleToMapStage.size shuffleToMapStage.clearOldValues(cleanupTime) logInfo("shuffleToMapStage " + sizeBefore + " --> " + shuffleToMapStage.size) - + sizeBefore = pendingTasks.size pendingTasks.clearOldValues(cleanupTime) logInfo("pendingTasks " + sizeBefore + " --> " + pendingTasks.size) diff --git a/core/src/main/scala/spark/scheduler/DAGSchedulerEvent.scala b/core/src/main/scala/spark/scheduler/DAGSchedulerEvent.scala index b34fa78c07..acad915f13 100644 --- a/core/src/main/scala/spark/scheduler/DAGSchedulerEvent.scala +++ b/core/src/main/scala/spark/scheduler/DAGSchedulerEvent.scala @@ -1,8 +1,12 @@ package spark.scheduler +import java.util.Properties + +import spark.scheduler.cluster.TaskInfo import scala.collection.mutable.Map import spark._ +import spark.executor.TaskMetrics /** * Types of events that can be handled by the DAGScheduler. The DAGScheduler uses an event queue @@ -18,16 +22,23 @@ private[spark] case class JobSubmitted( partitions: Array[Int], allowLocal: Boolean, callSite: String, - listener: JobListener) + listener: JobListener, + properties: Properties = null) extends DAGSchedulerEvent private[spark] case class CompletionEvent( task: Task[_], reason: TaskEndReason, result: Any, - accumUpdates: Map[Long, Any]) + accumUpdates: Map[Long, Any], + taskInfo: TaskInfo, + taskMetrics: TaskMetrics) extends DAGSchedulerEvent +private[spark] case class ExecutorGained(execId: String, hostPort: String) extends DAGSchedulerEvent { + Utils.checkHostPort(hostPort, "Required hostport") +} + private[spark] case class ExecutorLost(execId: String) extends DAGSchedulerEvent private[spark] case class TaskSetFailed(taskSet: TaskSet, reason: String) extends DAGSchedulerEvent diff --git a/core/src/main/scala/spark/scheduler/InputFormatInfo.scala b/core/src/main/scala/spark/scheduler/InputFormatInfo.scala new file mode 100644 index 0000000000..287f731787 --- /dev/null +++ b/core/src/main/scala/spark/scheduler/InputFormatInfo.scala @@ -0,0 +1,156 @@ +package spark.scheduler + +import spark.Logging +import scala.collection.immutable.Set +import org.apache.hadoop.mapred.{FileInputFormat, JobConf} +import org.apache.hadoop.util.ReflectionUtils +import org.apache.hadoop.mapreduce.Job +import org.apache.hadoop.conf.Configuration +import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet} +import scala.collection.JavaConversions._ + + +/** + * Parses and holds information about inputFormat (and files) specified as a parameter. + */ +class InputFormatInfo(val configuration: Configuration, val inputFormatClazz: Class[_], + val path: String) extends Logging { + + var mapreduceInputFormat: Boolean = false + var mapredInputFormat: Boolean = false + + validate() + + override def toString(): String = { + "InputFormatInfo " + super.toString + " .. inputFormatClazz " + inputFormatClazz + ", path : " + path + } + + override def hashCode(): Int = { + var hashCode = inputFormatClazz.hashCode + hashCode = hashCode * 31 + path.hashCode + hashCode + } + + // Since we are not doing canonicalization of path, this can be wrong : like relative vs absolute path + // .. which is fine, this is best case effort to remove duplicates - right ? + override def equals(other: Any): Boolean = other match { + case that: InputFormatInfo => { + // not checking config - that should be fine, right ? + this.inputFormatClazz == that.inputFormatClazz && + this.path == that.path + } + case _ => false + } + + private def validate() { + logDebug("validate InputFormatInfo : " + inputFormatClazz + ", path " + path) + + try { + if (classOf[org.apache.hadoop.mapreduce.InputFormat[_, _]].isAssignableFrom(inputFormatClazz)) { + logDebug("inputformat is from mapreduce package") + mapreduceInputFormat = true + } + else if (classOf[org.apache.hadoop.mapred.InputFormat[_, _]].isAssignableFrom(inputFormatClazz)) { + logDebug("inputformat is from mapred package") + mapredInputFormat = true + } + else { + throw new IllegalArgumentException("Specified inputformat " + inputFormatClazz + + " is NOT a supported input format ? does not implement either of the supported hadoop api's") + } + } + catch { + case e: ClassNotFoundException => { + throw new IllegalArgumentException("Specified inputformat " + inputFormatClazz + " cannot be found ?", e) + } + } + } + + + // This method does not expect failures, since validate has already passed ... + private def prefLocsFromMapreduceInputFormat(): Set[SplitInfo] = { + val conf = new JobConf(configuration) + FileInputFormat.setInputPaths(conf, path) + + val instance: org.apache.hadoop.mapreduce.InputFormat[_, _] = + ReflectionUtils.newInstance(inputFormatClazz.asInstanceOf[Class[_]], conf).asInstanceOf[ + org.apache.hadoop.mapreduce.InputFormat[_, _]] + val job = new Job(conf) + + val retval = new ArrayBuffer[SplitInfo]() + val list = instance.getSplits(job) + for (split <- list) { + retval ++= SplitInfo.toSplitInfo(inputFormatClazz, path, split) + } + + return retval.toSet + } + + // This method does not expect failures, since validate has already passed ... + private def prefLocsFromMapredInputFormat(): Set[SplitInfo] = { + val jobConf = new JobConf(configuration) + FileInputFormat.setInputPaths(jobConf, path) + + val instance: org.apache.hadoop.mapred.InputFormat[_, _] = + ReflectionUtils.newInstance(inputFormatClazz.asInstanceOf[Class[_]], jobConf).asInstanceOf[ + org.apache.hadoop.mapred.InputFormat[_, _]] + + val retval = new ArrayBuffer[SplitInfo]() + instance.getSplits(jobConf, jobConf.getNumMapTasks()).foreach( + elem => retval ++= SplitInfo.toSplitInfo(inputFormatClazz, path, elem) + ) + + return retval.toSet + } + + private def findPreferredLocations(): Set[SplitInfo] = { + logDebug("mapreduceInputFormat : " + mapreduceInputFormat + ", mapredInputFormat : " + mapredInputFormat + + ", inputFormatClazz : " + inputFormatClazz) + if (mapreduceInputFormat) { + return prefLocsFromMapreduceInputFormat() + } + else { + assert(mapredInputFormat) + return prefLocsFromMapredInputFormat() + } + } +} + + + + +object InputFormatInfo { + /** + Computes the preferred locations based on input(s) and returned a location to block map. + Typical use of this method for allocation would follow some algo like this + (which is what we currently do in YARN branch) : + a) For each host, count number of splits hosted on that host. + b) Decrement the currently allocated containers on that host. + c) Compute rack info for each host and update rack -> count map based on (b). + d) Allocate nodes based on (c) + e) On the allocation result, ensure that we dont allocate "too many" jobs on a single node + (even if data locality on that is very high) : this is to prevent fragility of job if a single + (or small set of) hosts go down. + + go to (a) until required nodes are allocated. + + If a node 'dies', follow same procedure. + + PS: I know the wording here is weird, hopefully it makes some sense ! + */ + def computePreferredLocations(formats: Seq[InputFormatInfo]): HashMap[String, HashSet[SplitInfo]] = { + + val nodeToSplit = new HashMap[String, HashSet[SplitInfo]] + for (inputSplit <- formats) { + val splits = inputSplit.findPreferredLocations() + + for (split <- splits){ + val location = split.hostLocation + val set = nodeToSplit.getOrElseUpdate(location, new HashSet[SplitInfo]) + set += split + } + } + + nodeToSplit + } +} diff --git a/core/src/main/scala/spark/scheduler/JobLogger.scala b/core/src/main/scala/spark/scheduler/JobLogger.scala new file mode 100644 index 0000000000..178bfaba3d --- /dev/null +++ b/core/src/main/scala/spark/scheduler/JobLogger.scala @@ -0,0 +1,306 @@ +package spark.scheduler
+
+import java.io.PrintWriter
+import java.io.File
+import java.io.FileNotFoundException
+import java.text.SimpleDateFormat
+import java.util.{Date, Properties}
+import java.util.concurrent.LinkedBlockingQueue
+import scala.collection.mutable.{Map, HashMap, ListBuffer}
+import scala.io.Source
+import spark._
+import spark.executor.TaskMetrics
+import spark.scheduler.cluster.TaskInfo
+
+// Used to record runtime information for each job, including RDD graph
+// tasks' start/stop shuffle information and information from outside
+
+class JobLogger(val logDirName: String) extends SparkListener with Logging {
+ private val logDir =
+ if (System.getenv("SPARK_LOG_DIR") != null)
+ System.getenv("SPARK_LOG_DIR")
+ else
+ "/tmp/spark"
+ private val jobIDToPrintWriter = new HashMap[Int, PrintWriter]
+ private val stageIDToJobID = new HashMap[Int, Int]
+ private val jobIDToStages = new HashMap[Int, ListBuffer[Stage]]
+ private val DATE_FORMAT = new SimpleDateFormat("yyyy/MM/dd HH:mm:ss")
+ private val eventQueue = new LinkedBlockingQueue[SparkListenerEvents]
+
+ createLogDir()
+ def this() = this(String.valueOf(System.currentTimeMillis()))
+
+ def getLogDir = logDir
+ def getJobIDtoPrintWriter = jobIDToPrintWriter
+ def getStageIDToJobID = stageIDToJobID
+ def getJobIDToStages = jobIDToStages
+ def getEventQueue = eventQueue
+
+ new Thread("JobLogger") {
+ setDaemon(true)
+ override def run() {
+ while (true) {
+ val event = eventQueue.take
+ logDebug("Got event of type " + event.getClass.getName)
+ event match {
+ case SparkListenerJobStart(job, properties) =>
+ processJobStartEvent(job, properties)
+ case SparkListenerStageSubmitted(stage, taskSize) =>
+ processStageSubmittedEvent(stage, taskSize)
+ case StageCompleted(stageInfo) =>
+ processStageCompletedEvent(stageInfo)
+ case SparkListenerJobEnd(job, result) =>
+ processJobEndEvent(job, result)
+ case SparkListenerTaskEnd(task, reason, taskInfo, taskMetrics) =>
+ processTaskEndEvent(task, reason, taskInfo, taskMetrics)
+ case _ =>
+ }
+ }
+ }
+ }.start()
+
+ // Create a folder for log files, the folder's name is the creation time of the jobLogger
+ protected def createLogDir() {
+ val dir = new File(logDir + "/" + logDirName + "/")
+ if (dir.exists()) {
+ return
+ }
+ if (dir.mkdirs() == false) {
+ logError("create log directory error:" + logDir + "/" + logDirName + "/")
+ }
+ }
+
+ // Create a log file for one job, the file name is the jobID
+ protected def createLogWriter(jobID: Int) {
+ try{
+ val fileWriter = new PrintWriter(logDir + "/" + logDirName + "/" + jobID)
+ jobIDToPrintWriter += (jobID -> fileWriter)
+ } catch {
+ case e: FileNotFoundException => e.printStackTrace()
+ }
+ }
+
+ // Close log file, and clean the stage relationship in stageIDToJobID
+ protected def closeLogWriter(jobID: Int) =
+ jobIDToPrintWriter.get(jobID).foreach { fileWriter =>
+ fileWriter.close()
+ jobIDToStages.get(jobID).foreach(_.foreach{ stage =>
+ stageIDToJobID -= stage.id
+ })
+ jobIDToPrintWriter -= jobID
+ jobIDToStages -= jobID
+ }
+
+ // Write log information to log file, withTime parameter controls whether to recored
+ // time stamp for the information
+ protected def jobLogInfo(jobID: Int, info: String, withTime: Boolean = true) {
+ var writeInfo = info
+ if (withTime) {
+ val date = new Date(System.currentTimeMillis())
+ writeInfo = DATE_FORMAT.format(date) + ": " +info
+ }
+ jobIDToPrintWriter.get(jobID).foreach(_.println(writeInfo))
+ }
+
+ protected def stageLogInfo(stageID: Int, info: String, withTime: Boolean = true) =
+ stageIDToJobID.get(stageID).foreach(jobID => jobLogInfo(jobID, info, withTime))
+
+ protected def buildJobDep(jobID: Int, stage: Stage) {
+ if (stage.priority == jobID) {
+ jobIDToStages.get(jobID) match {
+ case Some(stageList) => stageList += stage
+ case None => val stageList = new ListBuffer[Stage]
+ stageList += stage
+ jobIDToStages += (jobID -> stageList)
+ }
+ stageIDToJobID += (stage.id -> jobID)
+ stage.parents.foreach(buildJobDep(jobID, _))
+ }
+ }
+
+ protected def recordStageDep(jobID: Int) {
+ def getRddsInStage(rdd: RDD[_]): ListBuffer[RDD[_]] = {
+ var rddList = new ListBuffer[RDD[_]]
+ rddList += rdd
+ rdd.dependencies.foreach{ dep => dep match {
+ case shufDep: ShuffleDependency[_,_] =>
+ case _ => rddList ++= getRddsInStage(dep.rdd)
+ }
+ }
+ rddList
+ }
+ jobIDToStages.get(jobID).foreach {_.foreach { stage =>
+ var depRddDesc: String = ""
+ getRddsInStage(stage.rdd).foreach { rdd =>
+ depRddDesc += rdd.id + ","
+ }
+ var depStageDesc: String = ""
+ stage.parents.foreach { stage =>
+ depStageDesc += "(" + stage.id + "," + stage.shuffleDep.get.shuffleId + ")"
+ }
+ jobLogInfo(jobID, "STAGE_ID=" + stage.id + " RDD_DEP=(" +
+ depRddDesc.substring(0, depRddDesc.length - 1) + ")" +
+ " STAGE_DEP=" + depStageDesc, false)
+ }
+ }
+ }
+
+ // Generate indents and convert to String
+ protected def indentString(indent: Int) = {
+ val sb = new StringBuilder()
+ for (i <- 1 to indent) {
+ sb.append(" ")
+ }
+ sb.toString()
+ }
+
+ protected def getRddName(rdd: RDD[_]) = {
+ var rddName = rdd.getClass.getName
+ if (rdd.name != null) {
+ rddName = rdd.name
+ }
+ rddName
+ }
+
+ protected def recordRddInStageGraph(jobID: Int, rdd: RDD[_], indent: Int) {
+ val rddInfo = "RDD_ID=" + rdd.id + "(" + getRddName(rdd) + "," + rdd.generator + ")"
+ jobLogInfo(jobID, indentString(indent) + rddInfo, false)
+ rdd.dependencies.foreach{ dep => dep match {
+ case shufDep: ShuffleDependency[_,_] =>
+ val depInfo = "SHUFFLE_ID=" + shufDep.shuffleId
+ jobLogInfo(jobID, indentString(indent + 1) + depInfo, false)
+ case _ => recordRddInStageGraph(jobID, dep.rdd, indent + 1)
+ }
+ }
+ }
+
+ protected def recordStageDepGraph(jobID: Int, stage: Stage, indent: Int = 0) {
+ var stageInfo: String = ""
+ if (stage.isShuffleMap) {
+ stageInfo = "STAGE_ID=" + stage.id + " MAP_STAGE SHUFFLE_ID=" +
+ stage.shuffleDep.get.shuffleId
+ }else{
+ stageInfo = "STAGE_ID=" + stage.id + " RESULT_STAGE"
+ }
+ if (stage.priority == jobID) {
+ jobLogInfo(jobID, indentString(indent) + stageInfo, false)
+ recordRddInStageGraph(jobID, stage.rdd, indent)
+ stage.parents.foreach(recordStageDepGraph(jobID, _, indent + 2))
+ } else
+ jobLogInfo(jobID, indentString(indent) + stageInfo + " JOB_ID=" + stage.priority, false)
+ }
+
+ // Record task metrics into job log files
+ protected def recordTaskMetrics(stageID: Int, status: String,
+ taskInfo: TaskInfo, taskMetrics: TaskMetrics) {
+ val info = " TID=" + taskInfo.taskId + " STAGE_ID=" + stageID +
+ " START_TIME=" + taskInfo.launchTime + " FINISH_TIME=" + taskInfo.finishTime +
+ " EXECUTOR_ID=" + taskInfo.executorId + " HOST=" + taskMetrics.hostname
+ val executorRunTime = " EXECUTOR_RUN_TIME=" + taskMetrics.executorRunTime
+ val readMetrics =
+ taskMetrics.shuffleReadMetrics match {
+ case Some(metrics) =>
+ " SHUFFLE_FINISH_TIME=" + metrics.shuffleFinishTime +
+ " BLOCK_FETCHED_TOTAL=" + metrics.totalBlocksFetched +
+ " BLOCK_FETCHED_LOCAL=" + metrics.localBlocksFetched +
+ " BLOCK_FETCHED_REMOTE=" + metrics.remoteBlocksFetched +
+ " REMOTE_FETCH_WAIT_TIME=" + metrics.fetchWaitTime +
+ " REMOTE_FETCH_TIME=" + metrics.remoteFetchTime +
+ " REMOTE_BYTES_READ=" + metrics.remoteBytesRead
+ case None => ""
+ }
+ val writeMetrics =
+ taskMetrics.shuffleWriteMetrics match {
+ case Some(metrics) =>
+ " SHUFFLE_BYTES_WRITTEN=" + metrics.shuffleBytesWritten
+ case None => ""
+ }
+ stageLogInfo(stageID, status + info + executorRunTime + readMetrics + writeMetrics)
+ }
+
+ override def onStageSubmitted(stageSubmitted: SparkListenerStageSubmitted) {
+ eventQueue.put(stageSubmitted)
+ }
+
+ protected def processStageSubmittedEvent(stage: Stage, taskSize: Int) {
+ stageLogInfo(stage.id, "STAGE_ID=" + stage.id + " STATUS=SUBMITTED" + " TASK_SIZE=" + taskSize)
+ }
+
+ override def onStageCompleted(stageCompleted: StageCompleted) {
+ eventQueue.put(stageCompleted)
+ }
+
+ protected def processStageCompletedEvent(stageInfo: StageInfo) {
+ stageLogInfo(stageInfo.stage.id, "STAGE_ID=" +
+ stageInfo.stage.id + " STATUS=COMPLETED")
+
+ }
+
+ override def onTaskEnd(taskEnd: SparkListenerTaskEnd) {
+ eventQueue.put(taskEnd)
+ }
+
+ protected def processTaskEndEvent(task: Task[_], reason: TaskEndReason,
+ taskInfo: TaskInfo, taskMetrics: TaskMetrics) {
+ var taskStatus = ""
+ task match {
+ case resultTask: ResultTask[_, _] => taskStatus = "TASK_TYPE=RESULT_TASK"
+ case shuffleMapTask: ShuffleMapTask => taskStatus = "TASK_TYPE=SHUFFLE_MAP_TASK"
+ }
+ reason match {
+ case Success => taskStatus += " STATUS=SUCCESS"
+ recordTaskMetrics(task.stageId, taskStatus, taskInfo, taskMetrics)
+ case Resubmitted =>
+ taskStatus += " STATUS=RESUBMITTED TID=" + taskInfo.taskId +
+ " STAGE_ID=" + task.stageId
+ stageLogInfo(task.stageId, taskStatus)
+ case FetchFailed(bmAddress, shuffleId, mapId, reduceId) =>
+ taskStatus += " STATUS=FETCHFAILED TID=" + taskInfo.taskId + " STAGE_ID=" +
+ task.stageId + " SHUFFLE_ID=" + shuffleId + " MAP_ID=" +
+ mapId + " REDUCE_ID=" + reduceId
+ stageLogInfo(task.stageId, taskStatus)
+ case OtherFailure(message) =>
+ taskStatus += " STATUS=FAILURE TID=" + taskInfo.taskId +
+ " STAGE_ID=" + task.stageId + " INFO=" + message
+ stageLogInfo(task.stageId, taskStatus)
+ case _ =>
+ }
+ }
+
+ override def onJobEnd(jobEnd: SparkListenerJobEnd) {
+ eventQueue.put(jobEnd)
+ }
+
+ protected def processJobEndEvent(job: ActiveJob, reason: JobResult) {
+ var info = "JOB_ID=" + job.runId
+ reason match {
+ case JobSucceeded => info += " STATUS=SUCCESS"
+ case JobFailed(exception) =>
+ info += " STATUS=FAILED REASON="
+ exception.getMessage.split("\\s+").foreach(info += _ + "_")
+ case _ =>
+ }
+ jobLogInfo(job.runId, info.substring(0, info.length - 1).toUpperCase)
+ closeLogWriter(job.runId)
+ }
+
+ protected def recordJobProperties(jobID: Int, properties: Properties) {
+ if(properties != null) {
+ val annotation = properties.getProperty("spark.job.annotation", "")
+ jobLogInfo(jobID, annotation, false)
+ }
+ }
+
+ override def onJobStart(jobStart: SparkListenerJobStart) {
+ eventQueue.put(jobStart)
+ }
+
+ protected def processJobStartEvent(job: ActiveJob, properties: Properties) {
+ createLogWriter(job.runId)
+ recordJobProperties(job.runId, properties)
+ buildJobDep(job.runId, job.finalStage)
+ recordStageDep(job.runId)
+ recordStageDepGraph(job.runId, job.finalStage)
+ jobLogInfo(job.runId, "JOB_ID=" + job.runId + " STATUS=STARTED")
+ }
+}
diff --git a/core/src/main/scala/spark/scheduler/ResultTask.scala b/core/src/main/scala/spark/scheduler/ResultTask.scala index 1721f78f48..83166bce22 100644 --- a/core/src/main/scala/spark/scheduler/ResultTask.scala +++ b/core/src/main/scala/spark/scheduler/ResultTask.scala @@ -70,8 +70,16 @@ private[spark] class ResultTask[T, U]( rdd.partitions(partition) } + private val preferredLocs: Seq[String] = if (locs == null) Nil else locs.toSet.toSeq + + { + // DEBUG code + preferredLocs.foreach (hostPort => Utils.checkHost(Utils.parseHostPort(hostPort)._1, "preferredLocs : " + preferredLocs)) + } + override def run(attemptId: Long): U = { val context = new TaskContext(stageId, partition, attemptId) + metrics = Some(context.taskMetrics) try { func(context, rdd.iterator(split, context)) } finally { @@ -79,7 +87,7 @@ private[spark] class ResultTask[T, U]( } } - override def preferredLocations: Seq[String] = locs + override def preferredLocations: Seq[String] = preferredLocs override def toString = "ResultTask(" + stageId + ", " + partition + ")" diff --git a/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala b/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala index 59ee3c0a09..95647389c3 100644 --- a/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala +++ b/core/src/main/scala/spark/scheduler/ShuffleMapTask.scala @@ -13,8 +13,10 @@ import com.ning.compress.lzf.LZFInputStream import com.ning.compress.lzf.LZFOutputStream import spark._ +import spark.executor.ShuffleWriteMetrics import spark.storage._ -import util.{TimeStampedHashMap, MetadataCleaner} +import spark.util.{TimeStampedHashMap, MetadataCleaner} + private[spark] object ShuffleMapTask { @@ -76,13 +78,20 @@ private[spark] class ShuffleMapTask( var rdd: RDD[_], var dep: ShuffleDependency[_,_], var partition: Int, - @transient var locs: Seq[String]) + @transient private var locs: Seq[String]) extends Task[MapStatus](stageId) with Externalizable with Logging { protected def this() = this(0, null, null, 0, null) + @transient private val preferredLocs: Seq[String] = if (locs == null) Nil else locs.toSet.toSeq + + { + // DEBUG code + preferredLocs.foreach (hostPort => Utils.checkHost(Utils.parseHostPort(hostPort)._1, "preferredLocs : " + preferredLocs)) + } + var split = if (rdd == null) { null } else { @@ -119,34 +128,59 @@ private[spark] class ShuffleMapTask( val numOutputSplits = dep.partitioner.numPartitions val taskContext = new TaskContext(stageId, partition, attemptId) + metrics = Some(taskContext.taskMetrics) + + val blockManager = SparkEnv.get.blockManager + var shuffle: ShuffleBlocks = null + var buckets: ShuffleWriterGroup = null + try { - // Partition the map output. - val buckets = Array.fill(numOutputSplits)(new ArrayBuffer[(Any, Any)]) + // Obtain all the block writers for shuffle blocks. + val ser = SparkEnv.get.serializerManager.get(dep.serializerClass) + shuffle = blockManager.shuffleBlockManager.forShuffle(dep.shuffleId, numOutputSplits, ser) + buckets = shuffle.acquireWriters(partition) + + // Write the map output to its associated buckets. for (elem <- rdd.iterator(split, taskContext)) { val pair = elem.asInstanceOf[(Any, Any)] val bucketId = dep.partitioner.getPartition(pair._1) - buckets(bucketId) += pair + buckets.writers(bucketId).write(pair) } - val compressedSizes = new Array[Byte](numOutputSplits) - - val blockManager = SparkEnv.get.blockManager - for (i <- 0 until numOutputSplits) { - val blockId = "shuffle_" + dep.shuffleId + "_" + partition + "_" + i - // Get a Scala iterator from Java map - val iter: Iterator[(Any, Any)] = buckets(i).iterator - val size = blockManager.put(blockId, iter, StorageLevel.DISK_ONLY, false) - compressedSizes(i) = MapOutputTracker.compressSize(size) + // Commit the writes. Get the size of each bucket block (total block size). + var totalBytes = 0L + val compressedSizes: Array[Byte] = buckets.writers.map { writer: BlockObjectWriter => + writer.commit() + writer.close() + val size = writer.size() + totalBytes += size + MapOutputTracker.compressSize(size) } + // Update shuffle metrics. + val shuffleMetrics = new ShuffleWriteMetrics + shuffleMetrics.shuffleBytesWritten = totalBytes + metrics.get.shuffleWriteMetrics = Some(shuffleMetrics) + return new MapStatus(blockManager.blockManagerId, compressedSizes) + } catch { case e: Exception => + // If there is an exception from running the task, revert the partial writes + // and throw the exception upstream to Spark. + if (buckets != null) { + buckets.writers.foreach(_.revertPartialWrites()) + } + throw e } finally { + // Release the writers back to the shuffle block manager. + if (shuffle != null && buckets != null) { + shuffle.releaseWriters(buckets) + } // Execute the callbacks on task completion. taskContext.executeOnCompleteCallbacks() } } - override def preferredLocations: Seq[String] = locs + override def preferredLocations: Seq[String] = preferredLocs override def toString = "ShuffleMapTask(%d, %d)".format(stageId, partition) } diff --git a/core/src/main/scala/spark/scheduler/SparkListener.scala b/core/src/main/scala/spark/scheduler/SparkListener.scala new file mode 100644 index 0000000000..bac984b5c9 --- /dev/null +++ b/core/src/main/scala/spark/scheduler/SparkListener.scala @@ -0,0 +1,178 @@ +package spark.scheduler + +import java.util.Properties +import spark.scheduler.cluster.TaskInfo +import spark.util.Distribution +import spark.{Logging, SparkContext, TaskEndReason, Utils} +import spark.executor.TaskMetrics + +sealed trait SparkListenerEvents + +case class SparkListenerStageSubmitted(stage: Stage, taskSize: Int) extends SparkListenerEvents + +case class StageCompleted(val stageInfo: StageInfo) extends SparkListenerEvents + +case class SparkListenerTaskEnd(task: Task[_], reason: TaskEndReason, taskInfo: TaskInfo, + taskMetrics: TaskMetrics) extends SparkListenerEvents + +case class SparkListenerJobStart(job: ActiveJob, properties: Properties = null) + extends SparkListenerEvents + +case class SparkListenerJobEnd(job: ActiveJob, jobResult: JobResult) + extends SparkListenerEvents + +trait SparkListener { + /** + * Called when a stage is completed, with information on the completed stage + */ + def onStageCompleted(stageCompleted: StageCompleted) { } + + /** + * Called when a stage is submitted + */ + def onStageSubmitted(stageSubmitted: SparkListenerStageSubmitted) { } + + /** + * Called when a task ends + */ + def onTaskEnd(taskEnd: SparkListenerTaskEnd) { } + + /** + * Called when a job starts + */ + def onJobStart(jobStart: SparkListenerJobStart) { } + + /** + * Called when a job ends + */ + def onJobEnd(jobEnd: SparkListenerJobEnd) { } + +} + +/** + * Simple SparkListener that logs a few summary statistics when each stage completes + */ +class StatsReportListener extends SparkListener with Logging { + override def onStageCompleted(stageCompleted: StageCompleted) { + import spark.scheduler.StatsReportListener._ + implicit val sc = stageCompleted + this.logInfo("Finished stage: " + stageCompleted.stageInfo) + showMillisDistribution("task runtime:", (info, _) => Some(info.duration)) + + //shuffle write + showBytesDistribution("shuffle bytes written:",(_,metric) => metric.shuffleWriteMetrics.map{_.shuffleBytesWritten}) + + //fetch & io + showMillisDistribution("fetch wait time:",(_, metric) => metric.shuffleReadMetrics.map{_.fetchWaitTime}) + showBytesDistribution("remote bytes read:", (_, metric) => metric.shuffleReadMetrics.map{_.remoteBytesRead}) + showBytesDistribution("task result size:", (_, metric) => Some(metric.resultSize)) + + //runtime breakdown + val runtimePcts = stageCompleted.stageInfo.taskInfos.map{ + case (info, metrics) => RuntimePercentage(info.duration, metrics) + } + showDistribution("executor (non-fetch) time pct: ", Distribution(runtimePcts.map{_.executorPct * 100}), "%2.0f %%") + showDistribution("fetch wait time pct: ", Distribution(runtimePcts.flatMap{_.fetchPct.map{_ * 100}}), "%2.0f %%") + showDistribution("other time pct: ", Distribution(runtimePcts.map{_.other * 100}), "%2.0f %%") + } + +} + +object StatsReportListener extends Logging { + + //for profiling, the extremes are more interesting + val percentiles = Array[Int](0,5,10,25,50,75,90,95,100) + val probabilities = percentiles.map{_ / 100.0} + val percentilesHeader = "\t" + percentiles.mkString("%\t") + "%" + + def extractDoubleDistribution(stage:StageCompleted, getMetric: (TaskInfo,TaskMetrics) => Option[Double]): Option[Distribution] = { + Distribution(stage.stageInfo.taskInfos.flatMap{ + case ((info,metric)) => getMetric(info, metric)}) + } + + //is there some way to setup the types that I can get rid of this completely? + def extractLongDistribution(stage:StageCompleted, getMetric: (TaskInfo,TaskMetrics) => Option[Long]): Option[Distribution] = { + extractDoubleDistribution(stage, (info, metric) => getMetric(info,metric).map{_.toDouble}) + } + + def showDistribution(heading: String, d: Distribution, formatNumber: Double => String) { + val stats = d.statCounter + logInfo(heading + stats) + val quantiles = d.getQuantiles(probabilities).map{formatNumber} + logInfo(percentilesHeader) + logInfo("\t" + quantiles.mkString("\t")) + } + + def showDistribution(heading: String, dOpt: Option[Distribution], formatNumber: Double => String) { + dOpt.foreach { d => showDistribution(heading, d, formatNumber)} + } + + def showDistribution(heading: String, dOpt: Option[Distribution], format:String) { + def f(d:Double) = format.format(d) + showDistribution(heading, dOpt, f _) + } + + def showDistribution(heading:String, format: String, getMetric: (TaskInfo,TaskMetrics) => Option[Double]) + (implicit stage: StageCompleted) { + showDistribution(heading, extractDoubleDistribution(stage, getMetric), format) + } + + def showBytesDistribution(heading:String, getMetric: (TaskInfo,TaskMetrics) => Option[Long]) + (implicit stage: StageCompleted) { + showBytesDistribution(heading, extractLongDistribution(stage, getMetric)) + } + + def showBytesDistribution(heading: String, dOpt: Option[Distribution]) { + dOpt.foreach{dist => showBytesDistribution(heading, dist)} + } + + def showBytesDistribution(heading: String, dist: Distribution) { + showDistribution(heading, dist, (d => Utils.memoryBytesToString(d.toLong)): Double => String) + } + + def showMillisDistribution(heading: String, dOpt: Option[Distribution]) { + showDistribution(heading, dOpt, (d => StatsReportListener.millisToString(d.toLong)): Double => String) + } + + def showMillisDistribution(heading: String, getMetric: (TaskInfo, TaskMetrics) => Option[Long]) + (implicit stage: StageCompleted) { + showMillisDistribution(heading, extractLongDistribution(stage, getMetric)) + } + + + + val seconds = 1000L + val minutes = seconds * 60 + val hours = minutes * 60 + + /** + * reformat a time interval in milliseconds to a prettier format for output + */ + def millisToString(ms: Long) = { + val (size, units) = + if (ms > hours) { + (ms.toDouble / hours, "hours") + } else if (ms > minutes) { + (ms.toDouble / minutes, "min") + } else if (ms > seconds) { + (ms.toDouble / seconds, "s") + } else { + (ms.toDouble, "ms") + } + "%.1f %s".format(size, units) + } +} + + + +case class RuntimePercentage(executorPct: Double, fetchPct: Option[Double], other: Double) +object RuntimePercentage { + def apply(totalTime: Long, metrics: TaskMetrics): RuntimePercentage = { + val denom = totalTime.toDouble + val fetchTime = metrics.shuffleReadMetrics.map{_.fetchWaitTime} + val fetch = fetchTime.map{_ / denom} + val exec = (metrics.executorRunTime - fetchTime.getOrElse(0l)) / denom + val other = 1.0 - (exec + fetch.getOrElse(0d)) + RuntimePercentage(exec, fetch, other) + } +} diff --git a/core/src/main/scala/spark/scheduler/SplitInfo.scala b/core/src/main/scala/spark/scheduler/SplitInfo.scala new file mode 100644 index 0000000000..6abfb7a1f7 --- /dev/null +++ b/core/src/main/scala/spark/scheduler/SplitInfo.scala @@ -0,0 +1,61 @@ +package spark.scheduler + +import collection.mutable.ArrayBuffer + +// information about a specific split instance : handles both split instances. +// So that we do not need to worry about the differences. +class SplitInfo(val inputFormatClazz: Class[_], val hostLocation: String, val path: String, + val length: Long, val underlyingSplit: Any) { + override def toString(): String = { + "SplitInfo " + super.toString + " .. inputFormatClazz " + inputFormatClazz + + ", hostLocation : " + hostLocation + ", path : " + path + + ", length : " + length + ", underlyingSplit " + underlyingSplit + } + + override def hashCode(): Int = { + var hashCode = inputFormatClazz.hashCode + hashCode = hashCode * 31 + hostLocation.hashCode + hashCode = hashCode * 31 + path.hashCode + // ignore overflow ? It is hashcode anyway ! + hashCode = hashCode * 31 + (length & 0x7fffffff).toInt + hashCode + } + + // This is practically useless since most of the Split impl's dont seem to implement equals :-( + // So unless there is identity equality between underlyingSplits, it will always fail even if it + // is pointing to same block. + override def equals(other: Any): Boolean = other match { + case that: SplitInfo => { + this.hostLocation == that.hostLocation && + this.inputFormatClazz == that.inputFormatClazz && + this.path == that.path && + this.length == that.length && + // other split specific checks (like start for FileSplit) + this.underlyingSplit == that.underlyingSplit + } + case _ => false + } +} + +object SplitInfo { + + def toSplitInfo(inputFormatClazz: Class[_], path: String, + mapredSplit: org.apache.hadoop.mapred.InputSplit): Seq[SplitInfo] = { + val retval = new ArrayBuffer[SplitInfo]() + val length = mapredSplit.getLength + for (host <- mapredSplit.getLocations) { + retval += new SplitInfo(inputFormatClazz, host, path, length, mapredSplit) + } + retval + } + + def toSplitInfo(inputFormatClazz: Class[_], path: String, + mapreduceSplit: org.apache.hadoop.mapreduce.InputSplit): Seq[SplitInfo] = { + val retval = new ArrayBuffer[SplitInfo]() + val length = mapreduceSplit.getLength + for (host <- mapreduceSplit.getLocations) { + retval += new SplitInfo(inputFormatClazz, host, path, length, mapreduceSplit) + } + retval + } +} diff --git a/core/src/main/scala/spark/scheduler/Stage.scala b/core/src/main/scala/spark/scheduler/Stage.scala index 552061e46b..7fc9e13fd9 100644 --- a/core/src/main/scala/spark/scheduler/Stage.scala +++ b/core/src/main/scala/spark/scheduler/Stage.scala @@ -26,7 +26,7 @@ private[spark] class Stage( val parents: List[Stage], val priority: Int) extends Logging { - + val isShuffleMap = shuffleDep != None val numPartitions = rdd.partitions.size val outputLocs = Array.fill[List[MapStatus]](numPartitions)(Nil) @@ -60,7 +60,7 @@ private[spark] class Stage( numAvailableOutputs -= 1 } } - + def removeOutputsOnExecutor(execId: String) { var becameUnavailable = false for (partition <- 0 until numPartitions) { diff --git a/core/src/main/scala/spark/scheduler/StageInfo.scala b/core/src/main/scala/spark/scheduler/StageInfo.scala new file mode 100644 index 0000000000..8d83ff10c4 --- /dev/null +++ b/core/src/main/scala/spark/scheduler/StageInfo.scala @@ -0,0 +1,12 @@ +package spark.scheduler + +import spark.scheduler.cluster.TaskInfo +import scala.collection._ +import spark.executor.TaskMetrics + +case class StageInfo( + val stage: Stage, + val taskInfos: mutable.Buffer[(TaskInfo, TaskMetrics)] = mutable.Buffer[(TaskInfo, TaskMetrics)]() +) { + override def toString = stage.rdd.toString +}
\ No newline at end of file diff --git a/core/src/main/scala/spark/scheduler/Task.scala b/core/src/main/scala/spark/scheduler/Task.scala index ef987fdeb6..a6462c6968 100644 --- a/core/src/main/scala/spark/scheduler/Task.scala +++ b/core/src/main/scala/spark/scheduler/Task.scala @@ -1,12 +1,12 @@ package spark.scheduler -import scala.collection.mutable.HashMap -import spark.serializer.{SerializerInstance, Serializer} +import spark.serializer.SerializerInstance import java.io.{DataInputStream, DataOutputStream} import java.nio.ByteBuffer import it.unimi.dsi.fastutil.io.FastByteArrayOutputStream import spark.util.ByteBufferInputStream import scala.collection.mutable.HashMap +import spark.executor.TaskMetrics /** * A task to execute on a worker node. @@ -16,6 +16,9 @@ private[spark] abstract class Task[T](val stageId: Int) extends Serializable { def preferredLocations: Seq[String] = Nil var generation: Long = -1 // Map output tracker generation. Will be set by TaskScheduler. + + var metrics: Option[TaskMetrics] = None + } /** diff --git a/core/src/main/scala/spark/scheduler/TaskResult.scala b/core/src/main/scala/spark/scheduler/TaskResult.scala index 9a54d0e854..6de0aa7adf 100644 --- a/core/src/main/scala/spark/scheduler/TaskResult.scala +++ b/core/src/main/scala/spark/scheduler/TaskResult.scala @@ -3,13 +3,14 @@ package spark.scheduler import java.io._ import scala.collection.mutable.Map +import spark.executor.TaskMetrics // Task result. Also contains updates to accumulator variables. // TODO: Use of distributed cache to return result is a hack to get around // what seems to be a bug with messages over 60KB in libprocess; fix it private[spark] -class TaskResult[T](var value: T, var accumUpdates: Map[Long, Any]) extends Externalizable { - def this() = this(null.asInstanceOf[T], null) +class TaskResult[T](var value: T, var accumUpdates: Map[Long, Any], var metrics: TaskMetrics) extends Externalizable { + def this() = this(null.asInstanceOf[T], null, null) override def writeExternal(out: ObjectOutput) { out.writeObject(value) @@ -18,6 +19,7 @@ class TaskResult[T](var value: T, var accumUpdates: Map[Long, Any]) extends Exte out.writeLong(key) out.writeObject(value) } + out.writeObject(metrics) } override def readExternal(in: ObjectInput) { @@ -31,5 +33,6 @@ class TaskResult[T](var value: T, var accumUpdates: Map[Long, Any]) extends Exte accumUpdates(in.readLong()) = in.readObject() } } + metrics = in.readObject().asInstanceOf[TaskMetrics] } } diff --git a/core/src/main/scala/spark/scheduler/TaskScheduler.scala b/core/src/main/scala/spark/scheduler/TaskScheduler.scala index d549b184b0..7787b54762 100644 --- a/core/src/main/scala/spark/scheduler/TaskScheduler.scala +++ b/core/src/main/scala/spark/scheduler/TaskScheduler.scala @@ -10,6 +10,10 @@ package spark.scheduler private[spark] trait TaskScheduler { def start(): Unit + // Invoked after system has successfully initialized (typically in spark context). + // Yarn uses this to bootstrap allocation of resources based on preferred locations, wait for slave registerations, etc. + def postStartHook() { } + // Disconnect from the cluster. def stop(): Unit diff --git a/core/src/main/scala/spark/scheduler/TaskSchedulerListener.scala b/core/src/main/scala/spark/scheduler/TaskSchedulerListener.scala index 9fcef86e46..b75d3736cf 100644 --- a/core/src/main/scala/spark/scheduler/TaskSchedulerListener.scala +++ b/core/src/main/scala/spark/scheduler/TaskSchedulerListener.scala @@ -1,15 +1,21 @@ package spark.scheduler +import spark.scheduler.cluster.TaskInfo import scala.collection.mutable.Map import spark.TaskEndReason +import spark.executor.TaskMetrics /** * Interface for getting events back from the TaskScheduler. */ private[spark] trait TaskSchedulerListener { // A task has finished or failed. - def taskEnded(task: Task[_], reason: TaskEndReason, result: Any, accumUpdates: Map[Long, Any]): Unit + def taskEnded(task: Task[_], reason: TaskEndReason, result: Any, accumUpdates: Map[Long, Any], + taskInfo: TaskInfo, taskMetrics: TaskMetrics): Unit + + // A node was added to the cluster. + def executorGained(execId: String, hostPort: String): Unit // A node was lost from the cluster. def executorLost(execId: String): Unit diff --git a/core/src/main/scala/spark/scheduler/TaskSet.scala b/core/src/main/scala/spark/scheduler/TaskSet.scala index a3002ca477..e4b5fcaedb 100644 --- a/core/src/main/scala/spark/scheduler/TaskSet.scala +++ b/core/src/main/scala/spark/scheduler/TaskSet.scala @@ -1,11 +1,18 @@ package spark.scheduler +import java.util.Properties + /** * A set of tasks submitted together to the low-level TaskScheduler, usually representing * missing partitions of a particular stage. */ -private[spark] class TaskSet(val tasks: Array[Task[_]], val stageId: Int, val attempt: Int, val priority: Int) { - val id: String = stageId + "." + attempt +private[spark] class TaskSet( + val tasks: Array[Task[_]], + val stageId: Int, + val attempt: Int, + val priority: Int, + val properties: Properties) { + val id: String = stageId + "." + attempt override def toString: String = "TaskSet " + id } diff --git a/core/src/main/scala/spark/scheduler/cluster/ClusterScheduler.scala b/core/src/main/scala/spark/scheduler/cluster/ClusterScheduler.scala index 1e4fbdb874..3a0c29b27f 100644 --- a/core/src/main/scala/spark/scheduler/cluster/ClusterScheduler.scala +++ b/core/src/main/scala/spark/scheduler/cluster/ClusterScheduler.scala @@ -1,6 +1,6 @@ package spark.scheduler.cluster -import java.io.{File, FileInputStream, FileOutputStream} +import java.lang.{Boolean => JBoolean} import scala.collection.mutable.ArrayBuffer import scala.collection.mutable.HashMap @@ -11,6 +11,7 @@ import spark.TaskState.TaskState import spark.scheduler._ import java.nio.ByteBuffer import java.util.concurrent.atomic.AtomicLong +import java.util.{TimerTask, Timer} /** * The main TaskScheduler implementation, for running tasks on a cluster. Clients should first call @@ -22,25 +23,64 @@ private[spark] class ClusterScheduler(val sc: SparkContext) // How often to check for speculative tasks val SPECULATION_INTERVAL = System.getProperty("spark.speculation.interval", "100").toLong + // Threshold above which we warn user initial TaskSet may be starved + val STARVATION_TIMEOUT = System.getProperty("spark.starvation.timeout", "15000").toLong + // How often to revive offers in case there are pending tasks - that is how often to try to get + // tasks scheduled in case there are nodes available : default 0 is to disable it - to preserve existing behavior + // Note that this is required due to delayed scheduling due to data locality waits, etc. + // TODO: rename property ? + val TASK_REVIVAL_INTERVAL = System.getProperty("spark.tasks.revive.interval", "0").toLong + + /* + This property controls how aggressive we should be to modulate waiting for node local task scheduling. + To elaborate, currently there is a time limit (3 sec def) to ensure that spark attempts to wait for node locality of tasks before + scheduling on other nodes. We have modified this in yarn branch such that offers to task set happen in prioritized order : + node-local, rack-local and then others + But once all available node local (and no pref) tasks are scheduled, instead of waiting for 3 sec before + scheduling to other nodes (which degrades performance for time sensitive tasks and on larger clusters), we can + modulate that : to also allow rack local nodes or any node. The default is still set to HOST - so that previous behavior is + maintained. This is to allow tuning the tension between pulling rdd data off node and scheduling computation asap. + + TODO: rename property ? The value is one of + - NODE_LOCAL (default, no change w.r.t current behavior), + - RACK_LOCAL and + - ANY + + Note that this property makes more sense when used in conjugation with spark.tasks.revive.interval > 0 : else it is not very effective. + + Additional Note: For non trivial clusters, there is a 4x - 5x reduction in running time (in some of our experiments) based on whether + it is left at default NODE_LOCAL, RACK_LOCAL (if cluster is configured to be rack aware) or ANY. + If cluster is rack aware, then setting it to RACK_LOCAL gives best tradeoff and a 3x - 4x performance improvement while minimizing IO impact. + Also, it brings down the variance in running time drastically. + */ + val TASK_SCHEDULING_AGGRESSION = TaskLocality.parse(System.getProperty("spark.tasks.schedule.aggression", "NODE_LOCAL")) val activeTaskSets = new HashMap[String, TaskSetManager] - var activeTaskSetsQueue = new ArrayBuffer[TaskSetManager] val taskIdToTaskSetId = new HashMap[Long, String] val taskIdToExecutorId = new HashMap[Long, String] val taskSetTaskIds = new HashMap[String, HashSet[Long]] + @volatile private var hasReceivedTask = false + @volatile private var hasLaunchedTask = false + private val starvationTimer = new Timer(true) + // Incrementing Mesos task IDs val nextTaskId = new AtomicLong(0) // Which executor IDs we have executors on val activeExecutorIds = new HashSet[String] + // TODO: We might want to remove this and merge it with execId datastructures - but later. + // Which hosts in the cluster are alive (contains hostPort's) - used for process local and node local task locality. + private val hostPortsAlive = new HashSet[String] + private val hostToAliveHostPorts = new HashMap[String, HashSet[String]] + // The set of executors we have on each host; this is used to compute hostsAlive, which // in turn is used to decide when we can attain data locality on a given host - val executorsByHost = new HashMap[String, HashSet[String]] + private val executorsByHostPort = new HashMap[String, HashSet[String]] - val executorIdToHost = new HashMap[String, String] + private val executorIdToHostPort = new HashMap[String, String] // JAR server, if any JARs were added by the user to the SparkContext var jarServer: HttpServer = null @@ -55,24 +95,50 @@ private[spark] class ClusterScheduler(val sc: SparkContext) val mapOutputTracker = SparkEnv.get.mapOutputTracker + var schedulableBuilder: SchedulableBuilder = null + var rootPool: Pool = null + override def setListener(listener: TaskSchedulerListener) { this.listener = listener } def initialize(context: SchedulerBackend) { backend = context + //default scheduler is FIFO + val schedulingMode = System.getProperty("spark.cluster.schedulingmode", "FIFO") + //temporarily set rootPool name to empty + rootPool = new Pool("", SchedulingMode.withName(schedulingMode), 0, 0) + schedulableBuilder = { + schedulingMode match { + case "FIFO" => + new FIFOSchedulableBuilder(rootPool) + case "FAIR" => + new FairSchedulableBuilder(rootPool) + } + } + schedulableBuilder.buildPools() + // resolve executorId to hostPort mapping. + def executorToHostPort(executorId: String, defaultHostPort: String): String = { + executorIdToHostPort.getOrElse(executorId, defaultHostPort) + } + + // Unfortunately, this means that SparkEnv is indirectly referencing ClusterScheduler + // Will that be a design violation ? + SparkEnv.get.executorIdToHostPort = Some(executorToHostPort) } + def newTaskId(): Long = nextTaskId.getAndIncrement() override def start() { backend.start() - if (System.getProperty("spark.speculation", "false") == "true") { + if (JBoolean.getBoolean("spark.speculation")) { new Thread("ClusterScheduler speculation check") { setDaemon(true) override def run() { + logInfo("Starting speculative execution thread") while (true) { try { Thread.sleep(SPECULATION_INTERVAL) @@ -84,16 +150,51 @@ private[spark] class ClusterScheduler(val sc: SparkContext) } }.start() } + + + // Change to always run with some default if TASK_REVIVAL_INTERVAL <= 0 ? + if (TASK_REVIVAL_INTERVAL > 0) { + new Thread("ClusterScheduler task offer revival check") { + setDaemon(true) + + override def run() { + logInfo("Starting speculative task offer revival thread") + while (true) { + try { + Thread.sleep(TASK_REVIVAL_INTERVAL) + } catch { + case e: InterruptedException => {} + } + + if (hasPendingTasks()) backend.reviveOffers() + } + } + }.start() + } } override def submitTasks(taskSet: TaskSet) { val tasks = taskSet.tasks logInfo("Adding task set " + taskSet.id + " with " + tasks.length + " tasks") this.synchronized { - val manager = new TaskSetManager(this, taskSet) + val manager = new ClusterTaskSetManager(this, taskSet) activeTaskSets(taskSet.id) = manager - activeTaskSetsQueue += manager + schedulableBuilder.addTaskSetManager(manager, manager.taskSet.properties) taskSetTaskIds(taskSet.id) = new HashSet[Long]() + + if (hasReceivedTask == false) { + starvationTimer.scheduleAtFixedRate(new TimerTask() { + override def run() { + if (!hasLaunchedTask) { + logWarning("Initial job has not accepted any resources; " + + "check your cluster UI to ensure that workers are registered") + } else { + this.cancel() + } + } + }, STARVATION_TIMEOUT, STARVATION_TIMEOUT) + } + hasReceivedTask = true; } backend.reviveOffers() } @@ -101,7 +202,8 @@ private[spark] class ClusterScheduler(val sc: SparkContext) def taskSetFinished(manager: TaskSetManager) { this.synchronized { activeTaskSets -= manager.taskSet.id - activeTaskSetsQueue -= manager + manager.parent.removeSchedulable(manager) + logInfo("Remove TaskSet %s from pool %s".format(manager.taskSet.id, manager.parent.name)) taskIdToTaskSetId --= taskSetTaskIds(manager.taskSet.id) taskIdToExecutorId --= taskSetTaskIds(manager.taskSet.id) taskSetTaskIds.remove(manager.taskSet.id) @@ -118,19 +220,128 @@ private[spark] class ClusterScheduler(val sc: SparkContext) SparkEnv.set(sc.env) // Mark each slave as alive and remember its hostname for (o <- offers) { - executorIdToHost(o.executorId) = o.hostname + // DEBUG Code + Utils.checkHostPort(o.hostPort) + + executorIdToHostPort(o.executorId) = o.hostPort + if (! executorsByHostPort.contains(o.hostPort)) { + executorsByHostPort(o.hostPort) = new HashSet[String]() + } + + hostPortsAlive += o.hostPort + hostToAliveHostPorts.getOrElseUpdate(Utils.parseHostPort(o.hostPort)._1, new HashSet[String]).add(o.hostPort) + executorGained(o.executorId, o.hostPort) } // Build a list of tasks to assign to each slave val tasks = offers.map(o => new ArrayBuffer[TaskDescription](o.cores)) + // merge availableCpus into nodeToAvailableCpus block ? val availableCpus = offers.map(o => o.cores).toArray + val nodeToAvailableCpus = { + val map = new HashMap[String, Int]() + for (offer <- offers) { + val hostPort = offer.hostPort + val cores = offer.cores + // DEBUG code + Utils.checkHostPort(hostPort) + + val host = Utils.parseHostPort(hostPort)._1 + + map.put(host, map.getOrElse(host, 0) + cores) + } + + map + } var launchedTask = false - for (manager <- activeTaskSetsQueue.sortBy(m => (m.taskSet.priority, m.taskSet.stageId))) { + val sortedTaskSetQueue = rootPool.getSortedTaskSetQueue() + for (manager <- sortedTaskSetQueue) + { + logInfo("parentName:%s,name:%s,runningTasks:%s".format(manager.parent.name, manager.name, manager.runningTasks)) + } + for (manager <- sortedTaskSetQueue) { + + // Split offers based on node local, rack local and off-rack tasks. + val processLocalOffers = new HashMap[String, ArrayBuffer[Int]]() + val nodeLocalOffers = new HashMap[String, ArrayBuffer[Int]]() + val rackLocalOffers = new HashMap[String, ArrayBuffer[Int]]() + val otherOffers = new HashMap[String, ArrayBuffer[Int]]() + + for (i <- 0 until offers.size) { + val hostPort = offers(i).hostPort + // DEBUG code + Utils.checkHostPort(hostPort) + + val numProcessLocalTasks = math.max(0, math.min(manager.numPendingTasksForHostPort(hostPort), availableCpus(i))) + if (numProcessLocalTasks > 0){ + val list = processLocalOffers.getOrElseUpdate(hostPort, new ArrayBuffer[Int]) + for (j <- 0 until numProcessLocalTasks) list += i + } + + val host = Utils.parseHostPort(hostPort)._1 + val numNodeLocalTasks = math.max(0, + // Remove process local tasks (which are also host local btw !) from this + math.min(manager.numPendingTasksForHost(hostPort) - numProcessLocalTasks, nodeToAvailableCpus(host))) + if (numNodeLocalTasks > 0){ + val list = nodeLocalOffers.getOrElseUpdate(host, new ArrayBuffer[Int]) + for (j <- 0 until numNodeLocalTasks) list += i + } + + val numRackLocalTasks = math.max(0, + // Remove node local tasks (which are also rack local btw !) from this + math.min(manager.numRackLocalPendingTasksForHost(hostPort) - numProcessLocalTasks - numNodeLocalTasks, nodeToAvailableCpus(host))) + if (numRackLocalTasks > 0){ + val list = rackLocalOffers.getOrElseUpdate(host, new ArrayBuffer[Int]) + for (j <- 0 until numRackLocalTasks) list += i + } + if (numNodeLocalTasks <= 0 && numRackLocalTasks <= 0){ + // add to others list - spread even this across cluster. + val list = otherOffers.getOrElseUpdate(host, new ArrayBuffer[Int]) + list += i + } + } + + val offersPriorityList = new ArrayBuffer[Int]( + processLocalOffers.size + nodeLocalOffers.size + rackLocalOffers.size + otherOffers.size) + + // First process local, then host local, then rack, then others + + // numNodeLocalOffers contains count of both process local and host offers. + val numNodeLocalOffers = { + val processLocalPriorityList = ClusterScheduler.prioritizeContainers(processLocalOffers) + offersPriorityList ++= processLocalPriorityList + + val nodeLocalPriorityList = ClusterScheduler.prioritizeContainers(nodeLocalOffers) + offersPriorityList ++= nodeLocalPriorityList + + processLocalPriorityList.size + nodeLocalPriorityList.size + } + val numRackLocalOffers = { + val rackLocalPriorityList = ClusterScheduler.prioritizeContainers(rackLocalOffers) + offersPriorityList ++= rackLocalPriorityList + rackLocalPriorityList.size + } + offersPriorityList ++= ClusterScheduler.prioritizeContainers(otherOffers) + + var lastLoop = false + val lastLoopIndex = TASK_SCHEDULING_AGGRESSION match { + case TaskLocality.NODE_LOCAL => numNodeLocalOffers + case TaskLocality.RACK_LOCAL => numRackLocalOffers + numNodeLocalOffers + case TaskLocality.ANY => offersPriorityList.size + } + do { launchedTask = false - for (i <- 0 until offers.size) { + var loopCount = 0 + for (i <- offersPriorityList) { val execId = offers(i).executorId - val host = offers(i).hostname - manager.slaveOffer(execId, host, availableCpus(i)) match { + val hostPort = offers(i).hostPort + + // If last loop and within the lastLoopIndex, expand scope - else use null (which will use default/existing) + val overrideLocality = if (lastLoop && loopCount < lastLoopIndex) TASK_SCHEDULING_AGGRESSION else null + + // If last loop, override waiting for host locality - we scheduled all local tasks already and there might be more available ... + loopCount += 1 + + manager.slaveOffer(execId, hostPort, availableCpus(i), overrideLocality) match { case Some(task) => tasks(i) += task val tid = task.taskId @@ -138,18 +349,34 @@ private[spark] class ClusterScheduler(val sc: SparkContext) taskSetTaskIds(manager.taskSet.id) += tid taskIdToExecutorId(tid) = execId activeExecutorIds += execId - if (!executorsByHost.contains(host)) { - executorsByHost(host) = new HashSet() - } - executorsByHost(host) += execId + executorsByHostPort(hostPort) += execId availableCpus(i) -= 1 launchedTask = true case None => {} + } + } + // Loop once more - when lastLoop = true, then we try to schedule task on all nodes irrespective of + // data locality (we still go in order of priority : but that would not change anything since + // if data local tasks had been available, we would have scheduled them already) + if (lastLoop) { + // prevent more looping + launchedTask = false + } else if (!lastLoop && !launchedTask) { + // Do this only if TASK_SCHEDULING_AGGRESSION != NODE_LOCAL + if (TASK_SCHEDULING_AGGRESSION != TaskLocality.NODE_LOCAL) { + // fudge launchedTask to ensure we loop once more + launchedTask = true + // dont loop anymore + lastLoop = true } } } while (launchedTask) } + + if (tasks.size > 0) { + hasLaunchedTask = true + } return tasks } } @@ -199,6 +426,7 @@ private[spark] class ClusterScheduler(val sc: SparkContext) backend.reviveOffers() } if (taskFailed) { + // Also revive offers if a task had failed for some reason other than host lost backend.reviveOffers() } @@ -232,29 +460,40 @@ private[spark] class ClusterScheduler(val sc: SparkContext) if (jarServer != null) { jarServer.stop() } + + // sleeping for an arbitrary 5 seconds : to ensure that messages are sent out. + // TODO: Do something better ! + Thread.sleep(5000L) } override def defaultParallelism() = backend.defaultParallelism() - + + // Check for speculatable tasks in all our active jobs. def checkSpeculatableTasks() { var shouldRevive = false synchronized { - for (ts <- activeTaskSetsQueue) { - shouldRevive |= ts.checkSpeculatableTasks() - } + shouldRevive = rootPool.checkSpeculatableTasks() } if (shouldRevive) { backend.reviveOffers() } } + // Check for pending tasks in all our active jobs. + def hasPendingTasks(): Boolean = { + synchronized { + rootPool.hasPendingTasks() + } + } + def executorLost(executorId: String, reason: ExecutorLossReason) { var failedExecutor: Option[String] = None + synchronized { if (activeExecutorIds.contains(executorId)) { - val host = executorIdToHost(executorId) - logError("Lost executor %s on %s: %s".format(executorId, host, reason)) + val hostPort = executorIdToHostPort(executorId) + logError("Lost executor %s on %s: %s".format(executorId, hostPort, reason)) removeExecutor(executorId) failedExecutor = Some(executorId) } else { @@ -272,19 +511,104 @@ private[spark] class ClusterScheduler(val sc: SparkContext) } } - /** Get a list of hosts that currently have executors */ - def hostsAlive: scala.collection.Set[String] = executorsByHost.keySet - /** Remove an executor from all our data structures and mark it as lost */ private def removeExecutor(executorId: String) { activeExecutorIds -= executorId - val host = executorIdToHost(executorId) - val execs = executorsByHost.getOrElse(host, new HashSet) + val hostPort = executorIdToHostPort(executorId) + if (hostPortsAlive.contains(hostPort)) { + // DEBUG Code + Utils.checkHostPort(hostPort) + + hostPortsAlive -= hostPort + hostToAliveHostPorts.getOrElseUpdate(Utils.parseHostPort(hostPort)._1, new HashSet[String]).remove(hostPort) + } + + val execs = executorsByHostPort.getOrElse(hostPort, new HashSet) execs -= executorId if (execs.isEmpty) { - executorsByHost -= host + executorsByHostPort -= hostPort + } + executorIdToHostPort -= executorId + rootPool.executorLost(executorId, hostPort) + } + + def executorGained(execId: String, hostPort: String) { + listener.executorGained(execId, hostPort) + } + + def getExecutorsAliveOnHost(host: String): Option[Set[String]] = { + Utils.checkHost(host) + + val retval = hostToAliveHostPorts.get(host) + if (retval.isDefined) { + return Some(retval.get.toSet) } - executorIdToHost -= executorId - activeTaskSetsQueue.foreach(_.executorLost(executorId, host)) + + None + } + + def isExecutorAliveOnHostPort(hostPort: String): Boolean = { + // Even if hostPort is a host, it does not matter - it is just a specific check. + // But we do have to ensure that only hostPort get into hostPortsAlive ! + // So no check against Utils.checkHostPort + hostPortsAlive.contains(hostPort) + } + + // By default, rack is unknown + def getRackForHost(value: String): Option[String] = None + + // By default, (cached) hosts for rack is unknown + def getCachedHostsForRack(rack: String): Option[Set[String]] = None +} + + +object ClusterScheduler { + + // Used to 'spray' available containers across the available set to ensure too many containers on same host + // are not used up. Used in yarn mode and in task scheduling (when there are multiple containers available + // to execute a task) + // For example: yarn can returns more containers than we would have requested under ANY, this method + // prioritizes how to use the allocated containers. + // flatten the map such that the array buffer entries are spread out across the returned value. + // given <host, list[container]> == <h1, [c1 .. c5]>, <h2, [c1 .. c3]>, <h3, [c1, c2]>, <h4, c1>, <h5, c1>, i + // the return value would be something like : h1c1, h2c1, h3c1, h4c1, h5c1, h1c2, h2c2, h3c2, h1c3, h2c3, h1c4, h1c5 + // We then 'use' the containers in this order (consuming only the top K from this list where + // K = number to be user). This is to ensure that if we have multiple eligible allocations, + // they dont end up allocating all containers on a small number of hosts - increasing probability of + // multiple container failure when a host goes down. + // Note, there is bias for keys with higher number of entries in value to be picked first (by design) + // Also note that invocation of this method is expected to have containers of same 'type' + // (host-local, rack-local, off-rack) and not across types : so that reordering is simply better from + // the available list - everything else being same. + // That is, we we first consume data local, then rack local and finally off rack nodes. So the + // prioritization from this method applies to within each category + def prioritizeContainers[K, T] (map: HashMap[K, ArrayBuffer[T]]): List[T] = { + val _keyList = new ArrayBuffer[K](map.size) + _keyList ++= map.keys + + // order keyList based on population of value in map + val keyList = _keyList.sortWith( + (left, right) => map.get(left).getOrElse(Set()).size > map.get(right).getOrElse(Set()).size + ) + + val retval = new ArrayBuffer[T](keyList.size * 2) + var index = 0 + var found = true + + while (found){ + found = false + for (key <- keyList) { + val containerList: ArrayBuffer[T] = map.get(key).getOrElse(null) + assert(containerList != null) + // Get the index'th entry for this host - if present + if (index < containerList.size){ + retval += containerList.apply(index) + found = true + } + } + index += 1 + } + + retval.toList } } diff --git a/core/src/main/scala/spark/scheduler/cluster/ClusterTaskSetManager.scala b/core/src/main/scala/spark/scheduler/cluster/ClusterTaskSetManager.scala new file mode 100644 index 0000000000..d72b0bfc9f --- /dev/null +++ b/core/src/main/scala/spark/scheduler/cluster/ClusterTaskSetManager.scala @@ -0,0 +1,747 @@ +package spark.scheduler.cluster + +import java.util.{HashMap => JHashMap, NoSuchElementException, Arrays} + +import scala.collection.mutable.ArrayBuffer +import scala.collection.mutable.HashMap +import scala.collection.mutable.HashSet +import scala.math.max +import scala.math.min + +import spark._ +import spark.scheduler._ +import spark.TaskState.TaskState +import java.nio.ByteBuffer + +private[spark] object TaskLocality extends Enumeration("PROCESS_LOCAL", "NODE_LOCAL", "RACK_LOCAL", "ANY") with Logging { + + // process local is expected to be used ONLY within tasksetmanager for now. + val PROCESS_LOCAL, NODE_LOCAL, RACK_LOCAL, ANY = Value + + type TaskLocality = Value + + def isAllowed(constraint: TaskLocality, condition: TaskLocality): Boolean = { + + // Must not be the constraint. + assert (constraint != TaskLocality.PROCESS_LOCAL) + + constraint match { + case TaskLocality.NODE_LOCAL => condition == TaskLocality.NODE_LOCAL + case TaskLocality.RACK_LOCAL => condition == TaskLocality.NODE_LOCAL || condition == TaskLocality.RACK_LOCAL + // For anything else, allow + case _ => true + } + } + + def parse(str: String): TaskLocality = { + // better way to do this ? + try { + val retval = TaskLocality.withName(str) + // Must not specify PROCESS_LOCAL ! + assert (retval != TaskLocality.PROCESS_LOCAL) + + retval + } catch { + case nEx: NoSuchElementException => { + logWarning("Invalid task locality specified '" + str + "', defaulting to NODE_LOCAL"); + // default to preserve earlier behavior + NODE_LOCAL + } + } + } +} + +/** + * Schedules the tasks within a single TaskSet in the ClusterScheduler. + */ +private[spark] class ClusterTaskSetManager( + sched: ClusterScheduler, + val taskSet: TaskSet) + extends TaskSetManager + with Logging { + + // Maximum time to wait to run a task in a preferred location (in ms) + val LOCALITY_WAIT = System.getProperty("spark.locality.wait", "3000").toLong + + // CPUs to request per task + val CPUS_PER_TASK = System.getProperty("spark.task.cpus", "1").toDouble + + // Maximum times a task is allowed to fail before failing the job + val MAX_TASK_FAILURES = 4 + + // Quantile of tasks at which to start speculation + val SPECULATION_QUANTILE = System.getProperty("spark.speculation.quantile", "0.75").toDouble + val SPECULATION_MULTIPLIER = System.getProperty("spark.speculation.multiplier", "1.5").toDouble + + // Serializer for closures and tasks. + val ser = SparkEnv.get.closureSerializer.newInstance() + + val tasks = taskSet.tasks + val numTasks = tasks.length + val copiesRunning = new Array[Int](numTasks) + val finished = new Array[Boolean](numTasks) + val numFailures = new Array[Int](numTasks) + val taskAttempts = Array.fill[List[TaskInfo]](numTasks)(Nil) + var tasksFinished = 0 + + var weight = 1 + var minShare = 0 + var runningTasks = 0 + var priority = taskSet.priority + var stageId = taskSet.stageId + var name = "TaskSet_"+taskSet.stageId.toString + var parent:Schedulable = null + + // Last time when we launched a preferred task (for delay scheduling) + var lastPreferredLaunchTime = System.currentTimeMillis + + // List of pending tasks for each node (process local to container). These collections are actually + // treated as stacks, in which new tasks are added to the end of the + // ArrayBuffer and removed from the end. This makes it faster to detect + // tasks that repeatedly fail because whenever a task failed, it is put + // back at the head of the stack. They are also only cleaned up lazily; + // when a task is launched, it remains in all the pending lists except + // the one that it was launched from, but gets removed from them later. + private val pendingTasksForHostPort = new HashMap[String, ArrayBuffer[Int]] + + // List of pending tasks for each node. + // Essentially, similar to pendingTasksForHostPort, except at host level + private val pendingTasksForHost = new HashMap[String, ArrayBuffer[Int]] + + // List of pending tasks for each node based on rack locality. + // Essentially, similar to pendingTasksForHost, except at rack level + private val pendingRackLocalTasksForHost = new HashMap[String, ArrayBuffer[Int]] + + // List containing pending tasks with no locality preferences + val pendingTasksWithNoPrefs = new ArrayBuffer[Int] + + // List containing all pending tasks (also used as a stack, as above) + val allPendingTasks = new ArrayBuffer[Int] + + // Tasks that can be speculated. Since these will be a small fraction of total + // tasks, we'll just hold them in a HashSet. + val speculatableTasks = new HashSet[Int] + + // Task index, start and finish time for each task attempt (indexed by task ID) + val taskInfos = new HashMap[Long, TaskInfo] + + // Did the job fail? + var failed = false + var causeOfFailure = "" + + // How frequently to reprint duplicate exceptions in full, in milliseconds + val EXCEPTION_PRINT_INTERVAL = + System.getProperty("spark.logging.exceptionPrintInterval", "10000").toLong + // Map of recent exceptions (identified by string representation and + // top stack frame) to duplicate count (how many times the same + // exception has appeared) and time the full exception was + // printed. This should ideally be an LRU map that can drop old + // exceptions automatically. + val recentExceptions = HashMap[String, (Int, Long)]() + + // Figure out the current map output tracker generation and set it on all tasks + val generation = sched.mapOutputTracker.getGeneration + logDebug("Generation for " + taskSet.id + ": " + generation) + for (t <- tasks) { + t.generation = generation + } + + // Add all our tasks to the pending lists. We do this in reverse order + // of task index so that tasks with low indices get launched first. + for (i <- (0 until numTasks).reverse) { + addPendingTask(i) + } + + // Note that it follows the hierarchy. + // if we search for NODE_LOCAL, the output will include PROCESS_LOCAL and + // if we search for RACK_LOCAL, it will include PROCESS_LOCAL & NODE_LOCAL + private def findPreferredLocations(_taskPreferredLocations: Seq[String], scheduler: ClusterScheduler, + taskLocality: TaskLocality.TaskLocality): HashSet[String] = { + + if (TaskLocality.PROCESS_LOCAL == taskLocality) { + // straight forward comparison ! Special case it. + val retval = new HashSet[String]() + scheduler.synchronized { + for (location <- _taskPreferredLocations) { + if (scheduler.isExecutorAliveOnHostPort(location)) { + retval += location + } + } + } + + return retval + } + + val taskPreferredLocations = + if (TaskLocality.NODE_LOCAL == taskLocality) { + _taskPreferredLocations + } else { + assert (TaskLocality.RACK_LOCAL == taskLocality) + // Expand set to include all 'seen' rack local hosts. + // This works since container allocation/management happens within master - so any rack locality information is updated in msater. + // Best case effort, and maybe sort of kludge for now ... rework it later ? + val hosts = new HashSet[String] + _taskPreferredLocations.foreach(h => { + val rackOpt = scheduler.getRackForHost(h) + if (rackOpt.isDefined) { + val hostsOpt = scheduler.getCachedHostsForRack(rackOpt.get) + if (hostsOpt.isDefined) { + hosts ++= hostsOpt.get + } + } + + // Ensure that irrespective of what scheduler says, host is always added ! + hosts += h + }) + + hosts + } + + val retval = new HashSet[String] + scheduler.synchronized { + for (prefLocation <- taskPreferredLocations) { + val aliveLocationsOpt = scheduler.getExecutorsAliveOnHost(Utils.parseHostPort(prefLocation)._1) + if (aliveLocationsOpt.isDefined) { + retval ++= aliveLocationsOpt.get + } + } + } + + retval + } + + // Add a task to all the pending-task lists that it should be on. + private def addPendingTask(index: Int) { + // We can infer hostLocalLocations from rackLocalLocations by joining it against tasks(index).preferredLocations (with appropriate + // hostPort <-> host conversion). But not doing it for simplicity sake. If this becomes a performance issue, modify it. + val processLocalLocations = findPreferredLocations(tasks(index).preferredLocations, sched, TaskLocality.PROCESS_LOCAL) + val hostLocalLocations = findPreferredLocations(tasks(index).preferredLocations, sched, TaskLocality.NODE_LOCAL) + val rackLocalLocations = findPreferredLocations(tasks(index).preferredLocations, sched, TaskLocality.RACK_LOCAL) + + if (rackLocalLocations.size == 0) { + // Current impl ensures this. + assert (processLocalLocations.size == 0) + assert (hostLocalLocations.size == 0) + pendingTasksWithNoPrefs += index + } else { + + // process local locality + for (hostPort <- processLocalLocations) { + // DEBUG Code + Utils.checkHostPort(hostPort) + + val hostPortList = pendingTasksForHostPort.getOrElseUpdate(hostPort, ArrayBuffer()) + hostPortList += index + } + + // host locality (includes process local) + for (hostPort <- hostLocalLocations) { + // DEBUG Code + Utils.checkHostPort(hostPort) + + val host = Utils.parseHostPort(hostPort)._1 + val hostList = pendingTasksForHost.getOrElseUpdate(host, ArrayBuffer()) + hostList += index + } + + // rack locality (includes process local and host local) + for (rackLocalHostPort <- rackLocalLocations) { + // DEBUG Code + Utils.checkHostPort(rackLocalHostPort) + + val rackLocalHost = Utils.parseHostPort(rackLocalHostPort)._1 + val list = pendingRackLocalTasksForHost.getOrElseUpdate(rackLocalHost, ArrayBuffer()) + list += index + } + } + + allPendingTasks += index + } + + // Return the pending tasks list for a given host port (process local), or an empty list if + // there is no map entry for that host + private def getPendingTasksForHostPort(hostPort: String): ArrayBuffer[Int] = { + // DEBUG Code + Utils.checkHostPort(hostPort) + pendingTasksForHostPort.getOrElse(hostPort, ArrayBuffer()) + } + + // Return the pending tasks list for a given host, or an empty list if + // there is no map entry for that host + private def getPendingTasksForHost(hostPort: String): ArrayBuffer[Int] = { + val host = Utils.parseHostPort(hostPort)._1 + pendingTasksForHost.getOrElse(host, ArrayBuffer()) + } + + // Return the pending tasks (rack level) list for a given host, or an empty list if + // there is no map entry for that host + private def getRackLocalPendingTasksForHost(hostPort: String): ArrayBuffer[Int] = { + val host = Utils.parseHostPort(hostPort)._1 + pendingRackLocalTasksForHost.getOrElse(host, ArrayBuffer()) + } + + // Number of pending tasks for a given host Port (which would be process local) + def numPendingTasksForHostPort(hostPort: String): Int = { + getPendingTasksForHostPort(hostPort).count( index => copiesRunning(index) == 0 && !finished(index) ) + } + + // Number of pending tasks for a given host (which would be data local) + def numPendingTasksForHost(hostPort: String): Int = { + getPendingTasksForHost(hostPort).count( index => copiesRunning(index) == 0 && !finished(index) ) + } + + // Number of pending rack local tasks for a given host + def numRackLocalPendingTasksForHost(hostPort: String): Int = { + getRackLocalPendingTasksForHost(hostPort).count( index => copiesRunning(index) == 0 && !finished(index) ) + } + + + // Dequeue a pending task from the given list and return its index. + // Return None if the list is empty. + // This method also cleans up any tasks in the list that have already + // been launched, since we want that to happen lazily. + private def findTaskFromList(list: ArrayBuffer[Int]): Option[Int] = { + while (!list.isEmpty) { + val index = list.last + list.trimEnd(1) + if (copiesRunning(index) == 0 && !finished(index)) { + return Some(index) + } + } + return None + } + + // Return a speculative task for a given host if any are available. The task should not have an + // attempt running on this host, in case the host is slow. In addition, if locality is set, the + // task must have a preference for this host/rack/no preferred locations at all. + private def findSpeculativeTask(hostPort: String, locality: TaskLocality.TaskLocality): Option[Int] = { + + assert (TaskLocality.isAllowed(locality, TaskLocality.NODE_LOCAL)) + speculatableTasks.retain(index => !finished(index)) // Remove finished tasks from set + + if (speculatableTasks.size > 0) { + val localTask = speculatableTasks.find { + index => + val locations = findPreferredLocations(tasks(index).preferredLocations, sched, TaskLocality.NODE_LOCAL) + val attemptLocs = taskAttempts(index).map(_.hostPort) + (locations.size == 0 || locations.contains(hostPort)) && !attemptLocs.contains(hostPort) + } + + if (localTask != None) { + speculatableTasks -= localTask.get + return localTask + } + + // check for rack locality + if (TaskLocality.isAllowed(locality, TaskLocality.RACK_LOCAL)) { + val rackTask = speculatableTasks.find { + index => + val locations = findPreferredLocations(tasks(index).preferredLocations, sched, TaskLocality.RACK_LOCAL) + val attemptLocs = taskAttempts(index).map(_.hostPort) + locations.contains(hostPort) && !attemptLocs.contains(hostPort) + } + + if (rackTask != None) { + speculatableTasks -= rackTask.get + return rackTask + } + } + + // Any task ... + if (TaskLocality.isAllowed(locality, TaskLocality.ANY)) { + // Check for attemptLocs also ? + val nonLocalTask = speculatableTasks.find(i => !taskAttempts(i).map(_.hostPort).contains(hostPort)) + if (nonLocalTask != None) { + speculatableTasks -= nonLocalTask.get + return nonLocalTask + } + } + } + return None + } + + // Dequeue a pending task for a given node and return its index. + // If localOnly is set to false, allow non-local tasks as well. + private def findTask(hostPort: String, locality: TaskLocality.TaskLocality): Option[Int] = { + val processLocalTask = findTaskFromList(getPendingTasksForHostPort(hostPort)) + if (processLocalTask != None) { + return processLocalTask + } + + val localTask = findTaskFromList(getPendingTasksForHost(hostPort)) + if (localTask != None) { + return localTask + } + + if (TaskLocality.isAllowed(locality, TaskLocality.RACK_LOCAL)) { + val rackLocalTask = findTaskFromList(getRackLocalPendingTasksForHost(hostPort)) + if (rackLocalTask != None) { + return rackLocalTask + } + } + + // Look for no pref tasks AFTER rack local tasks - this has side effect that we will get to failed tasks later rather than sooner. + // TODO: That code path needs to be revisited (adding to no prefs list when host:port goes down). + val noPrefTask = findTaskFromList(pendingTasksWithNoPrefs) + if (noPrefTask != None) { + return noPrefTask + } + + if (TaskLocality.isAllowed(locality, TaskLocality.ANY)) { + val nonLocalTask = findTaskFromList(allPendingTasks) + if (nonLocalTask != None) { + return nonLocalTask + } + } + + // Finally, if all else has failed, find a speculative task + return findSpeculativeTask(hostPort, locality) + } + + private def isProcessLocalLocation(task: Task[_], hostPort: String): Boolean = { + Utils.checkHostPort(hostPort) + + val locs = task.preferredLocations + + locs.contains(hostPort) + } + + private def isHostLocalLocation(task: Task[_], hostPort: String): Boolean = { + val locs = task.preferredLocations + + // If no preference, consider it as host local + if (locs.isEmpty) return true + + val host = Utils.parseHostPort(hostPort)._1 + locs.find(h => Utils.parseHostPort(h)._1 == host).isDefined + } + + // Does a host count as a rack local preferred location for a task? (assumes host is NOT preferred location). + // This is true if either the task has preferred locations and this host is one, or it has + // no preferred locations (in which we still count the launch as preferred). + private def isRackLocalLocation(task: Task[_], hostPort: String): Boolean = { + + val locs = task.preferredLocations + + val preferredRacks = new HashSet[String]() + for (preferredHost <- locs) { + val rack = sched.getRackForHost(preferredHost) + if (None != rack) preferredRacks += rack.get + } + + if (preferredRacks.isEmpty) return false + + val hostRack = sched.getRackForHost(hostPort) + + return None != hostRack && preferredRacks.contains(hostRack.get) + } + + // Respond to an offer of a single slave from the scheduler by finding a task + def slaveOffer(execId: String, hostPort: String, availableCpus: Double, overrideLocality: TaskLocality.TaskLocality = null): Option[TaskDescription] = { + + if (tasksFinished < numTasks && availableCpus >= CPUS_PER_TASK) { + // If explicitly specified, use that + val locality = if (overrideLocality != null) overrideLocality else { + // expand only if we have waited for more than LOCALITY_WAIT for a host local task ... + val time = System.currentTimeMillis + if (time - lastPreferredLaunchTime < LOCALITY_WAIT) TaskLocality.NODE_LOCAL else TaskLocality.ANY + } + + findTask(hostPort, locality) match { + case Some(index) => { + // Found a task; do some bookkeeping and return a Mesos task for it + val task = tasks(index) + val taskId = sched.newTaskId() + // Figure out whether this should count as a preferred launch + val taskLocality = + if (isProcessLocalLocation(task, hostPort)) TaskLocality.PROCESS_LOCAL else + if (isHostLocalLocation(task, hostPort)) TaskLocality.NODE_LOCAL else + if (isRackLocalLocation(task, hostPort)) TaskLocality.RACK_LOCAL else + TaskLocality.ANY + val prefStr = taskLocality.toString + logInfo("Starting task %s:%d as TID %s on slave %s: %s (%s)".format( + taskSet.id, index, taskId, execId, hostPort, prefStr)) + // Do various bookkeeping + copiesRunning(index) += 1 + val time = System.currentTimeMillis + val info = new TaskInfo(taskId, index, time, execId, hostPort, taskLocality) + taskInfos(taskId) = info + taskAttempts(index) = info :: taskAttempts(index) + if (TaskLocality.NODE_LOCAL == taskLocality) { + lastPreferredLaunchTime = time + } + // Serialize and return the task + val startTime = System.currentTimeMillis + val serializedTask = Task.serializeWithDependencies( + task, sched.sc.addedFiles, sched.sc.addedJars, ser) + val timeTaken = System.currentTimeMillis - startTime + increaseRunningTasks(1) + logInfo("Serialized task %s:%d as %d bytes in %d ms".format( + taskSet.id, index, serializedTask.limit, timeTaken)) + val taskName = "task %s:%d".format(taskSet.id, index) + return Some(new TaskDescription(taskId, execId, taskName, serializedTask)) + } + case _ => + } + } + return None + } + + def statusUpdate(tid: Long, state: TaskState, serializedData: ByteBuffer) { + state match { + case TaskState.FINISHED => + taskFinished(tid, state, serializedData) + case TaskState.LOST => + taskLost(tid, state, serializedData) + case TaskState.FAILED => + taskLost(tid, state, serializedData) + case TaskState.KILLED => + taskLost(tid, state, serializedData) + case _ => + } + } + + def taskFinished(tid: Long, state: TaskState, serializedData: ByteBuffer) { + val info = taskInfos(tid) + if (info.failed) { + // We might get two task-lost messages for the same task in coarse-grained Mesos mode, + // or even from Mesos itself when acks get delayed. + return + } + val index = info.index + info.markSuccessful() + decreaseRunningTasks(1) + if (!finished(index)) { + tasksFinished += 1 + logInfo("Finished TID %s in %d ms (progress: %d/%d)".format( + tid, info.duration, tasksFinished, numTasks)) + // Deserialize task result and pass it to the scheduler + try { + val result = ser.deserialize[TaskResult[_]](serializedData) + result.metrics.resultSize = serializedData.limit() + sched.listener.taskEnded(tasks(index), Success, result.value, result.accumUpdates, info, result.metrics) + } catch { + case cnf: ClassNotFoundException => + val loader = Thread.currentThread().getContextClassLoader + throw new SparkException("ClassNotFound with classloader: " + loader, cnf) + case ex => throw ex + } + // Mark finished and stop if we've finished all the tasks + finished(index) = true + if (tasksFinished == numTasks) { + sched.taskSetFinished(this) + } + } else { + logInfo("Ignoring task-finished event for TID " + tid + + " because task " + index + " is already finished") + } + } + + def taskLost(tid: Long, state: TaskState, serializedData: ByteBuffer) { + val info = taskInfos(tid) + if (info.failed) { + // We might get two task-lost messages for the same task in coarse-grained Mesos mode, + // or even from Mesos itself when acks get delayed. + return + } + val index = info.index + info.markFailed() + decreaseRunningTasks(1) + if (!finished(index)) { + logInfo("Lost TID %s (task %s:%d)".format(tid, taskSet.id, index)) + copiesRunning(index) -= 1 + // Check if the problem is a map output fetch failure. In that case, this + // task will never succeed on any node, so tell the scheduler about it. + if (serializedData != null && serializedData.limit() > 0) { + val reason = ser.deserialize[TaskEndReason](serializedData, getClass.getClassLoader) + reason match { + case fetchFailed: FetchFailed => + logInfo("Loss was due to fetch failure from " + fetchFailed.bmAddress) + sched.listener.taskEnded(tasks(index), fetchFailed, null, null, info, null) + finished(index) = true + tasksFinished += 1 + sched.taskSetFinished(this) + decreaseRunningTasks(runningTasks) + return + + case taskResultTooBig: TaskResultTooBigFailure => + logInfo("Loss was due to task %s result exceeding Akka frame size; " + + "aborting job".format(tid)) + abort("Task %s result exceeded Akka frame size".format(tid)) + return + + case ef: ExceptionFailure => + val key = ef.description + val now = System.currentTimeMillis + val (printFull, dupCount) = { + if (recentExceptions.contains(key)) { + val (dupCount, printTime) = recentExceptions(key) + if (now - printTime > EXCEPTION_PRINT_INTERVAL) { + recentExceptions(key) = (0, now) + (true, 0) + } else { + recentExceptions(key) = (dupCount + 1, printTime) + (false, dupCount + 1) + } + } else { + recentExceptions(key) = (0, now) + (true, 0) + } + } + if (printFull) { + val locs = ef.stackTrace.map(loc => "\tat %s".format(loc.toString)) + logInfo("Loss was due to %s\n%s\n%s".format( + ef.className, ef.description, locs.mkString("\n"))) + } else { + logInfo("Loss was due to %s [duplicate %d]".format(ef.description, dupCount)) + } + + case _ => {} + } + } + // On non-fetch failures, re-enqueue the task as pending for a max number of retries + addPendingTask(index) + // Count failed attempts only on FAILED and LOST state (not on KILLED) + if (state == TaskState.FAILED || state == TaskState.LOST) { + numFailures(index) += 1 + if (numFailures(index) > MAX_TASK_FAILURES) { + logError("Task %s:%d failed more than %d times; aborting job".format( + taskSet.id, index, MAX_TASK_FAILURES)) + abort("Task %s:%d failed more than %d times".format(taskSet.id, index, MAX_TASK_FAILURES)) + } + } + } else { + logInfo("Ignoring task-lost event for TID " + tid + + " because task " + index + " is already finished") + } + } + + def error(message: String) { + // Save the error message + abort("Error: " + message) + } + + def abort(message: String) { + failed = true + causeOfFailure = message + // TODO: Kill running tasks if we were not terminated due to a Mesos error + sched.listener.taskSetFailed(taskSet, message) + decreaseRunningTasks(runningTasks) + sched.taskSetFinished(this) + } + + override def increaseRunningTasks(taskNum: Int) { + runningTasks += taskNum + if (parent != null) { + parent.increaseRunningTasks(taskNum) + } + } + + override def decreaseRunningTasks(taskNum: Int) { + runningTasks -= taskNum + if (parent != null) { + parent.decreaseRunningTasks(taskNum) + } + } + + //TODO: for now we just find Pool not TaskSetManager, we can extend this function in future if needed + override def getSchedulableByName(name: String): Schedulable = { + return null + } + + override def addSchedulable(schedulable:Schedulable) { + //nothing + } + + override def removeSchedulable(schedulable:Schedulable) { + //nothing + } + + override def getSortedTaskSetQueue(): ArrayBuffer[TaskSetManager] = { + var sortedTaskSetQueue = new ArrayBuffer[TaskSetManager] + sortedTaskSetQueue += this + return sortedTaskSetQueue + } + + override def executorLost(execId: String, hostPort: String) { + logInfo("Re-queueing tasks for " + execId + " from TaskSet " + taskSet.id) + + // If some task has preferred locations only on hostname, and there are no more executors there, + // put it in the no-prefs list to avoid the wait from delay scheduling + + // host local tasks - should we push this to rack local or no pref list ? For now, preserving behavior and moving to + // no prefs list. Note, this was done due to impliations related to 'waiting' for data local tasks, etc. + // Note: NOT checking process local list - since host local list is super set of that. We need to ad to no prefs only if + // there is no host local node for the task (not if there is no process local node for the task) + for (index <- getPendingTasksForHost(Utils.parseHostPort(hostPort)._1)) { + // val newLocs = findPreferredLocations(tasks(index).preferredLocations, sched, TaskLocality.RACK_LOCAL) + val newLocs = findPreferredLocations(tasks(index).preferredLocations, sched, TaskLocality.NODE_LOCAL) + if (newLocs.isEmpty) { + pendingTasksWithNoPrefs += index + } + } + + // Re-enqueue any tasks that ran on the failed executor if this is a shuffle map stage + if (tasks(0).isInstanceOf[ShuffleMapTask]) { + for ((tid, info) <- taskInfos if info.executorId == execId) { + val index = taskInfos(tid).index + if (finished(index)) { + finished(index) = false + copiesRunning(index) -= 1 + tasksFinished -= 1 + addPendingTask(index) + // Tell the DAGScheduler that this task was resubmitted so that it doesn't think our + // stage finishes when a total of tasks.size tasks finish. + sched.listener.taskEnded(tasks(index), Resubmitted, null, null, info, null) + } + } + } + // Also re-enqueue any tasks that were running on the node + for ((tid, info) <- taskInfos if info.running && info.executorId == execId) { + taskLost(tid, TaskState.KILLED, null) + } + } + + /** + * Check for tasks to be speculated and return true if there are any. This is called periodically + * by the ClusterScheduler. + * + * TODO: To make this scale to large jobs, we need to maintain a list of running tasks, so that + * we don't scan the whole task set. It might also help to make this sorted by launch time. + */ + override def checkSpeculatableTasks(): Boolean = { + // Can't speculate if we only have one task, or if all tasks have finished. + if (numTasks == 1 || tasksFinished == numTasks) { + return false + } + var foundTasks = false + val minFinishedForSpeculation = (SPECULATION_QUANTILE * numTasks).floor.toInt + logDebug("Checking for speculative tasks: minFinished = " + minFinishedForSpeculation) + if (tasksFinished >= minFinishedForSpeculation) { + val time = System.currentTimeMillis() + val durations = taskInfos.values.filter(_.successful).map(_.duration).toArray + Arrays.sort(durations) + val medianDuration = durations(min((0.5 * numTasks).round.toInt, durations.size - 1)) + val threshold = max(SPECULATION_MULTIPLIER * medianDuration, 100) + // TODO: Threshold should also look at standard deviation of task durations and have a lower + // bound based on that. + logDebug("Task length threshold for speculation: " + threshold) + for ((tid, info) <- taskInfos) { + val index = info.index + if (!finished(index) && copiesRunning(index) == 1 && info.timeRunning(time) > threshold && + !speculatableTasks.contains(index)) { + logInfo( + "Marking task %s:%d (on %s) as speculatable because it ran more than %.0f ms".format( + taskSet.id, index, info.hostPort, threshold)) + speculatableTasks += index + foundTasks = true + } + } + } + return foundTasks + } + + override def hasPendingTasks(): Boolean = { + numTasks > 0 && tasksFinished < numTasks + } +} diff --git a/core/src/main/scala/spark/scheduler/cluster/Pool.scala b/core/src/main/scala/spark/scheduler/cluster/Pool.scala new file mode 100644 index 0000000000..941ba7a3f1 --- /dev/null +++ b/core/src/main/scala/spark/scheduler/cluster/Pool.scala @@ -0,0 +1,104 @@ +package spark.scheduler.cluster + +import scala.collection.mutable.ArrayBuffer +import scala.collection.mutable.HashMap + +import spark.Logging +import spark.scheduler.cluster.SchedulingMode.SchedulingMode + +/** + * An Schedulable entity that represent collection of Pools or TaskSetManagers + */ + +private[spark] class Pool( + val poolName: String, + val schedulingMode: SchedulingMode, + initMinShare: Int, + initWeight: Int) + extends Schedulable + with Logging { + + var schedulableQueue = new ArrayBuffer[Schedulable] + var schedulableNameToSchedulable = new HashMap[String, Schedulable] + + var weight = initWeight + var minShare = initMinShare + var runningTasks = 0 + + var priority = 0 + var stageId = 0 + var name = poolName + var parent:Schedulable = null + + var taskSetSchedulingAlgorithm: SchedulingAlgorithm = { + schedulingMode match { + case SchedulingMode.FAIR => + new FairSchedulingAlgorithm() + case SchedulingMode.FIFO => + new FIFOSchedulingAlgorithm() + } + } + + override def addSchedulable(schedulable: Schedulable) { + schedulableQueue += schedulable + schedulableNameToSchedulable(schedulable.name) = schedulable + schedulable.parent= this + } + + override def removeSchedulable(schedulable: Schedulable) { + schedulableQueue -= schedulable + schedulableNameToSchedulable -= schedulable.name + } + + override def getSchedulableByName(schedulableName: String): Schedulable = { + if (schedulableNameToSchedulable.contains(schedulableName)) { + return schedulableNameToSchedulable(schedulableName) + } + for (schedulable <- schedulableQueue) { + var sched = schedulable.getSchedulableByName(schedulableName) + if (sched != null) { + return sched + } + } + return null + } + + override def executorLost(executorId: String, host: String) { + schedulableQueue.foreach(_.executorLost(executorId, host)) + } + + override def checkSpeculatableTasks(): Boolean = { + var shouldRevive = false + for (schedulable <- schedulableQueue) { + shouldRevive |= schedulable.checkSpeculatableTasks() + } + return shouldRevive + } + + override def getSortedTaskSetQueue(): ArrayBuffer[TaskSetManager] = { + var sortedTaskSetQueue = new ArrayBuffer[TaskSetManager] + val sortedSchedulableQueue = schedulableQueue.sortWith(taskSetSchedulingAlgorithm.comparator) + for (schedulable <- sortedSchedulableQueue) { + sortedTaskSetQueue ++= schedulable.getSortedTaskSetQueue() + } + return sortedTaskSetQueue + } + + override def increaseRunningTasks(taskNum: Int) { + runningTasks += taskNum + if (parent != null) { + parent.increaseRunningTasks(taskNum) + } + } + + override def decreaseRunningTasks(taskNum: Int) { + runningTasks -= taskNum + if (parent != null) { + parent.decreaseRunningTasks(taskNum) + } + } + + override def hasPendingTasks(): Boolean = { + schedulableQueue.exists(_.hasPendingTasks()) + } +} diff --git a/core/src/main/scala/spark/scheduler/cluster/Schedulable.scala b/core/src/main/scala/spark/scheduler/cluster/Schedulable.scala new file mode 100644 index 0000000000..2dd9c0564f --- /dev/null +++ b/core/src/main/scala/spark/scheduler/cluster/Schedulable.scala @@ -0,0 +1,27 @@ +package spark.scheduler.cluster + +import scala.collection.mutable.ArrayBuffer + +/** + * An interface for schedulable entities. + * there are two type of Schedulable entities(Pools and TaskSetManagers) + */ +private[spark] trait Schedulable { + var parent: Schedulable + def weight: Int + def minShare: Int + def runningTasks: Int + def priority: Int + def stageId: Int + def name: String + + def increaseRunningTasks(taskNum: Int): Unit + def decreaseRunningTasks(taskNum: Int): Unit + def addSchedulable(schedulable: Schedulable): Unit + def removeSchedulable(schedulable: Schedulable): Unit + def getSchedulableByName(name: String): Schedulable + def executorLost(executorId: String, host: String): Unit + def checkSpeculatableTasks(): Boolean + def getSortedTaskSetQueue(): ArrayBuffer[TaskSetManager] + def hasPendingTasks(): Boolean +} diff --git a/core/src/main/scala/spark/scheduler/cluster/SchedulableBuilder.scala b/core/src/main/scala/spark/scheduler/cluster/SchedulableBuilder.scala new file mode 100644 index 0000000000..18cc15c2a5 --- /dev/null +++ b/core/src/main/scala/spark/scheduler/cluster/SchedulableBuilder.scala @@ -0,0 +1,115 @@ +package spark.scheduler.cluster + +import java.io.{File, FileInputStream, FileOutputStream} + +import scala.collection.mutable.ArrayBuffer +import scala.collection.mutable.ArrayBuffer +import scala.collection.mutable.HashMap +import scala.collection.mutable.HashSet +import scala.util.control.Breaks._ +import scala.xml._ + +import spark.Logging +import spark.scheduler.cluster.SchedulingMode.SchedulingMode + +import java.util.Properties + +/** + * An interface to build Schedulable tree + * buildPools: build the tree nodes(pools) + * addTaskSetManager: build the leaf nodes(TaskSetManagers) + */ +private[spark] trait SchedulableBuilder { + def buildPools() + def addTaskSetManager(manager: Schedulable, properties: Properties) +} + +private[spark] class FIFOSchedulableBuilder(val rootPool: Pool) extends SchedulableBuilder with Logging { + + override def buildPools() { + //nothing + } + + override def addTaskSetManager(manager: Schedulable, properties: Properties) { + rootPool.addSchedulable(manager) + } +} + +private[spark] class FairSchedulableBuilder(val rootPool: Pool) extends SchedulableBuilder with Logging { + + val schedulerAllocFile = System.getProperty("spark.fairscheduler.allocation.file","unspecified") + val FAIR_SCHEDULER_PROPERTIES = "spark.scheduler.cluster.fair.pool" + val DEFAULT_POOL_NAME = "default" + val MINIMUM_SHARES_PROPERTY = "minShare" + val SCHEDULING_MODE_PROPERTY = "schedulingMode" + val WEIGHT_PROPERTY = "weight" + val POOL_NAME_PROPERTY = "@name" + val POOLS_PROPERTY = "pool" + val DEFAULT_SCHEDULING_MODE = SchedulingMode.FIFO + val DEFAULT_MINIMUM_SHARE = 2 + val DEFAULT_WEIGHT = 1 + + override def buildPools() { + val file = new File(schedulerAllocFile) + if (file.exists()) { + val xml = XML.loadFile(file) + for (poolNode <- (xml \\ POOLS_PROPERTY)) { + + val poolName = (poolNode \ POOL_NAME_PROPERTY).text + var schedulingMode = DEFAULT_SCHEDULING_MODE + var minShare = DEFAULT_MINIMUM_SHARE + var weight = DEFAULT_WEIGHT + + val xmlSchedulingMode = (poolNode \ SCHEDULING_MODE_PROPERTY).text + if (xmlSchedulingMode != "") { + try { + schedulingMode = SchedulingMode.withName(xmlSchedulingMode) + } catch { + case e: Exception => logInfo("Error xml schedulingMode, using default schedulingMode") + } + } + + val xmlMinShare = (poolNode \ MINIMUM_SHARES_PROPERTY).text + if (xmlMinShare != "") { + minShare = xmlMinShare.toInt + } + + val xmlWeight = (poolNode \ WEIGHT_PROPERTY).text + if (xmlWeight != "") { + weight = xmlWeight.toInt + } + + val pool = new Pool(poolName, schedulingMode, minShare, weight) + rootPool.addSchedulable(pool) + logInfo("Create new pool with name:%s,schedulingMode:%s,minShare:%d,weight:%d".format( + poolName, schedulingMode, minShare, weight)) + } + } + + //finally create "default" pool + if (rootPool.getSchedulableByName(DEFAULT_POOL_NAME) == null) { + val pool = new Pool(DEFAULT_POOL_NAME, DEFAULT_SCHEDULING_MODE, DEFAULT_MINIMUM_SHARE, DEFAULT_WEIGHT) + rootPool.addSchedulable(pool) + logInfo("Create default pool with name:%s,schedulingMode:%s,minShare:%d,weight:%d".format( + DEFAULT_POOL_NAME, DEFAULT_SCHEDULING_MODE, DEFAULT_MINIMUM_SHARE, DEFAULT_WEIGHT)) + } + } + + override def addTaskSetManager(manager: Schedulable, properties: Properties) { + var poolName = DEFAULT_POOL_NAME + var parentPool = rootPool.getSchedulableByName(poolName) + if (properties != null) { + poolName = properties.getProperty(FAIR_SCHEDULER_PROPERTIES, DEFAULT_POOL_NAME) + parentPool = rootPool.getSchedulableByName(poolName) + if (parentPool == null) { + //we will create a new pool that user has configured in app instead of being defined in xml file + parentPool = new Pool(poolName,DEFAULT_SCHEDULING_MODE, DEFAULT_MINIMUM_SHARE, DEFAULT_WEIGHT) + rootPool.addSchedulable(parentPool) + logInfo("Create pool with name:%s,schedulingMode:%s,minShare:%d,weight:%d".format( + poolName, DEFAULT_SCHEDULING_MODE, DEFAULT_MINIMUM_SHARE, DEFAULT_WEIGHT)) + } + } + parentPool.addSchedulable(manager) + logInfo("Added task set " + manager.name + " tasks to pool "+poolName) + } +} diff --git a/core/src/main/scala/spark/scheduler/cluster/SchedulingAlgorithm.scala b/core/src/main/scala/spark/scheduler/cluster/SchedulingAlgorithm.scala new file mode 100644 index 0000000000..f33310a34a --- /dev/null +++ b/core/src/main/scala/spark/scheduler/cluster/SchedulingAlgorithm.scala @@ -0,0 +1,64 @@ +package spark.scheduler.cluster + +/** + * An interface for sort algorithm + * FIFO: FIFO algorithm between TaskSetManagers + * FS: FS algorithm between Pools, and FIFO or FS within Pools + */ +private[spark] trait SchedulingAlgorithm { + def comparator(s1: Schedulable, s2: Schedulable): Boolean +} + +private[spark] class FIFOSchedulingAlgorithm extends SchedulingAlgorithm { + override def comparator(s1: Schedulable, s2: Schedulable): Boolean = { + val priority1 = s1.priority + val priority2 = s2.priority + var res = math.signum(priority1 - priority2) + if (res == 0) { + val stageId1 = s1.stageId + val stageId2 = s2.stageId + res = math.signum(stageId1 - stageId2) + } + if (res < 0) { + return true + } else { + return false + } + } +} + +private[spark] class FairSchedulingAlgorithm extends SchedulingAlgorithm { + override def comparator(s1: Schedulable, s2: Schedulable): Boolean = { + val minShare1 = s1.minShare + val minShare2 = s2.minShare + val runningTasks1 = s1.runningTasks + val runningTasks2 = s2.runningTasks + val s1Needy = runningTasks1 < minShare1 + val s2Needy = runningTasks2 < minShare2 + val minShareRatio1 = runningTasks1.toDouble / math.max(minShare1, 1.0).toDouble + val minShareRatio2 = runningTasks2.toDouble / math.max(minShare2, 1.0).toDouble + val taskToWeightRatio1 = runningTasks1.toDouble / s1.weight.toDouble + val taskToWeightRatio2 = runningTasks2.toDouble / s2.weight.toDouble + var res:Boolean = true + var compare:Int = 0 + + if (s1Needy && !s2Needy) { + return true + } else if (!s1Needy && s2Needy) { + return false + } else if (s1Needy && s2Needy) { + compare = minShareRatio1.compareTo(minShareRatio2) + } else { + compare = taskToWeightRatio1.compareTo(taskToWeightRatio2) + } + + if (compare < 0) { + return true + } else if (compare > 0) { + return false + } else { + return s1.name < s2.name + } + } +} + diff --git a/core/src/main/scala/spark/scheduler/cluster/SchedulingMode.scala b/core/src/main/scala/spark/scheduler/cluster/SchedulingMode.scala new file mode 100644 index 0000000000..6e0c6793e0 --- /dev/null +++ b/core/src/main/scala/spark/scheduler/cluster/SchedulingMode.scala @@ -0,0 +1,7 @@ +package spark.scheduler.cluster + +object SchedulingMode extends Enumeration("FAIR","FIFO"){ + + type SchedulingMode = Value + val FAIR,FIFO = Value +} diff --git a/core/src/main/scala/spark/scheduler/cluster/SparkDeploySchedulerBackend.scala b/core/src/main/scala/spark/scheduler/cluster/SparkDeploySchedulerBackend.scala index bb289c9cf3..170ede0f44 100644 --- a/core/src/main/scala/spark/scheduler/cluster/SparkDeploySchedulerBackend.scala +++ b/core/src/main/scala/spark/scheduler/cluster/SparkDeploySchedulerBackend.scala @@ -31,7 +31,8 @@ private[spark] class SparkDeploySchedulerBackend( val command = Command("spark.executor.StandaloneExecutorBackend", args, sc.executorEnvs) val sparkHome = sc.getSparkHome().getOrElse( throw new IllegalArgumentException("must supply spark home for spark standalone")) - val appDesc = new ApplicationDescription(appName, maxCores, executorMemory, command, sparkHome) + val appDesc = new ApplicationDescription(appName, maxCores, executorMemory, command, sparkHome, + sc.ui.appUIAddress) client = new Client(sc.env.actorSystem, master, appDesc, this) client.start() @@ -57,9 +58,9 @@ private[spark] class SparkDeploySchedulerBackend( } } - override def executorAdded(executorId: String, workerId: String, host: String, cores: Int, memory: Int) { - logInfo("Granted executor ID %s on host %s with %d cores, %s RAM".format( - executorId, host, cores, Utils.memoryMegabytesToString(memory))) + override def executorAdded(executorId: String, workerId: String, hostPort: String, cores: Int, memory: Int) { + logInfo("Granted executor ID %s on hostPort %s with %d cores, %s RAM".format( + executorId, hostPort, cores, Utils.memoryMegabytesToString(memory))) } override def executorRemoved(executorId: String, message: String, exitStatus: Option[Int]) { diff --git a/core/src/main/scala/spark/scheduler/cluster/StandaloneClusterMessage.scala b/core/src/main/scala/spark/scheduler/cluster/StandaloneClusterMessage.scala index d766067824..3335294844 100644 --- a/core/src/main/scala/spark/scheduler/cluster/StandaloneClusterMessage.scala +++ b/core/src/main/scala/spark/scheduler/cluster/StandaloneClusterMessage.scala @@ -3,6 +3,7 @@ package spark.scheduler.cluster import spark.TaskState.TaskState import java.nio.ByteBuffer import spark.util.SerializableBuffer +import spark.Utils private[spark] sealed trait StandaloneClusterMessage extends Serializable @@ -19,8 +20,10 @@ case class RegisterExecutorFailed(message: String) extends StandaloneClusterMess // Executors to driver private[spark] -case class RegisterExecutor(executorId: String, host: String, cores: Int) - extends StandaloneClusterMessage +case class RegisterExecutor(executorId: String, hostPort: String, cores: Int) + extends StandaloneClusterMessage { + Utils.checkHostPort(hostPort, "Expected host port") +} private[spark] case class StatusUpdate(executorId: String, taskId: Long, state: TaskState, data: SerializableBuffer) diff --git a/core/src/main/scala/spark/scheduler/cluster/StandaloneSchedulerBackend.scala b/core/src/main/scala/spark/scheduler/cluster/StandaloneSchedulerBackend.scala index d606432572..004592a540 100644 --- a/core/src/main/scala/spark/scheduler/cluster/StandaloneSchedulerBackend.scala +++ b/core/src/main/scala/spark/scheduler/cluster/StandaloneSchedulerBackend.scala @@ -5,8 +5,9 @@ import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet} import akka.actor._ import akka.util.duration._ import akka.pattern.ask +import akka.util.Duration -import spark.{SparkException, Logging, TaskState} +import spark.{Utils, SparkException, Logging, TaskState} import akka.dispatch.Await import java.util.concurrent.atomic.AtomicInteger import akka.remote.{RemoteClientShutdown, RemoteClientDisconnected, RemoteClientLifeCycleEvent} @@ -24,12 +25,12 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor var totalCoreCount = new AtomicInteger(0) class DriverActor(sparkProperties: Seq[(String, String)]) extends Actor { - val executorActor = new HashMap[String, ActorRef] - val executorAddress = new HashMap[String, Address] - val executorHost = new HashMap[String, String] - val freeCores = new HashMap[String, Int] - val actorToExecutorId = new HashMap[ActorRef, String] - val addressToExecutorId = new HashMap[Address, String] + private val executorActor = new HashMap[String, ActorRef] + private val executorAddress = new HashMap[String, Address] + private val executorHostPort = new HashMap[String, String] + private val freeCores = new HashMap[String, Int] + private val actorToExecutorId = new HashMap[ActorRef, String] + private val addressToExecutorId = new HashMap[Address, String] override def preStart() { // Listen for remote client disconnection events, since they don't go through Akka's watch() @@ -37,7 +38,8 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor } def receive = { - case RegisterExecutor(executorId, host, cores) => + case RegisterExecutor(executorId, hostPort, cores) => + Utils.checkHostPort(hostPort, "Host port expected " + hostPort) if (executorActor.contains(executorId)) { sender ! RegisterExecutorFailed("Duplicate executor ID: " + executorId) } else { @@ -45,7 +47,7 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor sender ! RegisteredExecutor(sparkProperties) context.watch(sender) executorActor(executorId) = sender - executorHost(executorId) = host + executorHostPort(executorId) = hostPort freeCores(executorId) = cores executorAddress(executorId) = sender.path.address actorToExecutorId(sender) = executorId @@ -85,13 +87,13 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor // Make fake resource offers on all executors def makeOffers() { launchTasks(scheduler.resourceOffers( - executorHost.toArray.map {case (id, host) => new WorkerOffer(id, host, freeCores(id))})) + executorHostPort.toArray.map {case (id, hostPort) => new WorkerOffer(id, hostPort, freeCores(id))})) } // Make fake resource offers on just one executor def makeOffers(executorId: String) { launchTasks(scheduler.resourceOffers( - Seq(new WorkerOffer(executorId, executorHost(executorId), freeCores(executorId))))) + Seq(new WorkerOffer(executorId, executorHostPort(executorId), freeCores(executorId))))) } // Launch tasks returned by a set of resource offers @@ -110,9 +112,9 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor actorToExecutorId -= executorActor(executorId) addressToExecutorId -= executorAddress(executorId) executorActor -= executorId - executorHost -= executorId + executorHostPort -= executorId freeCores -= executorId - executorHost -= executorId + executorHostPort -= executorId totalCoreCount.addAndGet(-numCores) scheduler.executorLost(executorId, SlaveLost(reason)) } @@ -128,7 +130,7 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor while (iterator.hasNext) { val entry = iterator.next val (key, value) = (entry.getKey.toString, entry.getValue.toString) - if (key.startsWith("spark.")) { + if (key.startsWith("spark.") && !key.equals("spark.hostPort")) { properties += ((key, value)) } } @@ -136,10 +138,11 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor Props(new DriverActor(properties)), name = StandaloneSchedulerBackend.ACTOR_NAME) } + private val timeout = Duration.create(System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds") + override def stop() { try { if (driverActor != null) { - val timeout = 5.seconds val future = driverActor.ask(StopDriver)(timeout) Await.result(future, timeout) } @@ -153,12 +156,12 @@ class StandaloneSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Actor driverActor ! ReviveOffers } - override def defaultParallelism(): Int = math.max(totalCoreCount.get(), 2) + override def defaultParallelism() = Option(System.getProperty("spark.default.parallelism")) + .map(_.toInt).getOrElse(math.max(totalCoreCount.get(), 2)) // Called by subclasses when notified of a lost worker def removeExecutor(executorId: String, reason: String) { try { - val timeout = 5.seconds val future = driverActor.ask(RemoveExecutor(executorId, reason))(timeout) Await.result(future, timeout) } catch { diff --git a/core/src/main/scala/spark/scheduler/cluster/TaskInfo.scala b/core/src/main/scala/spark/scheduler/cluster/TaskInfo.scala index 0f975ce1eb..718f26bfbd 100644 --- a/core/src/main/scala/spark/scheduler/cluster/TaskInfo.scala +++ b/core/src/main/scala/spark/scheduler/cluster/TaskInfo.scala @@ -1,5 +1,7 @@ package spark.scheduler.cluster +import spark.Utils + /** * Information about a running task attempt inside a TaskSet. */ @@ -9,7 +11,11 @@ class TaskInfo( val index: Int, val launchTime: Long, val executorId: String, - val host: String) { + val hostPort: String, + val taskLocality: TaskLocality.TaskLocality) { + + Utils.checkHostPort(hostPort, "Expected hostport") + var finishTime: Long = 0 var failed = false diff --git a/core/src/main/scala/spark/scheduler/cluster/TaskSetManager.scala b/core/src/main/scala/spark/scheduler/cluster/TaskSetManager.scala index 3dabdd76b1..b4dd75d90f 100644 --- a/core/src/main/scala/spark/scheduler/cluster/TaskSetManager.scala +++ b/core/src/main/scala/spark/scheduler/cluster/TaskSetManager.scala @@ -1,429 +1,17 @@ package spark.scheduler.cluster -import java.util.Arrays -import java.util.{HashMap => JHashMap} - import scala.collection.mutable.ArrayBuffer -import scala.collection.mutable.HashMap -import scala.collection.mutable.HashSet -import scala.math.max -import scala.math.min - -import spark._ import spark.scheduler._ import spark.TaskState.TaskState import java.nio.ByteBuffer -/** - * Schedules the tasks within a single TaskSet in the ClusterScheduler. - */ -private[spark] class TaskSetManager(sched: ClusterScheduler, val taskSet: TaskSet) extends Logging { - - // Maximum time to wait to run a task in a preferred location (in ms) - val LOCALITY_WAIT = System.getProperty("spark.locality.wait", "3000").toLong - - // CPUs to request per task - val CPUS_PER_TASK = System.getProperty("spark.task.cpus", "1").toDouble - - // Maximum times a task is allowed to fail before failing the job - val MAX_TASK_FAILURES = 4 - - // Quantile of tasks at which to start speculation - val SPECULATION_QUANTILE = System.getProperty("spark.speculation.quantile", "0.75").toDouble - val SPECULATION_MULTIPLIER = System.getProperty("spark.speculation.multiplier", "1.5").toDouble - - // Serializer for closures and tasks. - val ser = SparkEnv.get.closureSerializer.newInstance() - - val priority = taskSet.priority - val tasks = taskSet.tasks - val numTasks = tasks.length - val copiesRunning = new Array[Int](numTasks) - val finished = new Array[Boolean](numTasks) - val numFailures = new Array[Int](numTasks) - val taskAttempts = Array.fill[List[TaskInfo]](numTasks)(Nil) - var tasksFinished = 0 - - // Last time when we launched a preferred task (for delay scheduling) - var lastPreferredLaunchTime = System.currentTimeMillis - - // List of pending tasks for each node. These collections are actually - // treated as stacks, in which new tasks are added to the end of the - // ArrayBuffer and removed from the end. This makes it faster to detect - // tasks that repeatedly fail because whenever a task failed, it is put - // back at the head of the stack. They are also only cleaned up lazily; - // when a task is launched, it remains in all the pending lists except - // the one that it was launched from, but gets removed from them later. - val pendingTasksForHost = new HashMap[String, ArrayBuffer[Int]] - - // List containing pending tasks with no locality preferences - val pendingTasksWithNoPrefs = new ArrayBuffer[Int] - - // List containing all pending tasks (also used as a stack, as above) - val allPendingTasks = new ArrayBuffer[Int] - - // Tasks that can be speculated. Since these will be a small fraction of total - // tasks, we'll just hold them in a HashSet. - val speculatableTasks = new HashSet[Int] - - // Task index, start and finish time for each task attempt (indexed by task ID) - val taskInfos = new HashMap[Long, TaskInfo] - - // Did the job fail? - var failed = false - var causeOfFailure = "" - - // How frequently to reprint duplicate exceptions in full, in milliseconds - val EXCEPTION_PRINT_INTERVAL = - System.getProperty("spark.logging.exceptionPrintInterval", "10000").toLong - // Map of recent exceptions (identified by string representation and - // top stack frame) to duplicate count (how many times the same - // exception has appeared) and time the full exception was - // printed. This should ideally be an LRU map that can drop old - // exceptions automatically. - val recentExceptions = HashMap[String, (Int, Long)]() - - // Figure out the current map output tracker generation and set it on all tasks - val generation = sched.mapOutputTracker.getGeneration - logDebug("Generation for " + taskSet.id + ": " + generation) - for (t <- tasks) { - t.generation = generation - } - - // Add all our tasks to the pending lists. We do this in reverse order - // of task index so that tasks with low indices get launched first. - for (i <- (0 until numTasks).reverse) { - addPendingTask(i) - } - - // Add a task to all the pending-task lists that it should be on. - private def addPendingTask(index: Int) { - val locations = tasks(index).preferredLocations.toSet & sched.hostsAlive - if (locations.size == 0) { - pendingTasksWithNoPrefs += index - } else { - for (host <- locations) { - val list = pendingTasksForHost.getOrElseUpdate(host, ArrayBuffer()) - list += index - } - } - allPendingTasks += index - } - - // Return the pending tasks list for a given host, or an empty list if - // there is no map entry for that host - private def getPendingTasksForHost(host: String): ArrayBuffer[Int] = { - pendingTasksForHost.getOrElse(host, ArrayBuffer()) - } - - // Dequeue a pending task from the given list and return its index. - // Return None if the list is empty. - // This method also cleans up any tasks in the list that have already - // been launched, since we want that to happen lazily. - private def findTaskFromList(list: ArrayBuffer[Int]): Option[Int] = { - while (!list.isEmpty) { - val index = list.last - list.trimEnd(1) - if (copiesRunning(index) == 0 && !finished(index)) { - return Some(index) - } - } - return None - } - - // Return a speculative task for a given host if any are available. The task should not have an - // attempt running on this host, in case the host is slow. In addition, if localOnly is set, the - // task must have a preference for this host (or no preferred locations at all). - private def findSpeculativeTask(host: String, localOnly: Boolean): Option[Int] = { - val hostsAlive = sched.hostsAlive - speculatableTasks.retain(index => !finished(index)) // Remove finished tasks from set - val localTask = speculatableTasks.find { - index => - val locations = tasks(index).preferredLocations.toSet & hostsAlive - val attemptLocs = taskAttempts(index).map(_.host) - (locations.size == 0 || locations.contains(host)) && !attemptLocs.contains(host) - } - if (localTask != None) { - speculatableTasks -= localTask.get - return localTask - } - if (!localOnly && speculatableTasks.size > 0) { - val nonLocalTask = speculatableTasks.find(i => !taskAttempts(i).map(_.host).contains(host)) - if (nonLocalTask != None) { - speculatableTasks -= nonLocalTask.get - return nonLocalTask - } - } - return None - } - - // Dequeue a pending task for a given node and return its index. - // If localOnly is set to false, allow non-local tasks as well. - private def findTask(host: String, localOnly: Boolean): Option[Int] = { - val localTask = findTaskFromList(getPendingTasksForHost(host)) - if (localTask != None) { - return localTask - } - val noPrefTask = findTaskFromList(pendingTasksWithNoPrefs) - if (noPrefTask != None) { - return noPrefTask - } - if (!localOnly) { - val nonLocalTask = findTaskFromList(allPendingTasks) - if (nonLocalTask != None) { - return nonLocalTask - } - } - // Finally, if all else has failed, find a speculative task - return findSpeculativeTask(host, localOnly) - } - - // Does a host count as a preferred location for a task? This is true if - // either the task has preferred locations and this host is one, or it has - // no preferred locations (in which we still count the launch as preferred). - private def isPreferredLocation(task: Task[_], host: String): Boolean = { - val locs = task.preferredLocations - return (locs.contains(host) || locs.isEmpty) - } - - // Respond to an offer of a single slave from the scheduler by finding a task - def slaveOffer(execId: String, host: String, availableCpus: Double): Option[TaskDescription] = { - if (tasksFinished < numTasks && availableCpus >= CPUS_PER_TASK) { - val time = System.currentTimeMillis - val localOnly = (time - lastPreferredLaunchTime < LOCALITY_WAIT) - - findTask(host, localOnly) match { - case Some(index) => { - // Found a task; do some bookkeeping and return a Mesos task for it - val task = tasks(index) - val taskId = sched.newTaskId() - // Figure out whether this should count as a preferred launch - val preferred = isPreferredLocation(task, host) - val prefStr = if (preferred) { - "preferred" - } else { - "non-preferred, not one of " + task.preferredLocations.mkString(", ") - } - logInfo("Starting task %s:%d as TID %s on executor %s: %s (%s)".format( - taskSet.id, index, taskId, execId, host, prefStr)) - // Do various bookkeeping - copiesRunning(index) += 1 - val info = new TaskInfo(taskId, index, time, execId, host) - taskInfos(taskId) = info - taskAttempts(index) = info :: taskAttempts(index) - if (preferred) { - lastPreferredLaunchTime = time - } - // Serialize and return the task - val startTime = System.currentTimeMillis - val serializedTask = Task.serializeWithDependencies( - task, sched.sc.addedFiles, sched.sc.addedJars, ser) - val timeTaken = System.currentTimeMillis - startTime - logInfo("Serialized task %s:%d as %d bytes in %d ms".format( - taskSet.id, index, serializedTask.limit, timeTaken)) - val taskName = "task %s:%d".format(taskSet.id, index) - return Some(new TaskDescription(taskId, execId, taskName, serializedTask)) - } - case _ => - } - } - return None - } - - def statusUpdate(tid: Long, state: TaskState, serializedData: ByteBuffer) { - state match { - case TaskState.FINISHED => - taskFinished(tid, state, serializedData) - case TaskState.LOST => - taskLost(tid, state, serializedData) - case TaskState.FAILED => - taskLost(tid, state, serializedData) - case TaskState.KILLED => - taskLost(tid, state, serializedData) - case _ => - } - } - - def taskFinished(tid: Long, state: TaskState, serializedData: ByteBuffer) { - val info = taskInfos(tid) - if (info.failed) { - // We might get two task-lost messages for the same task in coarse-grained Mesos mode, - // or even from Mesos itself when acks get delayed. - return - } - val index = info.index - info.markSuccessful() - if (!finished(index)) { - tasksFinished += 1 - logInfo("Finished TID %s in %d ms (progress: %d/%d)".format( - tid, info.duration, tasksFinished, numTasks)) - // Deserialize task result and pass it to the scheduler - val result = ser.deserialize[TaskResult[_]](serializedData, getClass.getClassLoader) - sched.listener.taskEnded(tasks(index), Success, result.value, result.accumUpdates) - // Mark finished and stop if we've finished all the tasks - finished(index) = true - if (tasksFinished == numTasks) { - sched.taskSetFinished(this) - } - } else { - logInfo("Ignoring task-finished event for TID " + tid + - " because task " + index + " is already finished") - } - } - - def taskLost(tid: Long, state: TaskState, serializedData: ByteBuffer) { - val info = taskInfos(tid) - if (info.failed) { - // We might get two task-lost messages for the same task in coarse-grained Mesos mode, - // or even from Mesos itself when acks get delayed. - return - } - val index = info.index - info.markFailed() - if (!finished(index)) { - logInfo("Lost TID %s (task %s:%d)".format(tid, taskSet.id, index)) - copiesRunning(index) -= 1 - // Check if the problem is a map output fetch failure. In that case, this - // task will never succeed on any node, so tell the scheduler about it. - if (serializedData != null && serializedData.limit() > 0) { - val reason = ser.deserialize[TaskEndReason](serializedData, getClass.getClassLoader) - reason match { - case fetchFailed: FetchFailed => - logInfo("Loss was due to fetch failure from " + fetchFailed.bmAddress) - sched.listener.taskEnded(tasks(index), fetchFailed, null, null) - finished(index) = true - tasksFinished += 1 - sched.taskSetFinished(this) - return - - case ef: ExceptionFailure => - val key = ef.exception.toString - val now = System.currentTimeMillis - val (printFull, dupCount) = { - if (recentExceptions.contains(key)) { - val (dupCount, printTime) = recentExceptions(key) - if (now - printTime > EXCEPTION_PRINT_INTERVAL) { - recentExceptions(key) = (0, now) - (true, 0) - } else { - recentExceptions(key) = (dupCount + 1, printTime) - (false, dupCount + 1) - } - } else { - recentExceptions(key) = (0, now) - (true, 0) - } - } - if (printFull) { - val locs = ef.exception.getStackTrace.map(loc => "\tat %s".format(loc.toString)) - logInfo("Loss was due to %s\n%s".format(ef.exception.toString, locs.mkString("\n"))) - } else { - logInfo("Loss was due to %s [duplicate %d]".format(ef.exception.toString, dupCount)) - } - - case _ => {} - } - } - // On non-fetch failures, re-enqueue the task as pending for a max number of retries - addPendingTask(index) - // Count failed attempts only on FAILED and LOST state (not on KILLED) - if (state == TaskState.FAILED || state == TaskState.LOST) { - numFailures(index) += 1 - if (numFailures(index) > MAX_TASK_FAILURES) { - logError("Task %s:%d failed more than %d times; aborting job".format( - taskSet.id, index, MAX_TASK_FAILURES)) - abort("Task %s:%d failed more than %d times".format(taskSet.id, index, MAX_TASK_FAILURES)) - } - } - } else { - logInfo("Ignoring task-lost event for TID " + tid + - " because task " + index + " is already finished") - } - } - - def error(message: String) { - // Save the error message - abort("Error: " + message) - } - - def abort(message: String) { - failed = true - causeOfFailure = message - // TODO: Kill running tasks if we were not terminated due to a Mesos error - sched.listener.taskSetFailed(taskSet, message) - sched.taskSetFinished(this) - } - - def executorLost(execId: String, hostname: String) { - logInfo("Re-queueing tasks for " + execId + " from TaskSet " + taskSet.id) - val newHostsAlive = sched.hostsAlive - // If some task has preferred locations only on hostname, and there are no more executors there, - // put it in the no-prefs list to avoid the wait from delay scheduling - if (!newHostsAlive.contains(hostname)) { - for (index <- getPendingTasksForHost(hostname)) { - val newLocs = tasks(index).preferredLocations.toSet & newHostsAlive - if (newLocs.isEmpty) { - pendingTasksWithNoPrefs += index - } - } - } - // Re-enqueue any tasks that ran on the failed executor if this is a shuffle map stage - if (tasks(0).isInstanceOf[ShuffleMapTask]) { - for ((tid, info) <- taskInfos if info.executorId == execId) { - val index = taskInfos(tid).index - if (finished(index)) { - finished(index) = false - copiesRunning(index) -= 1 - tasksFinished -= 1 - addPendingTask(index) - // Tell the DAGScheduler that this task was resubmitted so that it doesn't think our - // stage finishes when a total of tasks.size tasks finish. - sched.listener.taskEnded(tasks(index), Resubmitted, null, null) - } - } - } - // Also re-enqueue any tasks that were running on the node - for ((tid, info) <- taskInfos if info.running && info.executorId == execId) { - taskLost(tid, TaskState.KILLED, null) - } - } - - /** - * Check for tasks to be speculated and return true if there are any. This is called periodically - * by the ClusterScheduler. - * - * TODO: To make this scale to large jobs, we need to maintain a list of running tasks, so that - * we don't scan the whole task set. It might also help to make this sorted by launch time. - */ - def checkSpeculatableTasks(): Boolean = { - // Can't speculate if we only have one task, or if all tasks have finished. - if (numTasks == 1 || tasksFinished == numTasks) { - return false - } - var foundTasks = false - val minFinishedForSpeculation = (SPECULATION_QUANTILE * numTasks).floor.toInt - logDebug("Checking for speculative tasks: minFinished = " + minFinishedForSpeculation) - if (tasksFinished >= minFinishedForSpeculation) { - val time = System.currentTimeMillis() - val durations = taskInfos.values.filter(_.successful).map(_.duration).toArray - Arrays.sort(durations) - val medianDuration = durations(min((0.5 * numTasks).round.toInt, durations.size - 1)) - val threshold = max(SPECULATION_MULTIPLIER * medianDuration, 100) - // TODO: Threshold should also look at standard deviation of task durations and have a lower - // bound based on that. - logDebug("Task length threshold for speculation: " + threshold) - for ((tid, info) <- taskInfos) { - val index = info.index - if (!finished(index) && copiesRunning(index) == 1 && info.timeRunning(time) > threshold && - !speculatableTasks.contains(index)) { - logInfo( - "Marking task %s:%d (on %s) as speculatable because it ran more than %.0f ms".format( - taskSet.id, index, info.host, threshold)) - speculatableTasks += index - foundTasks = true - } - } - } - return foundTasks - } +private[spark] trait TaskSetManager extends Schedulable { + def taskSet: TaskSet + def slaveOffer(execId: String, hostPort: String, availableCpus: Double, + overrideLocality: TaskLocality.TaskLocality = null): Option[TaskDescription] + def numPendingTasksForHostPort(hostPort: String): Int + def numRackLocalPendingTasksForHost(hostPort :String): Int + def numPendingTasksForHost(hostPort: String): Int + def statusUpdate(tid: Long, state: TaskState, serializedData: ByteBuffer) + def error(message: String) } diff --git a/core/src/main/scala/spark/scheduler/cluster/WorkerOffer.scala b/core/src/main/scala/spark/scheduler/cluster/WorkerOffer.scala index 3c3afcbb14..c47824315c 100644 --- a/core/src/main/scala/spark/scheduler/cluster/WorkerOffer.scala +++ b/core/src/main/scala/spark/scheduler/cluster/WorkerOffer.scala @@ -4,5 +4,5 @@ package spark.scheduler.cluster * Represents free resources available on an executor. */ private[spark] -class WorkerOffer(val executorId: String, val hostname: String, val cores: Int) { +class WorkerOffer(val executorId: String, val hostPort: String, val cores: Int) { } diff --git a/core/src/main/scala/spark/scheduler/local/LocalScheduler.scala b/core/src/main/scala/spark/scheduler/local/LocalScheduler.scala index 482d1cc853..93d4318b29 100644 --- a/core/src/main/scala/spark/scheduler/local/LocalScheduler.scala +++ b/core/src/main/scala/spark/scheduler/local/LocalScheduler.scala @@ -1,21 +1,51 @@ package spark.scheduler.local import java.io.File -import java.net.URLClassLoader -import java.util.concurrent.Executors import java.util.concurrent.atomic.AtomicInteger +import java.nio.ByteBuffer +import scala.collection.mutable.ArrayBuffer import scala.collection.mutable.HashMap +import scala.collection.mutable.HashSet import spark._ -import executor.ExecutorURLClassLoader +import spark.TaskState.TaskState +import spark.executor.ExecutorURLClassLoader import spark.scheduler._ +import spark.scheduler.cluster._ +import akka.actor._ /** - * A simple TaskScheduler implementation that runs tasks locally in a thread pool. Optionally + * A FIFO or Fair TaskScheduler implementation that runs tasks locally in a thread pool. Optionally * the scheduler also allows each task to fail up to maxFailures times, which is useful for * testing fault recovery. */ -private[spark] class LocalScheduler(threads: Int, maxFailures: Int, sc: SparkContext) + +private[spark] case class LocalReviveOffers() +private[spark] case class LocalStatusUpdate(taskId: Long, state: TaskState, serializedData: ByteBuffer) + +private[spark] class LocalActor(localScheduler: LocalScheduler, var freeCores: Int) extends Actor with Logging { + def receive = { + case LocalReviveOffers => + launchTask(localScheduler.resourceOffer(freeCores)) + case LocalStatusUpdate(taskId, state, serializeData) => + freeCores += 1 + localScheduler.statusUpdate(taskId, state, serializeData) + launchTask(localScheduler.resourceOffer(freeCores)) + } + + def launchTask(tasks : Seq[TaskDescription]) { + for (task <- tasks) { + freeCores -= 1 + localScheduler.threadPool.submit(new Runnable { + def run() { + localScheduler.runTask(task.taskId,task.serializedTask) + } + }) + } + } +} + +private[spark] class LocalScheduler(threads: Int, val maxFailures: Int, val sc: SparkContext) extends TaskScheduler with Logging { @@ -31,79 +61,127 @@ private[spark] class LocalScheduler(threads: Int, maxFailures: Int, sc: SparkCon val classLoader = new ExecutorURLClassLoader(Array(), Thread.currentThread.getContextClassLoader) - // TODO: Need to take into account stage priority in scheduling + var schedulableBuilder: SchedulableBuilder = null + var rootPool: Pool = null + val activeTaskSets = new HashMap[String, TaskSetManager] + val taskIdToTaskSetId = new HashMap[Long, String] + val taskSetTaskIds = new HashMap[String, HashSet[Long]] + + var localActor: ActorRef = null + + override def start() { + //default scheduler is FIFO + val schedulingMode = System.getProperty("spark.cluster.schedulingmode", "FIFO") + //temporarily set rootPool name to empty + rootPool = new Pool("", SchedulingMode.withName(schedulingMode), 0, 0) + schedulableBuilder = { + schedulingMode match { + case "FIFO" => + new FIFOSchedulableBuilder(rootPool) + case "FAIR" => + new FairSchedulableBuilder(rootPool) + } + } + schedulableBuilder.buildPools() - override def start() { } + localActor = env.actorSystem.actorOf(Props(new LocalActor(this, threads)), "Test") + } override def setListener(listener: TaskSchedulerListener) { this.listener = listener } override def submitTasks(taskSet: TaskSet) { - val tasks = taskSet.tasks - val failCount = new Array[Int](tasks.size) - - def submitTask(task: Task[_], idInJob: Int) { - val myAttemptId = attemptId.getAndIncrement() - threadPool.submit(new Runnable { - def run() { - runTask(task, idInJob, myAttemptId) - } - }) + synchronized { + var manager = new LocalTaskSetManager(this, taskSet) + schedulableBuilder.addTaskSetManager(manager, manager.taskSet.properties) + activeTaskSets(taskSet.id) = manager + taskSetTaskIds(taskSet.id) = new HashSet[Long]() + localActor ! LocalReviveOffers } + } + + def resourceOffer(freeCores: Int): Seq[TaskDescription] = { + synchronized { + var freeCpuCores = freeCores + val tasks = new ArrayBuffer[TaskDescription](freeCores) + val sortedTaskSetQueue = rootPool.getSortedTaskSetQueue() + for (manager <- sortedTaskSetQueue) { + logDebug("parentName:%s,name:%s,runningTasks:%s".format(manager.parent.name, manager.name, manager.runningTasks)) + } - def runTask(task: Task[_], idInJob: Int, attemptId: Int) { - logInfo("Running " + task) - // Set the Spark execution environment for the worker thread - SparkEnv.set(env) - try { - Accumulators.clear() - Thread.currentThread().setContextClassLoader(classLoader) - - // Serialize and deserialize the task so that accumulators are changed to thread-local ones; - // this adds a bit of unnecessary overhead but matches how the Mesos Executor works. - val ser = SparkEnv.get.closureSerializer.newInstance() - val bytes = Task.serializeWithDependencies(task, sc.addedFiles, sc.addedJars, ser) - logInfo("Size of task " + idInJob + " is " + bytes.limit + " bytes") - val (taskFiles, taskJars, taskBytes) = Task.deserializeWithDependencies(bytes) - updateDependencies(taskFiles, taskJars) // Download any files added with addFile - val deserializedTask = ser.deserialize[Task[_]]( - taskBytes, Thread.currentThread.getContextClassLoader) - - // Run it - val result: Any = deserializedTask.run(attemptId) - - // Serialize and deserialize the result to emulate what the Mesos - // executor does. This is useful to catch serialization errors early - // on in development (so when users move their local Spark programs - // to the cluster, they don't get surprised by serialization errors). - val resultToReturn = ser.deserialize[Any](ser.serialize(result)) - val accumUpdates = ser.deserialize[collection.mutable.Map[Long, Any]]( - ser.serialize(Accumulators.values)) - logInfo("Finished " + task) - - // If the threadpool has not already been shutdown, notify DAGScheduler - if (!Thread.currentThread().isInterrupted) - listener.taskEnded(task, Success, resultToReturn, accumUpdates) - } catch { - case t: Throwable => { - logError("Exception in task " + idInJob, t) - failCount.synchronized { - failCount(idInJob) += 1 - if (failCount(idInJob) <= maxFailures) { - submitTask(task, idInJob) - } else { - // TODO: Do something nicer here to return all the way to the user - if (!Thread.currentThread().isInterrupted) - listener.taskEnded(task, new ExceptionFailure(t), null, null) + var launchTask = false + for (manager <- sortedTaskSetQueue) { + do { + launchTask = false + manager.slaveOffer(null,null,freeCpuCores) match { + case Some(task) => + tasks += task + taskIdToTaskSetId(task.taskId) = manager.taskSet.id + taskSetTaskIds(manager.taskSet.id) += task.taskId + freeCpuCores -= 1 + launchTask = true + case None => {} } - } - } + } while(launchTask) } + return tasks } + } - for ((task, i) <- tasks.zipWithIndex) { - submitTask(task, i) + def taskSetFinished(manager: TaskSetManager) { + synchronized { + activeTaskSets -= manager.taskSet.id + manager.parent.removeSchedulable(manager) + logInfo("Remove TaskSet %s from pool %s".format(manager.taskSet.id, manager.parent.name)) + taskIdToTaskSetId --= taskSetTaskIds(manager.taskSet.id) + taskSetTaskIds -= manager.taskSet.id + } + } + + def runTask(taskId: Long, bytes: ByteBuffer) { + logInfo("Running " + taskId) + val info = new TaskInfo(taskId, 0, System.currentTimeMillis(), "local", "local:1", TaskLocality.NODE_LOCAL) + // Set the Spark execution environment for the worker thread + SparkEnv.set(env) + val ser = SparkEnv.get.closureSerializer.newInstance() + try { + Accumulators.clear() + Thread.currentThread().setContextClassLoader(classLoader) + + // Serialize and deserialize the task so that accumulators are changed to thread-local ones; + // this adds a bit of unnecessary overhead but matches how the Mesos Executor works. + val (taskFiles, taskJars, taskBytes) = Task.deserializeWithDependencies(bytes) + updateDependencies(taskFiles, taskJars) // Download any files added with addFile + val deserStart = System.currentTimeMillis() + val deserializedTask = ser.deserialize[Task[_]]( + taskBytes, Thread.currentThread.getContextClassLoader) + val deserTime = System.currentTimeMillis() - deserStart + + // Run it + val result: Any = deserializedTask.run(taskId) + + // Serialize and deserialize the result to emulate what the Mesos + // executor does. This is useful to catch serialization errors early + // on in development (so when users move their local Spark programs + // to the cluster, they don't get surprised by serialization errors). + val serResult = ser.serialize(result) + deserializedTask.metrics.get.resultSize = serResult.limit() + val resultToReturn = ser.deserialize[Any](serResult) + val accumUpdates = ser.deserialize[collection.mutable.Map[Long, Any]]( + ser.serialize(Accumulators.values)) + logInfo("Finished " + taskId) + deserializedTask.metrics.get.executorRunTime = deserTime.toInt//info.duration.toInt //close enough + deserializedTask.metrics.get.executorDeserializeTime = deserTime.toInt + + val taskResult = new TaskResult(result, accumUpdates, deserializedTask.metrics.getOrElse(null)) + val serializedResult = ser.serialize(taskResult) + localActor ! LocalStatusUpdate(taskId, TaskState.FINISHED, serializedResult) + } catch { + case t: Throwable => { + val failure = new ExceptionFailure(t.getClass.getName, t.toString, t.getStackTrace) + localActor ! LocalStatusUpdate(taskId, TaskState.FAILED, ser.serialize(failure)) + } } } @@ -119,6 +197,7 @@ private[spark] class LocalScheduler(threads: Int, maxFailures: Int, sc: SparkCon Utils.fetchFile(name, new File(SparkFiles.getRootDirectory)) currentFiles(name) = timestamp } + for ((name, timestamp) <- newJars if currentJars.getOrElse(name, -1L) < timestamp) { logInfo("Fetching " + name + " with timestamp " + timestamp) Utils.fetchFile(name, new File(SparkFiles.getRootDirectory)) @@ -134,7 +213,16 @@ private[spark] class LocalScheduler(threads: Int, maxFailures: Int, sc: SparkCon } } - override def stop() { + def statusUpdate(taskId :Long, state: TaskState, serializedData: ByteBuffer) { + synchronized { + val taskSetId = taskIdToTaskSetId(taskId) + val taskSetManager = activeTaskSets(taskSetId) + taskSetTaskIds(taskSetId) -= taskId + taskSetManager.statusUpdate(taskId, state, serializedData) + } + } + + override def stop() { threadPool.shutdownNow() } diff --git a/core/src/main/scala/spark/scheduler/local/LocalTaskSetManager.scala b/core/src/main/scala/spark/scheduler/local/LocalTaskSetManager.scala new file mode 100644 index 0000000000..70b69bb26f --- /dev/null +++ b/core/src/main/scala/spark/scheduler/local/LocalTaskSetManager.scala @@ -0,0 +1,172 @@ +package spark.scheduler.local + +import java.io.File +import java.util.concurrent.atomic.AtomicInteger +import java.nio.ByteBuffer +import scala.collection.mutable.ArrayBuffer +import scala.collection.mutable.HashMap +import scala.collection.mutable.HashSet + +import spark._ +import spark.TaskState.TaskState +import spark.scheduler._ +import spark.scheduler.cluster._ + +private[spark] class LocalTaskSetManager(sched: LocalScheduler, val taskSet: TaskSet) extends TaskSetManager with Logging { + var parent: Schedulable = null + var weight: Int = 1 + var minShare: Int = 0 + var runningTasks: Int = 0 + var priority: Int = taskSet.priority + var stageId: Int = taskSet.stageId + var name: String = "TaskSet_"+taskSet.stageId.toString + + + var failCount = new Array[Int](taskSet.tasks.size) + val taskInfos = new HashMap[Long, TaskInfo] + val numTasks = taskSet.tasks.size + var numFinished = 0 + val ser = SparkEnv.get.closureSerializer.newInstance() + val copiesRunning = new Array[Int](numTasks) + val finished = new Array[Boolean](numTasks) + val numFailures = new Array[Int](numTasks) + val MAX_TASK_FAILURES = sched.maxFailures + + def increaseRunningTasks(taskNum: Int): Unit = { + runningTasks += taskNum + if (parent != null) { + parent.increaseRunningTasks(taskNum) + } + } + + def decreaseRunningTasks(taskNum: Int): Unit = { + runningTasks -= taskNum + if (parent != null) { + parent.decreaseRunningTasks(taskNum) + } + } + + def addSchedulable(schedulable: Schedulable): Unit = { + //nothing + } + + def removeSchedulable(schedulable: Schedulable): Unit = { + //nothing + } + + def getSchedulableByName(name: String): Schedulable = { + return null + } + + def executorLost(executorId: String, host: String): Unit = { + //nothing + } + + def checkSpeculatableTasks(): Boolean = { + return true + } + + def getSortedTaskSetQueue(): ArrayBuffer[TaskSetManager] = { + var sortedTaskSetQueue = new ArrayBuffer[TaskSetManager] + sortedTaskSetQueue += this + return sortedTaskSetQueue + } + + def hasPendingTasks(): Boolean = { + return true + } + + def findTask(): Option[Int] = { + for (i <- 0 to numTasks-1) { + if (copiesRunning(i) == 0 && !finished(i)) { + return Some(i) + } + } + return None + } + + def slaveOffer(execId: String, hostPort: String, availableCpus: Double, overrideLocality: TaskLocality.TaskLocality = null): Option[TaskDescription] = { + SparkEnv.set(sched.env) + logDebug("availableCpus:%d,numFinished:%d,numTasks:%d".format(availableCpus.toInt, numFinished, numTasks)) + if (availableCpus > 0 && numFinished < numTasks) { + findTask() match { + case Some(index) => + val taskId = sched.attemptId.getAndIncrement() + val task = taskSet.tasks(index) + val info = new TaskInfo(taskId, index, System.currentTimeMillis(), "local", "local:1", TaskLocality.NODE_LOCAL) + taskInfos(taskId) = info + val bytes = Task.serializeWithDependencies(task, sched.sc.addedFiles, sched.sc.addedJars, ser) + logInfo("Size of task " + taskId + " is " + bytes.limit + " bytes") + val taskName = "task %s:%d".format(taskSet.id, index) + copiesRunning(index) += 1 + increaseRunningTasks(1) + return Some(new TaskDescription(taskId, null, taskName, bytes)) + case None => {} + } + } + return None + } + + def numPendingTasksForHostPort(hostPort: String): Int = { + return 0 + } + + def numRackLocalPendingTasksForHost(hostPort :String): Int = { + return 0 + } + + def numPendingTasksForHost(hostPort: String): Int = { + return 0 + } + + def statusUpdate(tid: Long, state: TaskState, serializedData: ByteBuffer) { + state match { + case TaskState.FINISHED => + taskEnded(tid, state, serializedData) + case TaskState.FAILED => + taskFailed(tid, state, serializedData) + case _ => {} + } + } + + def taskEnded(tid: Long, state: TaskState, serializedData: ByteBuffer) { + val info = taskInfos(tid) + val index = info.index + val task = taskSet.tasks(index) + info.markSuccessful() + val result = ser.deserialize[TaskResult[_]](serializedData, getClass.getClassLoader) + result.metrics.resultSize = serializedData.limit() + sched.listener.taskEnded(task, Success, result.value, result.accumUpdates, info, result.metrics) + numFinished += 1 + decreaseRunningTasks(1) + finished(index) = true + if (numFinished == numTasks) { + sched.taskSetFinished(this) + } + } + + def taskFailed(tid: Long, state: TaskState, serializedData: ByteBuffer) { + val info = taskInfos(tid) + val index = info.index + val task = taskSet.tasks(index) + info.markFailed() + decreaseRunningTasks(1) + val reason: ExceptionFailure = ser.deserialize[ExceptionFailure](serializedData, getClass.getClassLoader) + if (!finished(index)) { + copiesRunning(index) -= 1 + numFailures(index) += 1 + val locs = reason.stackTrace.map(loc => "\tat %s".format(loc.toString)) + logInfo("Loss was due to %s\n%s\n%s".format(reason.className, reason.description, locs.mkString("\n"))) + if (numFailures(index) > MAX_TASK_FAILURES) { + val errorMessage = "Task %s:%d failed more than %d times; aborting job %s".format(taskSet.id, index, 4, reason.description) + decreaseRunningTasks(runningTasks) + sched.listener.taskSetFailed(taskSet, errorMessage) + // need to delete failed Taskset from schedule queue + sched.taskSetFinished(this) + } + } + } + + def error(message: String) { + } +} diff --git a/core/src/main/scala/spark/serializer/Serializer.scala b/core/src/main/scala/spark/serializer/Serializer.scala index 50b086125a..2ad73b711d 100644 --- a/core/src/main/scala/spark/serializer/Serializer.scala +++ b/core/src/main/scala/spark/serializer/Serializer.scala @@ -1,10 +1,13 @@ package spark.serializer -import java.nio.ByteBuffer import java.io.{EOFException, InputStream, OutputStream} +import java.nio.ByteBuffer + import it.unimi.dsi.fastutil.io.FastByteArrayOutputStream + import spark.util.ByteBufferInputStream + /** * A serializer. Because some serialization libraries are not thread safe, this class is used to * create [[spark.serializer.SerializerInstance]] objects that do the actual serialization and are @@ -14,6 +17,7 @@ trait Serializer { def newInstance(): SerializerInstance } + /** * An instance of a serializer, for use by one thread at a time. */ @@ -45,6 +49,7 @@ trait SerializerInstance { } } + /** * A stream for writing serialized objects. */ @@ -61,6 +66,7 @@ trait SerializationStream { } } + /** * A stream for reading serialized objects. */ @@ -72,40 +78,18 @@ trait DeserializationStream { * Read the elements of this stream through an iterator. This can only be called once, as * reading each element will consume data from the input source. */ - def asIterator: Iterator[Any] = new Iterator[Any] { - var gotNext = false - var finished = false - var nextValue: Any = null - - private def getNext() { + def asIterator: Iterator[Any] = new spark.util.NextIterator[Any] { + override protected def getNext() = { try { - nextValue = readObject[Any]() + readObject[Any]() } catch { case eof: EOFException => finished = true } - gotNext = true } - override def hasNext: Boolean = { - if (!gotNext) { - getNext() - } - if (finished) { - close() - } - !finished - } - - override def next(): Any = { - if (!gotNext) { - getNext() - } - if (finished) { - throw new NoSuchElementException("End of stream") - } - gotNext = false - nextValue + override protected def close() { + DeserializationStream.this.close() } } } diff --git a/core/src/main/scala/spark/serializer/SerializerManager.scala b/core/src/main/scala/spark/serializer/SerializerManager.scala new file mode 100644 index 0000000000..60b2aac797 --- /dev/null +++ b/core/src/main/scala/spark/serializer/SerializerManager.scala @@ -0,0 +1,45 @@ +package spark.serializer + +import java.util.concurrent.ConcurrentHashMap + + +/** + * A service that returns a serializer object given the serializer's class name. If a previous + * instance of the serializer object has been created, the get method returns that instead of + * creating a new one. + */ +private[spark] class SerializerManager { + + private val serializers = new ConcurrentHashMap[String, Serializer] + private var _default: Serializer = _ + + def default = _default + + def setDefault(clsName: String): Serializer = { + _default = get(clsName) + _default + } + + def get(clsName: String): Serializer = { + if (clsName == null) { + default + } else { + var serializer = serializers.get(clsName) + if (serializer != null) { + // If the serializer has been created previously, reuse that. + serializer + } else this.synchronized { + // Otherwise, create a new one. But make sure no other thread has attempted + // to create another new one at the same time. + serializer = serializers.get(clsName) + if (serializer == null) { + val clsLoader = Thread.currentThread.getContextClassLoader + serializer = + Class.forName(clsName, true, clsLoader).newInstance().asInstanceOf[Serializer] + serializers.put(clsName, serializer) + } + serializer + } + } + } +} diff --git a/core/src/main/scala/spark/storage/BlockException.scala b/core/src/main/scala/spark/storage/BlockException.scala new file mode 100644 index 0000000000..f275d476df --- /dev/null +++ b/core/src/main/scala/spark/storage/BlockException.scala @@ -0,0 +1,5 @@ +package spark.storage + +private[spark] +case class BlockException(blockId: String, message: String) extends Exception(message) + diff --git a/core/src/main/scala/spark/storage/BlockFetchTracker.scala b/core/src/main/scala/spark/storage/BlockFetchTracker.scala new file mode 100644 index 0000000000..0718156b1b --- /dev/null +++ b/core/src/main/scala/spark/storage/BlockFetchTracker.scala @@ -0,0 +1,10 @@ +package spark.storage + +private[spark] trait BlockFetchTracker { + def totalBlocks : Int + def numLocalBlocks: Int + def numRemoteBlocks: Int + def remoteFetchTime : Long + def fetchWaitTime: Long + def remoteBytesRead : Long +} diff --git a/core/src/main/scala/spark/storage/BlockFetcherIterator.scala b/core/src/main/scala/spark/storage/BlockFetcherIterator.scala new file mode 100644 index 0000000000..bec876213e --- /dev/null +++ b/core/src/main/scala/spark/storage/BlockFetcherIterator.scala @@ -0,0 +1,330 @@ +package spark.storage + +import java.nio.ByteBuffer +import java.util.concurrent.LinkedBlockingQueue + +import scala.collection.mutable.ArrayBuffer +import scala.collection.mutable.HashSet +import scala.collection.mutable.Queue + +import io.netty.buffer.ByteBuf + +import spark.Logging +import spark.Utils +import spark.SparkException +import spark.network.BufferMessage +import spark.network.ConnectionManagerId +import spark.network.netty.ShuffleCopier +import spark.serializer.Serializer + + +/** + * A block fetcher iterator interface. There are two implementations: + * + * BasicBlockFetcherIterator: uses a custom-built NIO communication layer. + * NettyBlockFetcherIterator: uses Netty (OIO) as the communication layer. + * + * Eventually we would like the two to converge and use a single NIO-based communication layer, + * but extensive tests show that under some circumstances (e.g. large shuffles with lots of cores), + * NIO would perform poorly and thus the need for the Netty OIO one. + */ + +private[storage] +trait BlockFetcherIterator extends Iterator[(String, Option[Iterator[Any]])] + with Logging with BlockFetchTracker { + def initialize() +} + + +private[storage] +object BlockFetcherIterator { + + // A request to fetch one or more blocks, complete with their sizes + class FetchRequest(val address: BlockManagerId, val blocks: Seq[(String, Long)]) { + val size = blocks.map(_._2).sum + } + + // A result of a fetch. Includes the block ID, size in bytes, and a function to deserialize + // the block (since we want all deserializaton to happen in the calling thread); can also + // represent a fetch failure if size == -1. + class FetchResult(val blockId: String, val size: Long, val deserialize: () => Iterator[Any]) { + def failed: Boolean = size == -1 + } + + class BasicBlockFetcherIterator( + private val blockManager: BlockManager, + val blocksByAddress: Seq[(BlockManagerId, Seq[(String, Long)])], + serializer: Serializer) + extends BlockFetcherIterator { + + import blockManager._ + + private var _remoteBytesRead = 0l + private var _remoteFetchTime = 0l + private var _fetchWaitTime = 0l + + if (blocksByAddress == null) { + throw new IllegalArgumentException("BlocksByAddress is null") + } + + // Total number blocks fetched (local + remote). Also number of FetchResults expected + protected var _numBlocksToFetch = 0 + + protected var startTime = System.currentTimeMillis + + // This represents the number of local blocks, also counting zero-sized blocks + private var numLocal = 0 + // BlockIds for local blocks that need to be fetched. Excludes zero-sized blocks + protected val localBlocksToFetch = new ArrayBuffer[String]() + + // This represents the number of remote blocks, also counting zero-sized blocks + private var numRemote = 0 + // BlockIds for remote blocks that need to be fetched. Excludes zero-sized blocks + protected val remoteBlocksToFetch = new HashSet[String]() + + // A queue to hold our results. + protected val results = new LinkedBlockingQueue[FetchResult] + + // Queue of fetch requests to issue; we'll pull requests off this gradually to make sure that + // the number of bytes in flight is limited to maxBytesInFlight + private val fetchRequests = new Queue[FetchRequest] + + // Current bytes in flight from our requests + private var bytesInFlight = 0L + + protected def sendRequest(req: FetchRequest) { + logDebug("Sending request for %d blocks (%s) from %s".format( + req.blocks.size, Utils.memoryBytesToString(req.size), req.address.hostPort)) + val cmId = new ConnectionManagerId(req.address.host, req.address.port) + val blockMessageArray = new BlockMessageArray(req.blocks.map { + case (blockId, size) => BlockMessage.fromGetBlock(GetBlock(blockId)) + }) + bytesInFlight += req.size + val sizeMap = req.blocks.toMap // so we can look up the size of each blockID + val fetchStart = System.currentTimeMillis() + val future = connectionManager.sendMessageReliably(cmId, blockMessageArray.toBufferMessage) + future.onSuccess { + case Some(message) => { + val fetchDone = System.currentTimeMillis() + _remoteFetchTime += fetchDone - fetchStart + val bufferMessage = message.asInstanceOf[BufferMessage] + val blockMessageArray = BlockMessageArray.fromBufferMessage(bufferMessage) + for (blockMessage <- blockMessageArray) { + if (blockMessage.getType != BlockMessage.TYPE_GOT_BLOCK) { + throw new SparkException( + "Unexpected message " + blockMessage.getType + " received from " + cmId) + } + val blockId = blockMessage.getId + results.put(new FetchResult(blockId, sizeMap(blockId), + () => dataDeserialize(blockId, blockMessage.getData, serializer))) + _remoteBytesRead += req.size + logDebug("Got remote block " + blockId + " after " + Utils.getUsedTimeMs(startTime)) + } + } + case None => { + logError("Could not get block(s) from " + cmId) + for ((blockId, size) <- req.blocks) { + results.put(new FetchResult(blockId, -1, null)) + } + } + } + } + + protected def splitLocalRemoteBlocks(): ArrayBuffer[FetchRequest] = { + // Split local and remote blocks. Remote blocks are further split into FetchRequests of size + // at most maxBytesInFlight in order to limit the amount of data in flight. + val remoteRequests = new ArrayBuffer[FetchRequest] + for ((address, blockInfos) <- blocksByAddress) { + if (address == blockManagerId) { + numLocal = blockInfos.size + // Filter out zero-sized blocks + localBlocksToFetch ++= blockInfos.filter(_._2 != 0).map(_._1) + _numBlocksToFetch += localBlocksToFetch.size + } else { + numRemote += blockInfos.size + // Make our requests at least maxBytesInFlight / 5 in length; the reason to keep them + // smaller than maxBytesInFlight is to allow multiple, parallel fetches from up to 5 + // nodes, rather than blocking on reading output from one node. + val minRequestSize = math.max(maxBytesInFlight / 5, 1L) + logInfo("maxBytesInFlight: " + maxBytesInFlight + ", minRequest: " + minRequestSize) + val iterator = blockInfos.iterator + var curRequestSize = 0L + var curBlocks = new ArrayBuffer[(String, Long)] + while (iterator.hasNext) { + val (blockId, size) = iterator.next() + // Skip empty blocks + if (size > 0) { + curBlocks += ((blockId, size)) + remoteBlocksToFetch += blockId + _numBlocksToFetch += 1 + curRequestSize += size + } else if (size < 0) { + throw new BlockException(blockId, "Negative block size " + size) + } + if (curRequestSize >= minRequestSize) { + // Add this FetchRequest + remoteRequests += new FetchRequest(address, curBlocks) + curRequestSize = 0 + curBlocks = new ArrayBuffer[(String, Long)] + } + } + // Add in the final request + if (!curBlocks.isEmpty) { + remoteRequests += new FetchRequest(address, curBlocks) + } + } + } + logInfo("Getting " + _numBlocksToFetch + " non-zero-bytes blocks out of " + + totalBlocks + " blocks") + remoteRequests + } + + protected def getLocalBlocks() { + // Get the local blocks while remote blocks are being fetched. Note that it's okay to do + // these all at once because they will just memory-map some files, so they won't consume + // any memory that might exceed our maxBytesInFlight + for (id <- localBlocksToFetch) { + getLocalFromDisk(id, serializer) match { + case Some(iter) => { + // Pass 0 as size since it's not in flight + results.put(new FetchResult(id, 0, () => iter)) + logDebug("Got local block " + id) + } + case None => { + throw new BlockException(id, "Could not get block " + id + " from local machine") + } + } + } + } + + override def initialize() { + // Split local and remote blocks. + val remoteRequests = splitLocalRemoteBlocks() + // Add the remote requests into our queue in a random order + fetchRequests ++= Utils.randomize(remoteRequests) + + // Send out initial requests for blocks, up to our maxBytesInFlight + while (!fetchRequests.isEmpty && + (bytesInFlight == 0 || bytesInFlight + fetchRequests.front.size <= maxBytesInFlight)) { + sendRequest(fetchRequests.dequeue()) + } + + val numGets = remoteRequests.size - fetchRequests.size + logInfo("Started " + numGets + " remote gets in " + Utils.getUsedTimeMs(startTime)) + + // Get Local Blocks + startTime = System.currentTimeMillis + getLocalBlocks() + logDebug("Got local blocks in " + Utils.getUsedTimeMs(startTime) + " ms") + } + + //an iterator that will read fetched blocks off the queue as they arrive. + @volatile protected var resultsGotten = 0 + + override def hasNext: Boolean = resultsGotten < _numBlocksToFetch + + override def next(): (String, Option[Iterator[Any]]) = { + resultsGotten += 1 + val startFetchWait = System.currentTimeMillis() + val result = results.take() + val stopFetchWait = System.currentTimeMillis() + _fetchWaitTime += (stopFetchWait - startFetchWait) + if (! result.failed) bytesInFlight -= result.size + while (!fetchRequests.isEmpty && + (bytesInFlight == 0 || bytesInFlight + fetchRequests.front.size <= maxBytesInFlight)) { + sendRequest(fetchRequests.dequeue()) + } + (result.blockId, if (result.failed) None else Some(result.deserialize())) + } + + // Implementing BlockFetchTracker trait. + override def totalBlocks: Int = numLocal + numRemote + override def numLocalBlocks: Int = numLocal + override def numRemoteBlocks: Int = numRemote + override def remoteFetchTime: Long = _remoteFetchTime + override def fetchWaitTime: Long = _fetchWaitTime + override def remoteBytesRead: Long = _remoteBytesRead + } + // End of BasicBlockFetcherIterator + + class NettyBlockFetcherIterator( + blockManager: BlockManager, + blocksByAddress: Seq[(BlockManagerId, Seq[(String, Long)])], + serializer: Serializer) + extends BasicBlockFetcherIterator(blockManager, blocksByAddress, serializer) { + + import blockManager._ + + val fetchRequestsSync = new LinkedBlockingQueue[FetchRequest] + + private def startCopiers(numCopiers: Int): List[_ <: Thread] = { + (for ( i <- Range(0,numCopiers) ) yield { + val copier = new Thread { + override def run(){ + try { + while(!isInterrupted && !fetchRequestsSync.isEmpty) { + sendRequest(fetchRequestsSync.take()) + } + } catch { + case x: InterruptedException => logInfo("Copier Interrupted") + //case _ => throw new SparkException("Exception Throw in Shuffle Copier") + } + } + } + copier.start + copier + }).toList + } + + // keep this to interrupt the threads when necessary + private def stopCopiers() { + for (copier <- copiers) { + copier.interrupt() + } + } + + override protected def sendRequest(req: FetchRequest) { + + def putResult(blockId: String, blockSize: Long, blockData: ByteBuf) { + val fetchResult = new FetchResult(blockId, blockSize, + () => dataDeserialize(blockId, blockData.nioBuffer, serializer)) + results.put(fetchResult) + } + + logDebug("Sending request for %d blocks (%s) from %s".format( + req.blocks.size, Utils.memoryBytesToString(req.size), req.address.host)) + val cmId = new ConnectionManagerId(req.address.host, req.address.nettyPort) + val cpier = new ShuffleCopier + cpier.getBlocks(cmId, req.blocks, putResult) + logDebug("Sent request for remote blocks " + req.blocks + " from " + req.address.host ) + } + + private var copiers: List[_ <: Thread] = null + + override def initialize() { + // Split Local Remote Blocks and set numBlocksToFetch + val remoteRequests = splitLocalRemoteBlocks() + // Add the remote requests into our queue in a random order + for (request <- Utils.randomize(remoteRequests)) { + fetchRequestsSync.put(request) + } + + copiers = startCopiers(System.getProperty("spark.shuffle.copier.threads", "6").toInt) + logInfo("Started " + fetchRequestsSync.size + " remote gets in " + + Utils.getUsedTimeMs(startTime)) + + // Get Local Blocks + startTime = System.currentTimeMillis + getLocalBlocks() + logDebug("Got local blocks in " + Utils.getUsedTimeMs(startTime) + " ms") + } + + override def next(): (String, Option[Iterator[Any]]) = { + resultsGotten += 1 + val result = results.take() + // If all the results has been retrieved, copiers will exit automatically + (result.blockId, if (result.failed) None else Some(result.deserialize())) + } + } + // End of NettyBlockFetcherIterator +} diff --git a/core/src/main/scala/spark/storage/BlockManager.scala b/core/src/main/scala/spark/storage/BlockManager.scala index 2462721fb8..9b39d3aadf 100644 --- a/core/src/main/scala/spark/storage/BlockManager.scala +++ b/core/src/main/scala/spark/storage/BlockManager.scala @@ -2,10 +2,8 @@ package spark.storage import java.io.{InputStream, OutputStream} import java.nio.{ByteBuffer, MappedByteBuffer} -import java.util.concurrent.{ConcurrentHashMap, LinkedBlockingQueue} -import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet, Queue} -import scala.collection.JavaConversions._ +import scala.collection.mutable.{HashMap, ArrayBuffer, HashSet} import akka.actor.{ActorSystem, Cancellable, Props} import akka.dispatch.{Await, Future} @@ -16,7 +14,7 @@ import com.ning.compress.lzf.{LZFInputStream, LZFOutputStream} import it.unimi.dsi.fastutil.io.FastByteArrayOutputStream -import spark.{Logging, SizeEstimator, SparkEnv, SparkException, Utils} +import spark.{Logging, SparkEnv, SparkException, Utils} import spark.network._ import spark.serializer.Serializer import spark.util.{ByteBufferInputStream, IdGenerator, MetadataCleaner, TimeStampedHashMap} @@ -24,53 +22,89 @@ import spark.util.{ByteBufferInputStream, IdGenerator, MetadataCleaner, TimeStam import sun.nio.ch.DirectBuffer -private[spark] -case class BlockException(blockId: String, message: String, ex: Exception = null) -extends Exception(message) - -private[spark] -class BlockManager( +private[spark] class BlockManager( executorId: String, actorSystem: ActorSystem, val master: BlockManagerMaster, - val serializer: Serializer, + val defaultSerializer: Serializer, maxMemory: Long) extends Logging { - class BlockInfo(val level: StorageLevel, val tellMaster: Boolean) { - var pending: Boolean = true - var size: Long = -1L + private class BlockInfo(val level: StorageLevel, val tellMaster: Boolean) { + @volatile var pending: Boolean = true + @volatile var size: Long = -1L + @volatile var initThread: Thread = null + @volatile var failed = false + + setInitThread() + + private def setInitThread() { + // Set current thread as init thread - waitForReady will not block this thread + // (in case there is non trivial initialization which ends up calling waitForReady as part of + // initialization itself) + this.initThread = Thread.currentThread() + } - /** Wait for this BlockInfo to be marked as ready (i.e. block is finished writing) */ - def waitForReady() { - if (pending) { + /** + * Wait for this BlockInfo to be marked as ready (i.e. block is finished writing). + * Return true if the block is available, false otherwise. + */ + def waitForReady(): Boolean = { + if (initThread != Thread.currentThread() && pending) { synchronized { while (pending) this.wait() } } + !failed } /** Mark this BlockInfo as ready (i.e. block is finished writing) */ def markReady(sizeInBytes: Long) { - pending = false + assert (pending) size = sizeInBytes + initThread = null + failed = false + initThread = null + pending = false + synchronized { + this.notifyAll() + } + } + + /** Mark this BlockInfo as ready but failed */ + def markFailure() { + assert (pending) + size = 0 + initThread = null + failed = true + initThread = null + pending = false synchronized { this.notifyAll() } } } + val shuffleBlockManager = new ShuffleBlockManager(this) + private val blockInfo = new TimeStampedHashMap[String, BlockInfo] private[storage] val memoryStore: BlockStore = new MemoryStore(this, maxMemory) - private[storage] val diskStore: BlockStore = + private[storage] val diskStore: DiskStore = new DiskStore(this, System.getProperty("spark.local.dir", System.getProperty("java.io.tmpdir"))) + // If we use Netty for shuffle, start a new Netty-based shuffle sender service. + private val nettyPort: Int = { + val useNetty = System.getProperty("spark.shuffle.use.netty", "false").toBoolean + val nettyPortConfig = System.getProperty("spark.shuffle.sender.port", "0").toInt + if (useNetty) diskStore.startShuffleBlockSender(nettyPortConfig) else 0 + } + val connectionManager = new ConnectionManager(0) implicit val futureExecContext = connectionManager.futureExecContext val blockManagerId = BlockManagerId( - executorId, connectionManager.id.host, connectionManager.id.port) + executorId, connectionManager.id.host, connectionManager.id.port, nettyPort) // Max megabytes of data to keep in flight per reducer (to avoid over-allocating memory // for receiving shuffle outputs) @@ -86,9 +120,9 @@ class BlockManager( val heartBeatFrequency = BlockManager.getHeartBeatFrequencyFromSystemProperties - val host = System.getProperty("spark.hostname", Utils.localHostName()) + val hostPort = Utils.localHostPort() - val slaveActor = master.actorSystem.actorOf(Props(new BlockManagerSlaveActor(this)), + val slaveActor = actorSystem.actorOf(Props(new BlockManagerSlaveActor(this)), name = "BlockManagerActor" + BlockManager.ID_GENERATOR.next) // Pending reregistration action being executed asynchronously or null if none @@ -197,9 +231,12 @@ class BlockManager( * Tell the master about the current storage status of a block. This will send a block update * message reflecting the current status, *not* the desired storage level in its block info. * For example, a block with MEMORY_AND_DISK set might have fallen out to be only on disk. + * + * droppedMemorySize exists to account for when block is dropped from memory to disk (so it is still valid). + * This ensures that update in master will compensate for the increase in memory on slave. */ - def reportBlockStatus(blockId: String, info: BlockInfo) { - val needReregister = !tryToReportBlockStatus(blockId, info) + def reportBlockStatus(blockId: String, info: BlockInfo, droppedMemorySize: Long = 0L) { + val needReregister = !tryToReportBlockStatus(blockId, info, droppedMemorySize) if (needReregister) { logInfo("Got told to reregister updating block " + blockId) // Reregistering will report our new block for free. @@ -213,7 +250,7 @@ class BlockManager( * which will be true if the block was successfully recorded and false if * the slave needs to re-register. */ - private def tryToReportBlockStatus(blockId: String, info: BlockInfo): Boolean = { + private def tryToReportBlockStatus(blockId: String, info: BlockInfo, droppedMemorySize: Long = 0L): Boolean = { val (curLevel, inMemSize, onDiskSize, tellMaster) = info.synchronized { info.level match { case null => @@ -222,7 +259,7 @@ class BlockManager( val inMem = level.useMemory && memoryStore.contains(blockId) val onDisk = level.useDisk && diskStore.contains(blockId) val storageLevel = StorageLevel(onDisk, inMem, level.deserialized, level.replication) - val memSize = if (inMem) memoryStore.getSize(blockId) else 0L + val memSize = if (inMem) memoryStore.getSize(blockId) else droppedMemorySize val diskSize = if (onDisk) diskStore.getSize(blockId) else 0L (storageLevel, memSize, diskSize, info.tellMaster) } @@ -235,26 +272,24 @@ class BlockManager( } } - /** - * Get locations of the block. + * Get locations of an array of blocks. */ - def getLocations(blockId: String): Seq[String] = { + def getLocationBlockIds(blockIds: Array[String]): Array[Seq[BlockManagerId]] = { val startTimeMs = System.currentTimeMillis - var managers = master.getLocations(blockId) - val locations = managers.map(_.ip) - logDebug("Got block locations in " + Utils.getUsedTimeMs(startTimeMs)) - return locations + val locations = master.getLocations(blockIds).toArray + logDebug("Got multiple block location in " + Utils.getUsedTimeMs(startTimeMs)) + locations } /** - * Get locations of an array of blocks. + * A short-circuited method to get blocks directly from disk. This is used for getting + * shuffle blocks. It is safe to do so without a lock on block info since disk store + * never deletes (recent) items. */ - def getLocations(blockIds: Array[String]): Array[Seq[String]] = { - val startTimeMs = System.currentTimeMillis - val locations = master.getLocations(blockIds).map(_.map(_.ip).toSeq).toArray - logDebug("Got multiple block location in " + Utils.getUsedTimeMs(startTimeMs)) - return locations + def getLocalFromDisk(blockId: String, serializer: Serializer): Option[Iterator[Any]] = { + diskStore.getValues(blockId, serializer).orElse( + sys.error("Block " + blockId + " not found on disk, though it should be")) } /** @@ -262,22 +297,17 @@ class BlockManager( */ def getLocal(blockId: String): Option[Iterator[Any]] = { logDebug("Getting local block " + blockId) - - // As an optimization for map output fetches, if the block is for a shuffle, return it - // without acquiring a lock; the disk store never deletes (recent) items so this should work - if (blockId.startsWith("shuffle_")) { - return diskStore.getValues(blockId) match { - case Some(iterator) => - Some(iterator) - case None => - throw new Exception("Block " + blockId + " not found on disk, though it should be") - } - } - val info = blockInfo.get(blockId).orNull if (info != null) { info.synchronized { - info.waitForReady() // In case the block is still being put() by another thread + + // In the another thread is writing the block, wait for it to become ready. + if (!info.waitForReady()) { + // If we get here, the block write failed. + logWarning("Block " + blockId + " was marked as failure.") + return None + } + val level = info.level logDebug("Level for block " + blockId + " is " + level) @@ -317,6 +347,8 @@ class BlockManager( case Some(bytes) => // Put a copy of the block back in memory before returning it. Note that we can't // put the ByteBuffer returned by the disk store as that's a memory-mapped file. + // The use of rewind assumes this. + assert (0 == bytes.position()) val copyForMemory = ByteBuffer.allocate(bytes.limit) copyForMemory.put(bytes) memoryStore.putBytes(blockId, copyForMemory, level) @@ -350,7 +382,7 @@ class BlockManager( // As an optimization for map output fetches, if the block is for a shuffle, return it // without acquiring a lock; the disk store never deletes (recent) items so this should work - if (blockId.startsWith("shuffle_")) { + if (ShuffleBlockManager.isShuffle(blockId)) { return diskStore.getBytes(blockId) match { case Some(bytes) => Some(bytes) @@ -362,7 +394,14 @@ class BlockManager( val info = blockInfo.get(blockId).orNull if (info != null) { info.synchronized { - info.waitForReady() // In case the block is still being put() by another thread + + // In the another thread is writing the block, wait for it to become ready. + if (!info.waitForReady()) { + // If we get here, the block write failed. + logWarning("Block " + blockId + " was marked as failure.") + return None + } + val level = info.level logDebug("Level for block " + blockId + " is " + level) @@ -382,6 +421,7 @@ class BlockManager( // Read it as a byte buffer into memory first, then return it diskStore.getBytes(blockId) match { case Some(bytes) => + assert (0 == bytes.position()) if (level.useMemory) { if (level.deserialized) { memoryStore.putBytes(blockId, bytes, level) @@ -421,14 +461,13 @@ class BlockManager( for (loc <- locations) { logDebug("Getting remote block " + blockId + " from " + loc) val data = BlockManagerWorker.syncGetBlock( - GetBlock(blockId), ConnectionManagerId(loc.ip, loc.port)) + GetBlock(blockId), ConnectionManagerId(loc.host, loc.port)) if (data != null) { - logDebug("Data is not null: " + data) return Some(dataDeserialize(blockId, data)) } - logDebug("Data is null") + logDebug("The value of block " + blockId + " is null") } - logDebug("Data not found") + logDebug("Block " + blockId + " not found") return None } @@ -445,153 +484,19 @@ class BlockManager( * fashion as they're received. Expects a size in bytes to be provided for each block fetched, * so that we can control the maxMegabytesInFlight for the fetch. */ - def getMultiple(blocksByAddress: Seq[(BlockManagerId, Seq[(String, Long)])]) - : Iterator[(String, Option[Iterator[Any]])] = { - - if (blocksByAddress == null) { - throw new IllegalArgumentException("BlocksByAddress is null") - } - val totalBlocks = blocksByAddress.map(_._2.size).sum - logDebug("Getting " + totalBlocks + " blocks") - var startTime = System.currentTimeMillis - val localBlockIds = new ArrayBuffer[String]() - val remoteBlockIds = new HashSet[String]() - - // A result of a fetch. Includes the block ID, size in bytes, and a function to deserialize - // the block (since we want all deserializaton to happen in the calling thread); can also - // represent a fetch failure if size == -1. - class FetchResult(val blockId: String, val size: Long, val deserialize: () => Iterator[Any]) { - def failed: Boolean = size == -1 - } - - // A queue to hold our results. - val results = new LinkedBlockingQueue[FetchResult] - - // A request to fetch one or more blocks, complete with their sizes - class FetchRequest(val address: BlockManagerId, val blocks: Seq[(String, Long)]) { - val size = blocks.map(_._2).sum - } - - // Queue of fetch requests to issue; we'll pull requests off this gradually to make sure that - // the number of bytes in flight is limited to maxBytesInFlight - val fetchRequests = new Queue[FetchRequest] - - // Current bytes in flight from our requests - var bytesInFlight = 0L - - def sendRequest(req: FetchRequest) { - logDebug("Sending request for %d blocks (%s) from %s".format( - req.blocks.size, Utils.memoryBytesToString(req.size), req.address.ip)) - val cmId = new ConnectionManagerId(req.address.ip, req.address.port) - val blockMessageArray = new BlockMessageArray(req.blocks.map { - case (blockId, size) => BlockMessage.fromGetBlock(GetBlock(blockId)) - }) - bytesInFlight += req.size - val sizeMap = req.blocks.toMap // so we can look up the size of each blockID - val future = connectionManager.sendMessageReliably(cmId, blockMessageArray.toBufferMessage) - future.onSuccess { - case Some(message) => { - val bufferMessage = message.asInstanceOf[BufferMessage] - val blockMessageArray = BlockMessageArray.fromBufferMessage(bufferMessage) - for (blockMessage <- blockMessageArray) { - if (blockMessage.getType != BlockMessage.TYPE_GOT_BLOCK) { - throw new SparkException( - "Unexpected message " + blockMessage.getType + " received from " + cmId) - } - val blockId = blockMessage.getId - results.put(new FetchResult( - blockId, sizeMap(blockId), () => dataDeserialize(blockId, blockMessage.getData))) - logDebug("Got remote block " + blockId + " after " + Utils.getUsedTimeMs(startTime)) - } - } - case None => { - logError("Could not get block(s) from " + cmId) - for ((blockId, size) <- req.blocks) { - results.put(new FetchResult(blockId, -1, null)) - } - } - } - } + def getMultiple( + blocksByAddress: Seq[(BlockManagerId, Seq[(String, Long)])], serializer: Serializer) + : BlockFetcherIterator = { - // Partition local and remote blocks. Remote blocks are further split into FetchRequests of size - // at most maxBytesInFlight in order to limit the amount of data in flight. - val remoteRequests = new ArrayBuffer[FetchRequest] - for ((address, blockInfos) <- blocksByAddress) { - if (address == blockManagerId) { - localBlockIds ++= blockInfos.map(_._1) + val iter = + if (System.getProperty("spark.shuffle.use.netty", "false").toBoolean) { + new BlockFetcherIterator.NettyBlockFetcherIterator(this, blocksByAddress, serializer) } else { - remoteBlockIds ++= blockInfos.map(_._1) - // Make our requests at least maxBytesInFlight / 5 in length; the reason to keep them - // smaller than maxBytesInFlight is to allow multiple, parallel fetches from up to 5 - // nodes, rather than blocking on reading output from one node. - val minRequestSize = math.max(maxBytesInFlight / 5, 1L) - logInfo("maxBytesInFlight: " + maxBytesInFlight + ", minRequest: " + minRequestSize) - val iterator = blockInfos.iterator - var curRequestSize = 0L - var curBlocks = new ArrayBuffer[(String, Long)] - while (iterator.hasNext) { - val (blockId, size) = iterator.next() - curBlocks += ((blockId, size)) - curRequestSize += size - if (curRequestSize >= minRequestSize) { - // Add this FetchRequest - remoteRequests += new FetchRequest(address, curBlocks) - curRequestSize = 0 - curBlocks = new ArrayBuffer[(String, Long)] - } - } - // Add in the final request - if (!curBlocks.isEmpty) { - remoteRequests += new FetchRequest(address, curBlocks) - } + new BlockFetcherIterator.BasicBlockFetcherIterator(this, blocksByAddress, serializer) } - } - // Add the remote requests into our queue in a random order - fetchRequests ++= Utils.randomize(remoteRequests) - - // Send out initial requests for blocks, up to our maxBytesInFlight - while (!fetchRequests.isEmpty && - (bytesInFlight == 0 || bytesInFlight + fetchRequests.front.size <= maxBytesInFlight)) { - sendRequest(fetchRequests.dequeue()) - } - - val numGets = remoteBlockIds.size - fetchRequests.size - logInfo("Started " + numGets + " remote gets in " + Utils.getUsedTimeMs(startTime)) - // Get the local blocks while remote blocks are being fetched. Note that it's okay to do - // these all at once because they will just memory-map some files, so they won't consume - // any memory that might exceed our maxBytesInFlight - startTime = System.currentTimeMillis - for (id <- localBlockIds) { - getLocal(id) match { - case Some(iter) => { - results.put(new FetchResult(id, 0, () => iter)) // Pass 0 as size since it's not in flight - logDebug("Got local block " + id) - } - case None => { - throw new BlockException(id, "Could not get block " + id + " from local machine") - } - } - } - logDebug("Got local blocks in " + Utils.getUsedTimeMs(startTime) + " ms") - - // Return an iterator that will read fetched blocks off the queue as they arrive. - return new Iterator[(String, Option[Iterator[Any]])] { - var resultsGotten = 0 - - def hasNext: Boolean = resultsGotten < totalBlocks - - def next(): (String, Option[Iterator[Any]]) = { - resultsGotten += 1 - val result = results.take() - bytesInFlight -= result.size - while (!fetchRequests.isEmpty && - (bytesInFlight == 0 || bytesInFlight + fetchRequests.front.size <= maxBytesInFlight)) { - sendRequest(fetchRequests.dequeue()) - } - (result.blockId, if (result.failed) None else Some(result.deserialize())) - } - } + iter.initialize() + iter } def put(blockId: String, values: Iterator[Any], level: StorageLevel, tellMaster: Boolean) @@ -602,6 +507,22 @@ class BlockManager( } /** + * A short circuited method to get a block writer that can write data directly to disk. + * This is currently used for writing shuffle files out. Callers should handle error + * cases. + */ + def getDiskBlockWriter(blockId: String, serializer: Serializer, bufferSize: Int) + : BlockObjectWriter = { + val writer = diskStore.getBlockWriter(blockId, serializer, bufferSize) + writer.registerCloseEventHandler(() => { + val myInfo = new BlockInfo(StorageLevel.DISK_ONLY, false) + blockInfo.put(blockId, myInfo) + myInfo.markReady(writer.size()) + }) + writer + } + + /** * Put a new block of values to the block manager. Returns its (estimated) size in bytes. */ def put(blockId: String, values: ArrayBuffer[Any], level: StorageLevel, @@ -617,18 +538,26 @@ class BlockManager( throw new IllegalArgumentException("Storage level is null or invalid") } - val oldBlock = blockInfo.get(blockId).orNull - if (oldBlock != null) { - logWarning("Block " + blockId + " already exists on this machine; not re-adding it") - oldBlock.waitForReady() - return oldBlock.size - } - // Remember the block's storage level so that we can correctly drop it to disk if it needs // to be dropped right after it got put into memory. Note, however, that other threads will // not be able to get() this block until we call markReady on its BlockInfo. - val myInfo = new BlockInfo(level, tellMaster) - blockInfo.put(blockId, myInfo) + val myInfo = { + val tinfo = new BlockInfo(level, tellMaster) + // Do atomically ! + val oldBlockOpt = blockInfo.putIfAbsent(blockId, tinfo) + + if (oldBlockOpt.isDefined) { + if (oldBlockOpt.get.waitForReady()) { + logWarning("Block " + blockId + " already exists on this machine; not re-adding it") + return oldBlockOpt.get.size + } + + // TODO: So the block info exists - but previous attempt to load it (?) failed. What do we do now ? Retry on it ? + oldBlockOpt.get + } else { + tinfo + } + } val startTimeMs = System.currentTimeMillis @@ -648,36 +577,50 @@ class BlockManager( logTrace("Put for block " + blockId + " took " + Utils.getUsedTimeMs(startTimeMs) + " to get into synchronized block") - if (level.useMemory) { - // Save it just to memory first, even if it also has useDisk set to true; we will later - // drop it to disk if the memory store can't hold it. - val res = memoryStore.putValues(blockId, values, level, true) - size = res.size - res.data match { - case Right(newBytes) => bytesAfterPut = newBytes - case Left(newIterator) => valuesAfterPut = newIterator - } - } else { - // Save directly to disk. - val askForBytes = level.replication > 1 // Don't get back the bytes unless we replicate them - val res = diskStore.putValues(blockId, values, level, askForBytes) - size = res.size - res.data match { - case Right(newBytes) => bytesAfterPut = newBytes - case _ => + var marked = false + try { + if (level.useMemory) { + // Save it just to memory first, even if it also has useDisk set to true; we will later + // drop it to disk if the memory store can't hold it. + val res = memoryStore.putValues(blockId, values, level, true) + size = res.size + res.data match { + case Right(newBytes) => bytesAfterPut = newBytes + case Left(newIterator) => valuesAfterPut = newIterator + } + } else { + // Save directly to disk. + // Don't get back the bytes unless we replicate them. + val askForBytes = level.replication > 1 + val res = diskStore.putValues(blockId, values, level, askForBytes) + size = res.size + res.data match { + case Right(newBytes) => bytesAfterPut = newBytes + case _ => + } } - } - // Now that the block is in either the memory or disk store, let other threads read it, - // and tell the master about it. - myInfo.markReady(size) - if (tellMaster) { - reportBlockStatus(blockId, myInfo) + // Now that the block is in either the memory or disk store, let other threads read it, + // and tell the master about it. + marked = true + myInfo.markReady(size) + if (tellMaster) { + reportBlockStatus(blockId, myInfo) + } + } finally { + // If we failed at putting the block to memory/disk, notify other possible readers + // that it has failed, and then remove it from the block info map. + if (! marked) { + // Note that the remove must happen before markFailure otherwise another thread + // could've inserted a new BlockInfo before we remove it. + blockInfo.remove(blockId) + myInfo.markFailure() + logWarning("Putting block " + blockId + " failed") + } } } logDebug("Put block " + blockId + " locally took " + Utils.getUsedTimeMs(startTimeMs)) - // Replicate block if required if (level.replication > 1) { val remoteStartTime = System.currentTimeMillis @@ -714,16 +657,26 @@ class BlockManager( throw new IllegalArgumentException("Storage level is null or invalid") } - if (blockInfo.contains(blockId)) { - logWarning("Block " + blockId + " already exists on this machine; not re-adding it") - return - } - // Remember the block's storage level so that we can correctly drop it to disk if it needs // to be dropped right after it got put into memory. Note, however, that other threads will // not be able to get() this block until we call markReady on its BlockInfo. - val myInfo = new BlockInfo(level, tellMaster) - blockInfo.put(blockId, myInfo) + val myInfo = { + val tinfo = new BlockInfo(level, tellMaster) + // Do atomically ! + val oldBlockOpt = blockInfo.putIfAbsent(blockId, tinfo) + + if (oldBlockOpt.isDefined) { + if (oldBlockOpt.get.waitForReady()) { + logWarning("Block " + blockId + " already exists on this machine; not re-adding it") + return + } + + // TODO: So the block info exists - but previous attempt to load it (?) failed. What do we do now ? Retry on it ? + oldBlockOpt.get + } else { + tinfo + } + } val startTimeMs = System.currentTimeMillis @@ -742,28 +695,41 @@ class BlockManager( logDebug("PutBytes for block " + blockId + " took " + Utils.getUsedTimeMs(startTimeMs) + " to get into synchronized block") - if (level.useMemory) { - // Store it only in memory at first, even if useDisk is also set to true - bytes.rewind() - memoryStore.putBytes(blockId, bytes, level) - } else { - bytes.rewind() - diskStore.putBytes(blockId, bytes, level) - } + var marked = false + try { + if (level.useMemory) { + // Store it only in memory at first, even if useDisk is also set to true + bytes.rewind() + memoryStore.putBytes(blockId, bytes, level) + } else { + bytes.rewind() + diskStore.putBytes(blockId, bytes, level) + } - // Now that the block is in either the memory or disk store, let other threads read it, - // and tell the master about it. - myInfo.markReady(bytes.limit) - if (tellMaster) { - reportBlockStatus(blockId, myInfo) + // assert (0 == bytes.position(), "" + bytes) + + // Now that the block is in either the memory or disk store, let other threads read it, + // and tell the master about it. + marked = true + myInfo.markReady(bytes.limit) + if (tellMaster) { + reportBlockStatus(blockId, myInfo) + } + } finally { + // If we failed at putting the block to memory/disk, notify other possible readers + // that it has failed, and then remove it from the block info map. + if (! marked) { + // Note that the remove must happen before markFailure otherwise another thread + // could've inserted a new BlockInfo before we remove it. + blockInfo.remove(blockId) + myInfo.markFailure() + logWarning("Putting block " + blockId + " failed") + } } } // If replication had started, then wait for it to finish if (level.replication > 1) { - if (replicationFuture == null) { - throw new Exception("Unexpected") - } Await.ready(replicationFuture, Duration.Inf) } @@ -791,7 +757,7 @@ class BlockManager( logDebug("Try to replicate BlockId " + blockId + " once; The size of the data is " + data.limit() + " Bytes. To node: " + peer) if (!BlockManagerWorker.syncPutBlock(PutBlock(blockId, data, tLevel), - new ConnectionManagerId(peer.ip, peer.port))) { + new ConnectionManagerId(peer.host, peer.port))) { logError("Failed to call syncPutBlock to " + peer) } logDebug("Replicated BlockId " + blockId + " once used " + @@ -823,6 +789,14 @@ class BlockManager( val info = blockInfo.get(blockId).orNull if (info != null) { info.synchronized { + // required ? As of now, this will be invoked only for blocks which are ready + // But in case this changes in future, adding for consistency sake. + if (! info.waitForReady() ) { + // If we get here, the block write failed. + logWarning("Block " + blockId + " was marked as failure. Nothing to drop") + return + } + val level = info.level if (level.useDisk && !diskStore.contains(blockId)) { logInfo("Writing block " + blockId + " to disk") @@ -833,12 +807,13 @@ class BlockManager( diskStore.putBytes(blockId, bytes, level) } } + val droppedMemorySize = if (memoryStore.contains(blockId)) memoryStore.getSize(blockId) else 0L val blockWasRemoved = memoryStore.remove(blockId) if (!blockWasRemoved) { logWarning("Block " + blockId + " could not be dropped from memory as it does not exist") } if (info.tellMaster) { - reportBlockStatus(blockId, info) + reportBlockStatus(blockId, info, droppedMemorySize) } if (!level.useDisk) { // The block is completely gone from this node; forget it so we can put() it again later. @@ -851,9 +826,23 @@ class BlockManager( } /** + * Remove all blocks belonging to the given RDD. + * @return The number of blocks removed. + */ + def removeRdd(rddId: Int): Int = { + // TODO: Instead of doing a linear scan on the blockInfo map, create another map that maps + // from RDD.id to blocks. + logInfo("Removing RDD " + rddId) + val rddPrefix = "rdd_" + rddId + "_" + val blocksToRemove = blockInfo.filter(_._1.startsWith(rddPrefix)).map(_._1) + blocksToRemove.foreach(blockId => removeBlock(blockId, false)) + blocksToRemove.size + } + + /** * Remove a block from both memory and disk. */ - def removeBlock(blockId: String) { + def removeBlock(blockId: String, tellMaster: Boolean = true) { logInfo("Removing block " + blockId) val info = blockInfo.get(blockId).orNull if (info != null) info.synchronized { @@ -865,7 +854,7 @@ class BlockManager( "the disk or memory store") } blockInfo.remove(blockId) - if (info.tellMaster) { + if (tellMaster && info.tellMaster) { reportBlockStatus(blockId, info) } } else { @@ -898,7 +887,7 @@ class BlockManager( } def shouldCompress(blockId: String): Boolean = { - if (blockId.startsWith("shuffle_")) { + if (ShuffleBlockManager.isShuffle(blockId)) { compressShuffle } else if (blockId.startsWith("broadcast_")) { compressBroadcast @@ -913,7 +902,11 @@ class BlockManager( * Wrap an output stream for compression if block compression is enabled for its block type */ def wrapForCompression(blockId: String, s: OutputStream): OutputStream = { - if (shouldCompress(blockId)) new LZFOutputStream(s) else s + if (shouldCompress(blockId)) { + (new LZFOutputStream(s)).setFinishBlockOnFlush(true) + } else { + s + } } /** @@ -923,7 +916,10 @@ class BlockManager( if (shouldCompress(blockId)) new LZFInputStream(s) else s } - def dataSerialize(blockId: String, values: Iterator[Any]): ByteBuffer = { + def dataSerialize( + blockId: String, + values: Iterator[Any], + serializer: Serializer = defaultSerializer): ByteBuffer = { val byteStream = new FastByteArrayOutputStream(4096) val ser = serializer.newInstance() ser.serializeStream(wrapForCompression(blockId, byteStream)).writeAll(values).close() @@ -935,7 +931,10 @@ class BlockManager( * Deserializes a ByteBuffer into an iterator of values and disposes of it when the end of * the iterator is reached. */ - def dataDeserialize(blockId: String, bytes: ByteBuffer): Iterator[Any] = { + def dataDeserialize( + blockId: String, + bytes: ByteBuffer, + serializer: Serializer = defaultSerializer): Iterator[Any] = { bytes.rewind() val stream = wrapForCompression(blockId, new ByteBufferInputStream(bytes, true)) serializer.newInstance().deserializeStream(stream).asIterator @@ -946,7 +945,7 @@ class BlockManager( heartBeatTask.cancel() } connectionManager.stop() - master.actorSystem.stop(slaveActor) + actorSystem.stop(slaveActor) blockInfo.clear() memoryStore.clear() diskStore.clear() @@ -955,8 +954,8 @@ class BlockManager( } } -private[spark] -object BlockManager extends Logging { + +private[spark] object BlockManager extends Logging { val ID_GENERATOR = new IdGenerator @@ -966,7 +965,7 @@ object BlockManager extends Logging { } def getHeartBeatFrequencyFromSystemProperties: Long = - System.getProperty("spark.storage.blockManagerHeartBeatMs", "5000").toLong + System.getProperty("spark.storage.blockManagerTimeoutIntervalMs", "60000").toLong / 4 def getDisableHeartBeatsForTesting: Boolean = System.getProperty("spark.test.disableBlockManagerHeartBeat", "false").toBoolean @@ -985,4 +984,44 @@ object BlockManager extends Logging { } } } + + def blockIdsToExecutorLocations(blockIds: Array[String], env: SparkEnv, blockManagerMaster: BlockManagerMaster = null): HashMap[String, List[String]] = { + // env == null and blockManagerMaster != null is used in tests + assert (env != null || blockManagerMaster != null) + val locationBlockIds: Seq[Seq[BlockManagerId]] = + if (env != null) { + env.blockManager.getLocationBlockIds(blockIds) + } else { + blockManagerMaster.getLocations(blockIds) + } + + // Convert from block master locations to executor locations (we need that for task scheduling) + val executorLocations = new HashMap[String, List[String]]() + for (i <- 0 until blockIds.length) { + val blockId = blockIds(i) + val blockLocations = locationBlockIds(i) + + val executors = new HashSet[String]() + + if (env != null) { + for (bkLocation <- blockLocations) { + val executorHostPort = env.resolveExecutorIdToHostPort(bkLocation.executorId, bkLocation.host) + executors += executorHostPort + // logInfo("bkLocation = " + bkLocation + ", executorHostPort = " + executorHostPort) + } + } else { + // Typically while testing, etc - revert to simply using host. + for (bkLocation <- blockLocations) { + executors += bkLocation.host + // logInfo("bkLocation = " + bkLocation + ", executorHostPort = " + executorHostPort) + } + } + + executorLocations.put(blockId, executors.toSeq.toList) + } + + executorLocations + } + } + diff --git a/core/src/main/scala/spark/storage/BlockManagerId.scala b/core/src/main/scala/spark/storage/BlockManagerId.scala index f2f1e77d41..1e557d6148 100644 --- a/core/src/main/scala/spark/storage/BlockManagerId.scala +++ b/core/src/main/scala/spark/storage/BlockManagerId.scala @@ -2,51 +2,70 @@ package spark.storage import java.io.{Externalizable, IOException, ObjectInput, ObjectOutput} import java.util.concurrent.ConcurrentHashMap +import spark.Utils /** * This class represent an unique identifier for a BlockManager. * The first 2 constructors of this class is made private to ensure that - * BlockManagerId objects can be created only using the factory method in - * [[spark.storage.BlockManager$]]. This allows de-duplication of ID objects. + * BlockManagerId objects can be created only using the apply method in + * the companion object. This allows de-duplication of ID objects. * Also, constructor parameters are private to ensure that parameters cannot * be modified from outside this class. */ private[spark] class BlockManagerId private ( private var executorId_ : String, - private var ip_ : String, - private var port_ : Int + private var host_ : String, + private var port_ : Int, + private var nettyPort_ : Int ) extends Externalizable { - private def this() = this(null, null, 0) // For deserialization only + private def this() = this(null, null, 0, 0) // For deserialization only def executorId: String = executorId_ - def ip: String = ip_ + if (null != host_){ + Utils.checkHost(host_, "Expected hostname") + assert (port_ > 0) + } + + def hostPort: String = { + // DEBUG code + Utils.checkHost(host) + assert (port > 0) + + host + ":" + port + } + + def host: String = host_ def port: Int = port_ + def nettyPort: Int = nettyPort_ + override def writeExternal(out: ObjectOutput) { out.writeUTF(executorId_) - out.writeUTF(ip_) + out.writeUTF(host_) out.writeInt(port_) + out.writeInt(nettyPort_) } override def readExternal(in: ObjectInput) { executorId_ = in.readUTF() - ip_ = in.readUTF() + host_ = in.readUTF() port_ = in.readInt() + nettyPort_ = in.readInt() } @throws(classOf[IOException]) private def readResolve(): Object = BlockManagerId.getCachedBlockManagerId(this) - override def toString = "BlockManagerId(%s, %s, %d)".format(executorId, ip, port) + override def toString = "BlockManagerId(%s, %s, %d, %d)".format(executorId, host, port, nettyPort) - override def hashCode: Int = (executorId.hashCode * 41 + ip.hashCode) * 41 + port + override def hashCode: Int = (executorId.hashCode * 41 + host.hashCode) * 41 + port + nettyPort override def equals(that: Any) = that match { case id: BlockManagerId => - executorId == id.executorId && port == id.port && ip == id.ip + executorId == id.executorId && port == id.port && host == id.host && nettyPort == id.nettyPort case _ => false } @@ -55,8 +74,17 @@ private[spark] class BlockManagerId private ( private[spark] object BlockManagerId { - def apply(execId: String, ip: String, port: Int) = - getCachedBlockManagerId(new BlockManagerId(execId, ip, port)) + /** + * Returns a [[spark.storage.BlockManagerId]] for the given configuraiton. + * + * @param execId ID of the executor. + * @param host Host name of the block manager. + * @param port Port of the block manager. + * @param nettyPort Optional port for the Netty-based shuffle sender. + * @return A new [[spark.storage.BlockManagerId]]. + */ + def apply(execId: String, host: String, port: Int, nettyPort: Int) = + getCachedBlockManagerId(new BlockManagerId(execId, host, port, nettyPort)) def apply(in: ObjectInput) = { val obj = new BlockManagerId() @@ -67,11 +95,7 @@ private[spark] object BlockManagerId { val blockManagerIdCache = new ConcurrentHashMap[BlockManagerId, BlockManagerId]() def getCachedBlockManagerId(id: BlockManagerId): BlockManagerId = { - if (blockManagerIdCache.containsKey(id)) { - blockManagerIdCache.get(id) - } else { - blockManagerIdCache.put(id, id) - id - } + blockManagerIdCache.putIfAbsent(id, id) + blockManagerIdCache.get(id) } } diff --git a/core/src/main/scala/spark/storage/BlockManagerMaster.scala b/core/src/main/scala/spark/storage/BlockManagerMaster.scala index 7389bee150..58888b1ebb 100644 --- a/core/src/main/scala/spark/storage/BlockManagerMaster.scala +++ b/core/src/main/scala/spark/storage/BlockManagerMaster.scala @@ -1,46 +1,21 @@ package spark.storage -import java.io._ -import java.util.{HashMap => JHashMap} - -import scala.collection.JavaConverters._ -import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet} -import scala.util.Random - -import akka.actor.{Actor, ActorRef, ActorSystem, Props} -import akka.dispatch.Await +import akka.actor.ActorRef +import akka.dispatch.{Await, Future} import akka.pattern.ask -import akka.util.{Duration, Timeout} -import akka.util.duration._ +import akka.util.Duration + +import spark.{Logging, SparkException} -import spark.{Logging, SparkException, Utils} -private[spark] class BlockManagerMaster( - val actorSystem: ActorSystem, - isDriver: Boolean, - isLocal: Boolean, - driverIp: String, - driverPort: Int) - extends Logging { +private[spark] class BlockManagerMaster(var driverActor: ActorRef) extends Logging { val AKKA_RETRY_ATTEMPTS: Int = System.getProperty("spark.akka.num.retries", "3").toInt val AKKA_RETRY_INTERVAL_MS: Int = System.getProperty("spark.akka.retry.wait", "3000").toInt - val DRIVER_AKKA_ACTOR_NAME = "BlockMasterManager" - - val timeout = 10.seconds - var driverActor: ActorRef = { - if (isDriver) { - val driverActor = actorSystem.actorOf(Props(new BlockManagerMasterActor(isLocal)), - name = DRIVER_AKKA_ACTOR_NAME) - logInfo("Registered BlockManagerMaster Actor") - driverActor - } else { - val url = "akka://spark@%s:%s/user/%s".format(driverIp, driverPort, DRIVER_AKKA_ACTOR_NAME) - logInfo("Connecting to BlockManagerMaster: " + url) - actorSystem.actorFor(url) - } - } + val DRIVER_AKKA_ACTOR_NAME = "BlockManagerMaster" + + val timeout = Duration.create(System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds") /** Remove a dead executor from the driver actor. This is only called on the driver side. */ def removeExecutor(execId: String) { @@ -59,7 +34,7 @@ private[spark] class BlockManagerMaster( /** Register the BlockManager's id with the driver. */ def registerBlockManager( - blockManagerId: BlockManagerId, maxMemSize: Long, slaveActor: ActorRef) { + blockManagerId: BlockManagerId, maxMemSize: Long, slaveActor: ActorRef) { logInfo("Trying to register BlockManager") tell(RegisterBlockManager(blockManagerId, maxMemSize, slaveActor)) logInfo("Registered BlockManager") @@ -106,6 +81,19 @@ private[spark] class BlockManagerMaster( } /** + * Remove all blocks belonging to the given RDD. + */ + def removeRdd(rddId: Int, blocking: Boolean) { + val future = askDriverWithReply[Future[Seq[Int]]](RemoveRdd(rddId)) + future onFailure { + case e: Throwable => logError("Failed to remove RDD " + rddId, e) + } + if (blocking) { + Await.result(future, timeout) + } + } + + /** * Return the memory status for each block manager, in the form of a map from * the block manager's id to two long values. The first value is the maximum * amount of memory allocated for the block manager, while the second is the @@ -116,7 +104,7 @@ private[spark] class BlockManagerMaster( } def getStorageStatus: Array[StorageStatus] = { - askDriverWithReply[ArrayBuffer[StorageStatus]](GetStorageStatus).toArray + askDriverWithReply[Array[StorageStatus]](GetStorageStatus) } /** Stop the driver actor, called only on the Spark driver node */ @@ -153,7 +141,7 @@ private[spark] class BlockManagerMaster( val future = driverActor.ask(message)(timeout) val result = Await.result(future, timeout) if (result == null) { - throw new Exception("BlockManagerMaster returned null") + throw new SparkException("BlockManagerMaster returned null") } return result.asInstanceOf[T] } catch { diff --git a/core/src/main/scala/spark/storage/BlockManagerMasterActor.scala b/core/src/main/scala/spark/storage/BlockManagerMasterActor.scala index 2830bc6297..0d4384ba1f 100644 --- a/core/src/main/scala/spark/storage/BlockManagerMasterActor.scala +++ b/core/src/main/scala/spark/storage/BlockManagerMasterActor.scala @@ -2,15 +2,16 @@ package spark.storage import java.util.{HashMap => JHashMap} -import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet} +import scala.collection.mutable import scala.collection.JavaConversions._ -import scala.util.Random import akka.actor.{Actor, ActorRef, Cancellable} -import akka.util.{Duration, Timeout} +import akka.dispatch.Future +import akka.pattern.ask +import akka.util.Duration import akka.util.duration._ -import spark.{Logging, Utils} +import spark.{Logging, Utils, SparkException} /** * BlockManagerMasterActor is an actor on the master node to track statuses of @@ -21,13 +22,16 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { // Mapping from block manager id to the block manager's information. private val blockManagerInfo = - new HashMap[BlockManagerId, BlockManagerMasterActor.BlockManagerInfo] + new mutable.HashMap[BlockManagerId, BlockManagerMasterActor.BlockManagerInfo] // Mapping from executor ID to block manager ID. - private val blockManagerIdByExecutor = new HashMap[String, BlockManagerId] + private val blockManagerIdByExecutor = new mutable.HashMap[String, BlockManagerId] // Mapping from block id to the set of block managers that have the block. - private val blockLocations = new JHashMap[String, Pair[Int, HashSet[BlockManagerId]]] + private val blockLocations = new JHashMap[String, mutable.HashSet[BlockManagerId]] + + val akkaTimeout = Duration.create( + System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds") initLogging() @@ -35,7 +39,7 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { "" + (BlockManager.getHeartBeatFrequencyFromSystemProperties * 3)).toLong val checkTimeoutInterval = System.getProperty("spark.storage.blockManagerTimeoutIntervalMs", - "5000").toLong + "60000").toLong var timeoutCheckingTask: Cancellable = null @@ -50,28 +54,34 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { def receive = { case RegisterBlockManager(blockManagerId, maxMemSize, slaveActor) => register(blockManagerId, maxMemSize, slaveActor) + sender ! true case UpdateBlockInfo(blockManagerId, blockId, storageLevel, deserializedSize, size) => + // TODO: Ideally we want to handle all the message replies in receive instead of in the + // individual private methods. updateBlockInfo(blockManagerId, blockId, storageLevel, deserializedSize, size) case GetLocations(blockId) => - getLocations(blockId) + sender ! getLocations(blockId) case GetLocationsMultipleBlockIds(blockIds) => - getLocationsMultipleBlockIds(blockIds) + sender ! getLocationsMultipleBlockIds(blockIds) case GetPeers(blockManagerId, size) => - getPeersDeterministic(blockManagerId, size) - /*getPeers(blockManagerId, size)*/ + sender ! getPeers(blockManagerId, size) case GetMemoryStatus => - getMemoryStatus + sender ! memoryStatus case GetStorageStatus => - getStorageStatus + sender ! storageStatus + + case RemoveRdd(rddId) => + sender ! removeRdd(rddId) case RemoveBlock(blockId) => - removeBlock(blockId) + removeBlockFromWorkers(blockId) + sender ! true case RemoveExecutor(execId) => removeExecutor(execId) @@ -81,7 +91,7 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { logInfo("Stopping BlockManagerMaster") sender ! true if (timeoutCheckingTask != null) { - timeoutCheckingTask.cancel + timeoutCheckingTask.cancel() } context.stop(self) @@ -89,13 +99,36 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { expireDeadHosts() case HeartBeat(blockManagerId) => - heartBeat(blockManagerId) + sender ! heartBeat(blockManagerId) case other => - logInfo("Got unknown message: " + other) + logWarning("Got unknown message: " + other) + } + + private def removeRdd(rddId: Int): Future[Seq[Int]] = { + // First remove the metadata for the given RDD, and then asynchronously remove the blocks + // from the slaves. + + val prefix = "rdd_" + rddId + "_" + // Find all blocks for the given RDD, remove the block from both blockLocations and + // the blockManagerInfo that is tracking the blocks. + val blocks = blockLocations.keySet().filter(_.startsWith(prefix)) + blocks.foreach { blockId => + val bms: mutable.HashSet[BlockManagerId] = blockLocations.get(blockId) + bms.foreach(bm => blockManagerInfo.get(bm).foreach(_.removeBlock(blockId))) + blockLocations.remove(blockId) + } + + // Ask the slaves to remove the RDD, and put the result in a sequence of Futures. + // The dispatcher is used as an implicit argument into the Future sequence construction. + import context.dispatcher + val removeMsg = RemoveRdd(rddId) + Future.sequence(blockManagerInfo.values.map { bm => + bm.slaveActor.ask(removeMsg)(akkaTimeout).mapTo[Int] + }.toSeq) } - def removeBlockManager(blockManagerId: BlockManagerId) { + private def removeBlockManager(blockManagerId: BlockManagerId) { val info = blockManagerInfo(blockManagerId) // Remove the block manager from blockManagerIdByExecutor. @@ -106,7 +139,7 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { val iterator = info.blocks.keySet.iterator while (iterator.hasNext) { val blockId = iterator.next - val locations = blockLocations.get(blockId)._2 + val locations = blockLocations.get(blockId) locations -= blockManagerId if (locations.size == 0) { blockLocations.remove(locations) @@ -114,45 +147,41 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { } } - def expireDeadHosts() { + private def expireDeadHosts() { logTrace("Checking for hosts with no recent heart beats in BlockManagerMaster.") val now = System.currentTimeMillis() val minSeenTime = now - slaveTimeout - val toRemove = new HashSet[BlockManagerId] + val toRemove = new mutable.HashSet[BlockManagerId] for (info <- blockManagerInfo.values) { if (info.lastSeenMs < minSeenTime) { - logWarning("Removing BlockManager " + info.blockManagerId + " with no recent heart beats") + logWarning("Removing BlockManager " + info.blockManagerId + " with no recent heart beats: " + + (now - info.lastSeenMs) + "ms exceeds " + slaveTimeout + "ms") toRemove += info.blockManagerId } } toRemove.foreach(removeBlockManager) } - def removeExecutor(execId: String) { + private def removeExecutor(execId: String) { logInfo("Trying to remove executor " + execId + " from BlockManagerMaster.") blockManagerIdByExecutor.get(execId).foreach(removeBlockManager) - sender ! true } - def heartBeat(blockManagerId: BlockManagerId) { + private def heartBeat(blockManagerId: BlockManagerId): Boolean = { if (!blockManagerInfo.contains(blockManagerId)) { - if (blockManagerId.executorId == "<driver>" && !isLocal) { - sender ! true - } else { - sender ! false - } + blockManagerId.executorId == "<driver>" && !isLocal } else { blockManagerInfo(blockManagerId).updateLastSeenMs() - sender ! true + true } } // Remove a block from the slaves that have it. This can only be used to remove // blocks that the master knows about. - private def removeBlock(blockId: String) { - val block = blockLocations.get(blockId) - if (block != null) { - block._2.foreach { blockManagerId: BlockManagerId => + private def removeBlockFromWorkers(blockId: String) { + val locations = blockLocations.get(blockId) + if (locations != null) { + locations.foreach { blockManagerId: BlockManagerId => val blockManager = blockManagerInfo.get(blockManagerId) if (blockManager.isDefined) { // Remove the block from the slave's BlockManager. @@ -162,23 +191,20 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { } } } - sender ! true } // Return a map from the block manager id to max memory and remaining memory. - private def getMemoryStatus() { - val res = blockManagerInfo.map { case(blockManagerId, info) => + private def memoryStatus: Map[BlockManagerId, (Long, Long)] = { + blockManagerInfo.map { case(blockManagerId, info) => (blockManagerId, (info.maxMem, info.remainingMem)) }.toMap - sender ! res } - private def getStorageStatus() { - val res = blockManagerInfo.map { case(blockManagerId, info) => + private def storageStatus: Array[StorageStatus] = { + blockManagerInfo.map { case(blockManagerId, info) => import collection.JavaConverters._ StorageStatus(blockManagerId, info.maxMem, info.blocks.asScala.toMap) - } - sender ! res + }.toArray } private def register(id: BlockManagerId, maxMemSize: Long, slaveActor: ActorRef) { @@ -187,7 +213,8 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { } else if (!blockManagerInfo.contains(id)) { blockManagerIdByExecutor.get(id.executorId) match { case Some(manager) => - // A block manager of the same host name already exists + // A block manager of the same executor already exists. + // This should never happen. Let's just quit. logError("Got two different block manager registrations on " + id.executorId) System.exit(1) case None => @@ -196,7 +223,6 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { blockManagerInfo(id) = new BlockManagerMasterActor.BlockManagerInfo( id, System.currentTimeMillis(), maxMemSize, slaveActor) } - sender ! true } private def updateBlockInfo( @@ -225,12 +251,12 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { blockManagerInfo(blockManagerId).updateBlockInfo(blockId, storageLevel, memSize, diskSize) - var locations: HashSet[BlockManagerId] = null + var locations: mutable.HashSet[BlockManagerId] = null if (blockLocations.containsKey(blockId)) { - locations = blockLocations.get(blockId)._2 + locations = blockLocations.get(blockId) } else { - locations = new HashSet[BlockManagerId] - blockLocations.put(blockId, (storageLevel.replication, locations)) + locations = new mutable.HashSet[BlockManagerId] + blockLocations.put(blockId, locations) } if (storageLevel.isValid) { @@ -246,70 +272,24 @@ class BlockManagerMasterActor(val isLocal: Boolean) extends Actor with Logging { sender ! true } - private def getLocations(blockId: String) { - val startTimeMs = System.currentTimeMillis() - val tmp = " " + blockId + " " - if (blockLocations.containsKey(blockId)) { - var res: ArrayBuffer[BlockManagerId] = new ArrayBuffer[BlockManagerId] - res.appendAll(blockLocations.get(blockId)._2) - sender ! res.toSeq - } else { - var res: ArrayBuffer[BlockManagerId] = new ArrayBuffer[BlockManagerId] - sender ! res - } - } - - private def getLocationsMultipleBlockIds(blockIds: Array[String]) { - def getLocations(blockId: String): Seq[BlockManagerId] = { - val tmp = blockId - if (blockLocations.containsKey(blockId)) { - var res: ArrayBuffer[BlockManagerId] = new ArrayBuffer[BlockManagerId] - res.appendAll(blockLocations.get(blockId)._2) - return res.toSeq - } else { - var res: ArrayBuffer[BlockManagerId] = new ArrayBuffer[BlockManagerId] - return res.toSeq - } - } - - var res: ArrayBuffer[Seq[BlockManagerId]] = new ArrayBuffer[Seq[BlockManagerId]] - for (blockId <- blockIds) { - res.append(getLocations(blockId)) - } - sender ! res.toSeq + private def getLocations(blockId: String): Seq[BlockManagerId] = { + if (blockLocations.containsKey(blockId)) blockLocations.get(blockId).toSeq else Seq.empty } - private def getPeers(blockManagerId: BlockManagerId, size: Int) { - var peers: Array[BlockManagerId] = blockManagerInfo.keySet.toArray - var res: ArrayBuffer[BlockManagerId] = new ArrayBuffer[BlockManagerId] - res.appendAll(peers) - res -= blockManagerId - val rand = new Random(System.currentTimeMillis()) - while (res.length > size) { - res.remove(rand.nextInt(res.length)) - } - sender ! res.toSeq + private def getLocationsMultipleBlockIds(blockIds: Array[String]): Seq[Seq[BlockManagerId]] = { + blockIds.map(blockId => getLocations(blockId)) } - private def getPeersDeterministic(blockManagerId: BlockManagerId, size: Int) { - var peers: Array[BlockManagerId] = blockManagerInfo.keySet.toArray - var res: ArrayBuffer[BlockManagerId] = new ArrayBuffer[BlockManagerId] + private def getPeers(blockManagerId: BlockManagerId, size: Int): Seq[BlockManagerId] = { + val peers: Array[BlockManagerId] = blockManagerInfo.keySet.toArray val selfIndex = peers.indexOf(blockManagerId) if (selfIndex == -1) { - throw new Exception("Self index for " + blockManagerId + " not found") + throw new SparkException("Self index for " + blockManagerId + " not found") } // Note that this logic will select the same node multiple times if there aren't enough peers - var index = selfIndex - while (res.size < size) { - index += 1 - if (index == selfIndex) { - throw new Exception("More peer expected than available") - } - res += peers(index % peers.size) - } - sender ! res.toSeq + Array.tabulate[BlockManagerId](size) { i => peers((selfIndex + i + 1) % peers.length) }.toSeq } } @@ -332,8 +312,8 @@ object BlockManagerMasterActor { // Mapping from block id to its status. private val _blocks = new JHashMap[String, BlockStatus] - logInfo("Registering block manager %s:%d with %s RAM".format( - blockManagerId.ip, blockManagerId.port, Utils.memoryBytesToString(maxMem))) + logInfo("Registering block manager %s with %s RAM".format( + blockManagerId.hostPort, Utils.memoryBytesToString(maxMem))) def updateLastSeenMs() { _lastSeenMs = System.currentTimeMillis() @@ -358,13 +338,13 @@ object BlockManagerMasterActor { _blocks.put(blockId, BlockStatus(storageLevel, memSize, diskSize)) if (storageLevel.useMemory) { _remainingMem -= memSize - logInfo("Added %s in memory on %s:%d (size: %s, free: %s)".format( - blockId, blockManagerId.ip, blockManagerId.port, Utils.memoryBytesToString(memSize), + logInfo("Added %s in memory on %s (size: %s, free: %s)".format( + blockId, blockManagerId.hostPort, Utils.memoryBytesToString(memSize), Utils.memoryBytesToString(_remainingMem))) } if (storageLevel.useDisk) { - logInfo("Added %s on disk on %s:%d (size: %s)".format( - blockId, blockManagerId.ip, blockManagerId.port, Utils.memoryBytesToString(diskSize))) + logInfo("Added %s on disk on %s (size: %s)".format( + blockId, blockManagerId.hostPort, Utils.memoryBytesToString(diskSize))) } } else if (_blocks.containsKey(blockId)) { // If isValid is not true, drop the block. @@ -372,17 +352,24 @@ object BlockManagerMasterActor { _blocks.remove(blockId) if (blockStatus.storageLevel.useMemory) { _remainingMem += blockStatus.memSize - logInfo("Removed %s on %s:%d in memory (size: %s, free: %s)".format( - blockId, blockManagerId.ip, blockManagerId.port, Utils.memoryBytesToString(memSize), + logInfo("Removed %s on %s in memory (size: %s, free: %s)".format( + blockId, blockManagerId.hostPort, Utils.memoryBytesToString(memSize), Utils.memoryBytesToString(_remainingMem))) } if (blockStatus.storageLevel.useDisk) { - logInfo("Removed %s on %s:%d on disk (size: %s)".format( - blockId, blockManagerId.ip, blockManagerId.port, Utils.memoryBytesToString(diskSize))) + logInfo("Removed %s on %s on disk (size: %s)".format( + blockId, blockManagerId.hostPort, Utils.memoryBytesToString(diskSize))) } } } + def removeBlock(blockId: String) { + if (_blocks.containsKey(blockId)) { + _remainingMem += _blocks.get(blockId).memSize + _blocks.remove(blockId) + } + } + def remainingMem: Long = _remainingMem def lastSeenMs: Long = _lastSeenMs diff --git a/core/src/main/scala/spark/storage/BlockManagerMessages.scala b/core/src/main/scala/spark/storage/BlockManagerMessages.scala index 1494f90103..0010726c8d 100644 --- a/core/src/main/scala/spark/storage/BlockManagerMessages.scala +++ b/core/src/main/scala/spark/storage/BlockManagerMessages.scala @@ -16,6 +16,9 @@ sealed trait ToBlockManagerSlave private[spark] case class RemoveBlock(blockId: String) extends ToBlockManagerSlave +// Remove all blocks belonging to a specific RDD. +private[spark] case class RemoveRdd(rddId: Int) extends ToBlockManagerSlave + ////////////////////////////////////////////////////////////////////////////////// // Messages from slaves to the master. @@ -49,16 +52,16 @@ class UpdateBlockInfo( blockManagerId.writeExternal(out) out.writeUTF(blockId) storageLevel.writeExternal(out) - out.writeInt(memSize.toInt) - out.writeInt(diskSize.toInt) + out.writeLong(memSize) + out.writeLong(diskSize) } override def readExternal(in: ObjectInput) { blockManagerId = BlockManagerId(in) blockId = in.readUTF() storageLevel = StorageLevel(in) - memSize = in.readInt() - diskSize = in.readInt() + memSize = in.readLong() + diskSize = in.readLong() } } diff --git a/core/src/main/scala/spark/storage/BlockManagerSlaveActor.scala b/core/src/main/scala/spark/storage/BlockManagerSlaveActor.scala index f570cdc52d..b264d1deb5 100644 --- a/core/src/main/scala/spark/storage/BlockManagerSlaveActor.scala +++ b/core/src/main/scala/spark/storage/BlockManagerSlaveActor.scala @@ -11,6 +11,12 @@ import spark.{Logging, SparkException, Utils} */ class BlockManagerSlaveActor(blockManager: BlockManager) extends Actor { override def receive = { - case RemoveBlock(blockId) => blockManager.removeBlock(blockId) + + case RemoveBlock(blockId) => + blockManager.removeBlock(blockId) + + case RemoveRdd(rddId) => + val numBlocksRemoved = blockManager.removeRdd(rddId) + sender ! numBlocksRemoved } } diff --git a/core/src/main/scala/spark/storage/BlockManagerUI.scala b/core/src/main/scala/spark/storage/BlockManagerUI.scala index 9e6721ec17..e02281344a 100644 --- a/core/src/main/scala/spark/storage/BlockManagerUI.scala +++ b/core/src/main/scala/spark/storage/BlockManagerUI.scala @@ -1,7 +1,7 @@ package spark.storage import akka.actor.{ActorRef, ActorSystem} -import akka.util.Timeout +import akka.util.Duration import akka.util.duration._ import cc.spray.typeconversion.TwirlSupport._ import cc.spray.Directives @@ -19,20 +19,21 @@ class BlockManagerUI(val actorSystem: ActorSystem, blockManagerMaster: ActorRef, val STATIC_RESOURCE_DIR = "spark/deploy/static" - implicit val timeout = Timeout(10 seconds) + implicit val timeout = Duration.create(System.getProperty("spark.akka.askTimeout", "10").toLong, "seconds") + val host = Utils.localHostName() + val port = if (System.getProperty("spark.ui.port") != null) { + System.getProperty("spark.ui.port").toInt + } else { + // TODO: Unfortunately, it's not possible to pass port 0 to spray and figure out which + // random port it bound to, so we have to try to find a local one by creating a socket. + Utils.findFreePort() + } /** Start a HTTP server to run the Web interface */ def start() { try { - val port = if (System.getProperty("spark.ui.port") != null) { - System.getProperty("spark.ui.port").toInt - } else { - // TODO: Unfortunately, it's not possible to pass port 0 to spray and figure out which - // random port it bound to, so we have to try to find a local one by creating a socket. - Utils.findFreePort() - } AkkaUtils.startSprayServer(actorSystem, "0.0.0.0", port, handler, "BlockManagerHTTPServer") - logInfo("Started BlockManager web UI at http://%s:%d".format(Utils.localHostName(), port)) + logInfo("Started BlockManager web UI at http://%s:%d".format(host, port)) } catch { case e: Exception => logError("Failed to create BlockManager WebUI", e) @@ -73,4 +74,6 @@ class BlockManagerUI(val actorSystem: ActorSystem, blockManagerMaster: ActorRef, } } } + + private[spark] def appUIAddress = "http://" + host + ":" + port } diff --git a/core/src/main/scala/spark/storage/BlockManagerWorker.scala b/core/src/main/scala/spark/storage/BlockManagerWorker.scala index d2985559c1..3057ade233 100644 --- a/core/src/main/scala/spark/storage/BlockManagerWorker.scala +++ b/core/src/main/scala/spark/storage/BlockManagerWorker.scala @@ -2,13 +2,7 @@ package spark.storage import java.nio.ByteBuffer -import scala.actors._ -import scala.actors.Actor._ -import scala.actors.remote._ -import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet} -import scala.util.Random - -import spark.{Logging, Utils, SparkEnv} +import spark.{Logging, Utils} import spark.network._ /** @@ -19,7 +13,7 @@ import spark.network._ */ private[spark] class BlockManagerWorker(val blockManager: BlockManager) extends Logging { initLogging() - + blockManager.connectionManager.onReceiveMessage(onBlockMessageReceive) def onBlockMessageReceive(msg: Message, id: ConnectionManagerId): Option[Message] = { @@ -51,7 +45,7 @@ private[spark] class BlockManagerWorker(val blockManager: BlockManager) extends logDebug("Received [" + pB + "]") putBlock(pB.id, pB.data, pB.level) return None - } + } case BlockMessage.TYPE_GET_BLOCK => { val gB = new GetBlock(blockMessage.getId) logDebug("Received [" + gB + "]") @@ -88,30 +82,26 @@ private[spark] class BlockManagerWorker(val blockManager: BlockManager) extends private[spark] object BlockManagerWorker extends Logging { private var blockManagerWorker: BlockManagerWorker = null - private val DATA_TRANSFER_TIME_OUT_MS: Long = 500 - private val REQUEST_RETRY_INTERVAL_MS: Long = 1000 - + initLogging() - + def startBlockManagerWorker(manager: BlockManager) { blockManagerWorker = new BlockManagerWorker(manager) } - + def syncPutBlock(msg: PutBlock, toConnManagerId: ConnectionManagerId): Boolean = { val blockManager = blockManagerWorker.blockManager - val connectionManager = blockManager.connectionManager - val serializer = blockManager.serializer + val connectionManager = blockManager.connectionManager val blockMessage = BlockMessage.fromPutBlock(msg) val blockMessageArray = new BlockMessageArray(blockMessage) val resultMessage = connectionManager.sendMessageReliablySync( toConnManagerId, blockMessageArray.toBufferMessage) return (resultMessage != None) } - + def syncGetBlock(msg: GetBlock, toConnManagerId: ConnectionManagerId): ByteBuffer = { val blockManager = blockManagerWorker.blockManager - val connectionManager = blockManager.connectionManager - val serializer = blockManager.serializer + val connectionManager = blockManager.connectionManager val blockMessage = BlockMessage.fromGetBlock(msg) val blockMessageArray = new BlockMessageArray(blockMessage) val responseMessage = connectionManager.sendMessageReliablySync( diff --git a/core/src/main/scala/spark/storage/BlockMessageArray.scala b/core/src/main/scala/spark/storage/BlockMessageArray.scala index a25decb123..ee0c5ff9a2 100644 --- a/core/src/main/scala/spark/storage/BlockMessageArray.scala +++ b/core/src/main/scala/spark/storage/BlockMessageArray.scala @@ -115,6 +115,7 @@ private[spark] object BlockMessageArray { val newBuffer = ByteBuffer.allocate(totalSize) newBuffer.clear() bufferMessage.buffers.foreach(buffer => { + assert (0 == buffer.position()) newBuffer.put(buffer) buffer.rewind() }) diff --git a/core/src/main/scala/spark/storage/BlockObjectWriter.scala b/core/src/main/scala/spark/storage/BlockObjectWriter.scala new file mode 100644 index 0000000000..42e2b07d5c --- /dev/null +++ b/core/src/main/scala/spark/storage/BlockObjectWriter.scala @@ -0,0 +1,50 @@ +package spark.storage + +import java.nio.ByteBuffer + + +/** + * An interface for writing JVM objects to some underlying storage. This interface allows + * appending data to an existing block, and can guarantee atomicity in the case of faults + * as it allows the caller to revert partial writes. + * + * This interface does not support concurrent writes. + */ +abstract class BlockObjectWriter(val blockId: String) { + + var closeEventHandler: () => Unit = _ + + def open(): BlockObjectWriter + + def close() { + closeEventHandler() + } + + def isOpen: Boolean + + def registerCloseEventHandler(handler: () => Unit) { + closeEventHandler = handler + } + + /** + * Flush the partial writes and commit them as a single atomic block. Return the + * number of bytes written for this commit. + */ + def commit(): Long + + /** + * Reverts writes that haven't been flushed yet. Callers should invoke this function + * when there are runtime exceptions. + */ + def revertPartialWrites() + + /** + * Writes an object. + */ + def write(value: Any) + + /** + * Size of the valid writes, in bytes. + */ + def size(): Long +} diff --git a/core/src/main/scala/spark/storage/DiskStore.scala b/core/src/main/scala/spark/storage/DiskStore.scala index ddbf8821ad..da859eebcb 100644 --- a/core/src/main/scala/spark/storage/DiskStore.scala +++ b/core/src/main/scala/spark/storage/DiskStore.scala @@ -1,41 +1,126 @@ package spark.storage +import java.io.{File, FileOutputStream, OutputStream, RandomAccessFile} import java.nio.ByteBuffer -import java.io.{File, FileOutputStream, RandomAccessFile} +import java.nio.channels.FileChannel import java.nio.channels.FileChannel.MapMode import java.util.{Random, Date} import java.text.SimpleDateFormat -import it.unimi.dsi.fastutil.io.FastBufferedOutputStream - import scala.collection.mutable.ArrayBuffer -import spark.executor.ExecutorExitCode +import it.unimi.dsi.fastutil.io.FastBufferedOutputStream import spark.Utils +import spark.executor.ExecutorExitCode +import spark.serializer.{Serializer, SerializationStream} +import spark.Logging +import spark.network.netty.ShuffleSender +import spark.network.netty.PathResolver + /** * Stores BlockManager blocks on disk. */ private class DiskStore(blockManager: BlockManager, rootDirs: String) - extends BlockStore(blockManager) { + extends BlockStore(blockManager) with Logging { + + class DiskBlockObjectWriter(blockId: String, serializer: Serializer, bufferSize: Int) + extends BlockObjectWriter(blockId) { + + private val f: File = createFile(blockId /*, allowAppendExisting */) + + // The file channel, used for repositioning / truncating the file. + private var channel: FileChannel = null + private var bs: OutputStream = null + private var objOut: SerializationStream = null + private var lastValidPosition = 0L + private var initialized = false + + override def open(): DiskBlockObjectWriter = { + val fos = new FileOutputStream(f, true) + channel = fos.getChannel() + bs = blockManager.wrapForCompression(blockId, new FastBufferedOutputStream(fos, bufferSize)) + objOut = serializer.newInstance().serializeStream(bs) + initialized = true + this + } + + override def close() { + if (initialized) { + objOut.close() + bs.close() + channel = null + bs = null + objOut = null + } + // Invoke the close callback handler. + super.close() + } - val MAX_DIR_CREATION_ATTEMPTS: Int = 10 - val subDirsPerLocalDir = System.getProperty("spark.diskStore.subDirectories", "64").toInt + override def isOpen: Boolean = objOut != null + // Flush the partial writes, and set valid length to be the length of the entire file. + // Return the number of bytes written for this commit. + override def commit(): Long = { + if (initialized) { + // NOTE: Flush the serializer first and then the compressed/buffered output stream + objOut.flush() + bs.flush() + val prevPos = lastValidPosition + lastValidPosition = channel.position() + lastValidPosition - prevPos + } else { + // lastValidPosition is zero if stream is uninitialized + lastValidPosition + } + } + + override def revertPartialWrites() { + if (initialized) { + // Discard current writes. We do this by flushing the outstanding writes and + // truncate the file to the last valid position. + objOut.flush() + bs.flush() + channel.truncate(lastValidPosition) + } + } + + override def write(value: Any) { + if (!initialized) { + open() + } + objOut.writeObject(value) + } + + override def size(): Long = lastValidPosition + } + + private val MAX_DIR_CREATION_ATTEMPTS: Int = 10 + private val subDirsPerLocalDir = System.getProperty("spark.diskStore.subDirectories", "64").toInt + + private var shuffleSender : ShuffleSender = null // Create one local directory for each path mentioned in spark.local.dir; then, inside this // directory, create multiple subdirectories that we will hash files into, in order to avoid // having really large inodes at the top level. - val localDirs = createLocalDirs() - val subDirs = Array.fill(localDirs.length)(new Array[File](subDirsPerLocalDir)) + private val localDirs: Array[File] = createLocalDirs() + private val subDirs = Array.fill(localDirs.length)(new Array[File](subDirsPerLocalDir)) addShutdownHook() + def getBlockWriter(blockId: String, serializer: Serializer, bufferSize: Int) + : BlockObjectWriter = { + new DiskBlockObjectWriter(blockId, serializer, bufferSize) + } + override def getSize(blockId: String): Long = { getFile(blockId).length() } - override def putBytes(blockId: String, bytes: ByteBuffer, level: StorageLevel) { + override def putBytes(blockId: String, _bytes: ByteBuffer, level: StorageLevel) { + // So that we do not modify the input offsets ! + // duplicate does not copy buffer, so inexpensive + val bytes = _bytes.duplicate() logDebug("Attempting to put block " + blockId) val startTime = System.currentTimeMillis val file = createFile(blockId) @@ -49,6 +134,18 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) blockId, Utils.memoryBytesToString(bytes.limit), (finishTime - startTime))) } + private def getFileBytes(file: File): ByteBuffer = { + val length = file.length() + val channel = new RandomAccessFile(file, "r").getChannel() + val buffer = try { + channel.map(MapMode.READ_ONLY, 0, length) + } finally { + channel.close() + } + + buffer + } + override def putValues( blockId: String, values: ArrayBuffer[Any], @@ -61,18 +158,18 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) val file = createFile(blockId) val fileOut = blockManager.wrapForCompression(blockId, new FastBufferedOutputStream(new FileOutputStream(file))) - val objOut = blockManager.serializer.newInstance().serializeStream(fileOut) + val objOut = blockManager.defaultSerializer.newInstance().serializeStream(fileOut) objOut.writeAll(values.iterator) objOut.close() val length = file.length() + + val timeTaken = System.currentTimeMillis - startTime logDebug("Block %s stored as %s file on disk in %d ms".format( - blockId, Utils.memoryBytesToString(length), (System.currentTimeMillis - startTime))) + blockId, Utils.memoryBytesToString(length), timeTaken)) if (returnValues) { // Return a byte buffer for the contents of the file - val channel = new RandomAccessFile(file, "r").getChannel() - val buffer = channel.map(MapMode.READ_ONLY, 0, length) - channel.close() + val buffer = getFileBytes(file) PutResult(length, Right(buffer)) } else { PutResult(length, null) @@ -81,10 +178,7 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) override def getBytes(blockId: String): Option[ByteBuffer] = { val file = getFile(blockId) - val length = file.length().toInt - val channel = new RandomAccessFile(file, "r").getChannel() - val bytes = channel.map(MapMode.READ_ONLY, 0, length) - channel.close() + val bytes = getFileBytes(file) Some(bytes) } @@ -92,11 +186,18 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) getBytes(blockId).map(bytes => blockManager.dataDeserialize(blockId, bytes)) } + /** + * A version of getValues that allows a custom serializer. This is used as part of the + * shuffle short-circuit code. + */ + def getValues(blockId: String, serializer: Serializer): Option[Iterator[Any]] = { + getBytes(blockId).map(bytes => blockManager.dataDeserialize(blockId, bytes, serializer)) + } + override def remove(blockId: String): Boolean = { val file = getFile(blockId) if (file.exists()) { file.delete() - true } else { false } @@ -106,10 +207,13 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) getFile(blockId).exists() } - private def createFile(blockId: String): File = { + private def createFile(blockId: String, allowAppendExisting: Boolean = false): File = { val file = getFile(blockId) - if (file.exists()) { - throw new Exception("File for block " + blockId + " already exists on disk: " + file) + if (!allowAppendExisting && file.exists()) { + // NOTE(shivaram): Delete the file if it exists. This might happen if a ShuffleMap task + // was rescheduled on the same machine as the old task. + logWarning("File for block " + blockId + " already exists on disk: " + file + ". Deleting") + file.delete() } file } @@ -144,8 +248,8 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) private def createLocalDirs(): Array[File] = { logDebug("Creating local directories at root dirs '" + rootDirs + "'") val dateFormat = new SimpleDateFormat("yyyyMMddHHmmss") - rootDirs.split(",").map(rootDir => { - var foundLocalDir: Boolean = false + rootDirs.split(",").map { rootDir => + var foundLocalDir = false var localDir: File = null var localDirId: String = null var tries = 0 @@ -156,12 +260,11 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) localDirId = "%s-%04x".format(dateFormat.format(new Date), rand.nextInt(65536)) localDir = new File(rootDir, "spark-local-" + localDirId) if (!localDir.exists) { - localDir.mkdirs() - foundLocalDir = true + foundLocalDir = localDir.mkdirs() } } catch { case e: Exception => - logWarning("Attempt " + tries + " to create local dir failed", e) + logWarning("Attempt " + tries + " to create local dir " + localDir + " failed", e) } } if (!foundLocalDir) { @@ -171,19 +274,40 @@ private class DiskStore(blockManager: BlockManager, rootDirs: String) } logInfo("Created local directory at " + localDir) localDir - }) + } } private def addShutdownHook() { + localDirs.foreach(localDir => Utils.registerShutdownDeleteDir(localDir)) Runtime.getRuntime.addShutdownHook(new Thread("delete Spark local dirs") { override def run() { logDebug("Shutdown hook called") - try { - localDirs.foreach(localDir => Utils.deleteRecursively(localDir)) - } catch { - case t: Throwable => logError("Exception while deleting local spark dirs", t) + localDirs.foreach { localDir => + try { + if (!Utils.hasRootAsShutdownDeleteDir(localDir)) Utils.deleteRecursively(localDir) + } catch { + case t: Throwable => + logError("Exception while deleting local spark dir: " + localDir, t) + } + } + if (shuffleSender != null) { + shuffleSender.stop } } }) } + + private[storage] def startShuffleBlockSender(port: Int): Int = { + val pResolver = new PathResolver { + override def getAbsolutePath(blockId: String): String = { + if (!blockId.startsWith("shuffle_")) { + return null + } + DiskStore.this.getFile(blockId).getAbsolutePath() + } + } + shuffleSender = new ShuffleSender(port, pResolver) + logInfo("Created ShuffleSender binding to port : "+ shuffleSender.port) + shuffleSender.port + } } diff --git a/core/src/main/scala/spark/storage/MemoryStore.scala b/core/src/main/scala/spark/storage/MemoryStore.scala index ae88ff0bb1..eba5ee507f 100644 --- a/core/src/main/scala/spark/storage/MemoryStore.scala +++ b/core/src/main/scala/spark/storage/MemoryStore.scala @@ -31,9 +31,11 @@ private class MemoryStore(blockManager: BlockManager, maxMemory: Long) } } - override def putBytes(blockId: String, bytes: ByteBuffer, level: StorageLevel) { + override def putBytes(blockId: String, _bytes: ByteBuffer, level: StorageLevel) { + // Work on a duplicate - since the original input might be used elsewhere. + val bytes = _bytes.duplicate() + bytes.rewind() if (level.deserialized) { - bytes.rewind() val values = blockManager.dataDeserialize(blockId, bytes) val elements = new ArrayBuffer[Any] elements ++= values @@ -58,7 +60,7 @@ private class MemoryStore(blockManager: BlockManager, maxMemory: Long) } else { val bytes = blockManager.dataSerialize(blockId, values.iterator) tryToPut(blockId, bytes, bytes.limit, false) - PutResult(bytes.limit(), Right(bytes)) + PutResult(bytes.limit(), Right(bytes.duplicate())) } } diff --git a/core/src/main/scala/spark/storage/ShuffleBlockManager.scala b/core/src/main/scala/spark/storage/ShuffleBlockManager.scala new file mode 100644 index 0000000000..44638e0c2d --- /dev/null +++ b/core/src/main/scala/spark/storage/ShuffleBlockManager.scala @@ -0,0 +1,50 @@ +package spark.storage + +import spark.serializer.Serializer + + +private[spark] +class ShuffleWriterGroup(val id: Int, val writers: Array[BlockObjectWriter]) + + +private[spark] +trait ShuffleBlocks { + def acquireWriters(mapId: Int): ShuffleWriterGroup + def releaseWriters(group: ShuffleWriterGroup) +} + + +private[spark] +class ShuffleBlockManager(blockManager: BlockManager) { + + def forShuffle(shuffleId: Int, numBuckets: Int, serializer: Serializer): ShuffleBlocks = { + new ShuffleBlocks { + // Get a group of writers for a map task. + override def acquireWriters(mapId: Int): ShuffleWriterGroup = { + val bufferSize = System.getProperty("spark.shuffle.file.buffer.kb", "100").toInt * 1024 + val writers = Array.tabulate[BlockObjectWriter](numBuckets) { bucketId => + val blockId = ShuffleBlockManager.blockId(shuffleId, bucketId, mapId) + blockManager.getDiskBlockWriter(blockId, serializer, bufferSize) + } + new ShuffleWriterGroup(mapId, writers) + } + + override def releaseWriters(group: ShuffleWriterGroup) = { + // Nothing really to release here. + } + } + } +} + + +private[spark] +object ShuffleBlockManager { + + // Returns the block id for a given shuffle block. + def blockId(shuffleId: Int, bucketId: Int, groupId: Int): String = { + "shuffle_" + shuffleId + "_" + groupId + "_" + bucketId + } + + // Returns true if the block is a shuffle block. + def isShuffle(blockId: String): Boolean = blockId.startsWith("shuffle_") +} diff --git a/core/src/main/scala/spark/storage/StorageLevel.scala b/core/src/main/scala/spark/storage/StorageLevel.scala index 3b5a77ab22..cc0c354e7e 100644 --- a/core/src/main/scala/spark/storage/StorageLevel.scala +++ b/core/src/main/scala/spark/storage/StorageLevel.scala @@ -123,11 +123,7 @@ object StorageLevel { val storageLevelCache = new java.util.concurrent.ConcurrentHashMap[StorageLevel, StorageLevel]() private[spark] def getCachedStorageLevel(level: StorageLevel): StorageLevel = { - if (storageLevelCache.containsKey(level)) { - storageLevelCache.get(level) - } else { - storageLevelCache.put(level, level) - level - } + storageLevelCache.putIfAbsent(level, level) + storageLevelCache.get(level) } } diff --git a/core/src/main/scala/spark/storage/StorageUtils.scala b/core/src/main/scala/spark/storage/StorageUtils.scala index dec47a9d41..950c0cdf35 100644 --- a/core/src/main/scala/spark/storage/StorageUtils.scala +++ b/core/src/main/scala/spark/storage/StorageUtils.scala @@ -4,9 +4,9 @@ import spark.{Utils, SparkContext} import BlockManagerMasterActor.BlockStatus private[spark] -case class StorageStatus(blockManagerId: BlockManagerId, maxMem: Long, +case class StorageStatus(blockManagerId: BlockManagerId, maxMem: Long, blocks: Map[String, BlockStatus]) { - + def memUsed(blockPrefix: String = "") = { blocks.filterKeys(_.startsWith(blockPrefix)).values.map(_.memSize). reduceOption(_+_).getOrElse(0l) @@ -22,53 +22,62 @@ case class StorageStatus(blockManagerId: BlockManagerId, maxMem: Long, } case class RDDInfo(id: Int, name: String, storageLevel: StorageLevel, - numCachedPartitions: Int, numPartitions: Int, memSize: Long, diskSize: Long) { + numCachedPartitions: Int, numPartitions: Int, memSize: Long, diskSize: Long) + extends Ordered[RDDInfo] { override def toString = { import Utils.memoryBytesToString "RDD \"%s\" (%d) Storage: %s; CachedPartitions: %d; TotalPartitions: %d; MemorySize: %s; DiskSize: %s".format(name, id, storageLevel.toString, numCachedPartitions, numPartitions, memoryBytesToString(memSize), memoryBytesToString(diskSize)) } + + override def compare(that: RDDInfo) = { + this.id - that.id + } } /* Helper methods for storage-related objects */ private[spark] object StorageUtils { - /* Given the current storage status of the BlockManager, returns information for each RDD */ - def rddInfoFromStorageStatus(storageStatusList: Array[StorageStatus], + /* Given the current storage status of the BlockManager, returns information for each RDD */ + def rddInfoFromStorageStatus(storageStatusList: Array[StorageStatus], sc: SparkContext) : Array[RDDInfo] = { - rddInfoFromBlockStatusList(storageStatusList.flatMap(_.blocks).toMap, sc) + rddInfoFromBlockStatusList(storageStatusList.flatMap(_.blocks).toMap, sc) } - /* Given a list of BlockStatus objets, returns information for each RDD */ - def rddInfoFromBlockStatusList(infos: Map[String, BlockStatus], + /* Given a list of BlockStatus objets, returns information for each RDD */ + def rddInfoFromBlockStatusList(infos: Map[String, BlockStatus], sc: SparkContext) : Array[RDDInfo] = { // Group by rddId, ignore the partition name - val groupedRddBlocks = infos.groupBy { case(k, v) => + val groupedRddBlocks = infos.filterKeys(_.startsWith("rdd_")).groupBy { case(k, v) => k.substring(0,k.lastIndexOf('_')) }.mapValues(_.values.toArray) // For each RDD, generate an RDDInfo object - groupedRddBlocks.map { case(rddKey, rddBlocks) => - + val rddInfos = groupedRddBlocks.map { case (rddKey, rddBlocks) => // Add up memory and disk sizes val memSize = rddBlocks.map(_.memSize).reduce(_ + _) val diskSize = rddBlocks.map(_.diskSize).reduce(_ + _) // Find the id of the RDD, e.g. rdd_1 => 1 val rddId = rddKey.split("_").last.toInt - // Get the friendly name for the rdd, if available. - val rdd = sc.persistentRdds(rddId) - val rddName = Option(rdd.name).getOrElse(rddKey) - val rddStorageLevel = rdd.getStorageLevel - RDDInfo(rddId, rddName, rddStorageLevel, rddBlocks.length, rdd.partitions.size, memSize, diskSize) - }.toArray + // Get the friendly name and storage level for the RDD, if available + sc.persistentRdds.get(rddId).map { r => + val rddName = Option(r.name).getOrElse(rddKey) + val rddStorageLevel = r.getStorageLevel + RDDInfo(rddId, rddName, rddStorageLevel, rddBlocks.length, r.partitions.size, memSize, diskSize) + } + }.flatten.toArray + + scala.util.Sorting.quickSort(rddInfos) + + rddInfos } - /* Removes all BlockStatus object that are not part of a block prefix */ - def filterStorageStatusByPrefix(storageStatusList: Array[StorageStatus], + /* Removes all BlockStatus object that are not part of a block prefix */ + def filterStorageStatusByPrefix(storageStatusList: Array[StorageStatus], prefix: String) : Array[StorageStatus] = { storageStatusList.map { status => diff --git a/core/src/main/scala/spark/storage/ThreadingTest.scala b/core/src/main/scala/spark/storage/ThreadingTest.scala index a70d1c8e78..5c406e68cb 100644 --- a/core/src/main/scala/spark/storage/ThreadingTest.scala +++ b/core/src/main/scala/spark/storage/ThreadingTest.scala @@ -75,9 +75,8 @@ private[spark] object ThreadingTest { System.setProperty("spark.kryoserializer.buffer.mb", "1") val actorSystem = ActorSystem("test") val serializer = new KryoSerializer - val driverIp: String = System.getProperty("spark.driver.host", "localhost") - val driverPort: Int = System.getProperty("spark.driver.port", "7077").toInt - val blockManagerMaster = new BlockManagerMaster(actorSystem, true, true, driverIp, driverPort) + val blockManagerMaster = new BlockManagerMaster( + actorSystem.actorOf(Props(new BlockManagerMasterActor(true)))) val blockManager = new BlockManager( "<driver>", actorSystem, blockManagerMaster, serializer, 1024 * 1024) val producers = (1 to numProducers).map(i => new ProducerThread(blockManager, i)) diff --git a/core/src/main/scala/spark/util/AkkaUtils.scala b/core/src/main/scala/spark/util/AkkaUtils.scala index 30aec5a663..e93cc3b485 100644 --- a/core/src/main/scala/spark/util/AkkaUtils.scala +++ b/core/src/main/scala/spark/util/AkkaUtils.scala @@ -11,7 +11,7 @@ import cc.spray.{SprayCanRootService, HttpService} import cc.spray.can.server.HttpServer import cc.spray.io.pipelines.MessageHandlerDispatch.SingletonHandler import akka.dispatch.Await -import spark.SparkException +import spark.{Utils, SparkException} import java.util.concurrent.TimeoutException /** @@ -29,24 +29,30 @@ private[spark] object AkkaUtils { def createActorSystem(name: String, host: String, port: Int): (ActorSystem, Int) = { val akkaThreads = System.getProperty("spark.akka.threads", "4").toInt val akkaBatchSize = System.getProperty("spark.akka.batchSize", "15").toInt - val akkaTimeout = System.getProperty("spark.akka.timeout", "20").toInt + val akkaTimeout = System.getProperty("spark.akka.timeout", "60").toInt val akkaFrameSize = System.getProperty("spark.akka.frameSize", "10").toInt + val lifecycleEvents = if (System.getProperty("spark.akka.logLifecycleEvents", "false").toBoolean) "on" else "off" + // 10 seconds is the default akka timeout, but in a cluster, we need higher by default. + val akkaWriteTimeout = System.getProperty("spark.akka.writeTimeout", "30").toInt + val akkaConf = ConfigFactory.parseString(""" akka.daemonic = on akka.event-handlers = ["akka.event.slf4j.Slf4jEventHandler"] akka.stdout-loglevel = "ERROR" akka.actor.provider = "akka.remote.RemoteActorRefProvider" akka.remote.transport = "akka.remote.netty.NettyRemoteTransport" - akka.remote.log-remote-lifecycle-events = on akka.remote.netty.hostname = "%s" akka.remote.netty.port = %d akka.remote.netty.connection-timeout = %ds akka.remote.netty.message-frame-size = %d MiB akka.remote.netty.execution-pool-size = %d akka.actor.default-dispatcher.throughput = %d - """.format(host, port, akkaTimeout, akkaFrameSize, akkaThreads, akkaBatchSize)) + akka.remote.log-remote-lifecycle-events = %s + akka.remote.netty.write-timeout = %ds + """.format(host, port, akkaTimeout, akkaFrameSize, akkaThreads, akkaBatchSize, + lifecycleEvents, akkaWriteTimeout)) - val actorSystem = ActorSystem(name, akkaConf, getClass.getClassLoader) + val actorSystem = ActorSystem(name, akkaConf) // Figure out the port number we bound to, in case port was passed as 0. This is a bit of a // hack because Akka doesn't let you figure out the port through the public API yet. @@ -58,8 +64,9 @@ private[spark] object AkkaUtils { /** * Creates a Spray HTTP server bound to a given IP and port with a given Spray Route object to * handle requests. Returns the bound port or throws a SparkException on failure. + * TODO: Not changing ip to host here - is it required ? */ - def startSprayServer(actorSystem: ActorSystem, ip: String, port: Int, route: Route, + def startSprayServer(actorSystem: ActorSystem, ip: String, port: Int, route: Route, name: String = "HttpServer"): ActorRef = { val ioWorker = new IoWorker(actorSystem).start() val httpService = actorSystem.actorOf(Props(new HttpService(route))) diff --git a/core/src/main/scala/spark/util/BoundedPriorityQueue.scala b/core/src/main/scala/spark/util/BoundedPriorityQueue.scala new file mode 100644 index 0000000000..4bc5db8bb7 --- /dev/null +++ b/core/src/main/scala/spark/util/BoundedPriorityQueue.scala @@ -0,0 +1,45 @@ +package spark.util + +import java.io.Serializable +import java.util.{PriorityQueue => JPriorityQueue} +import scala.collection.generic.Growable +import scala.collection.JavaConverters._ + +/** + * Bounded priority queue. This class wraps the original PriorityQueue + * class and modifies it such that only the top K elements are retained. + * The top K elements are defined by an implicit Ordering[A]. + */ +class BoundedPriorityQueue[A](maxSize: Int)(implicit ord: Ordering[A]) + extends Iterable[A] with Growable[A] with Serializable { + + private val underlying = new JPriorityQueue[A](maxSize, ord) + + override def iterator: Iterator[A] = underlying.iterator.asScala + + override def ++=(xs: TraversableOnce[A]): this.type = { + xs.foreach { this += _ } + this + } + + override def +=(elem: A): this.type = { + if (size < maxSize) underlying.offer(elem) + else maybeReplaceLowest(elem) + this + } + + override def +=(elem1: A, elem2: A, elems: A*): this.type = { + this += elem1 += elem2 ++= elems + } + + override def clear() { underlying.clear() } + + private def maybeReplaceLowest(a: A): Boolean = { + val head = underlying.peek() + if (head != null && ord.gt(a, head)) { + underlying.poll() + underlying.offer(a) + } else false + } +} + diff --git a/core/src/main/scala/spark/util/CompletionIterator.scala b/core/src/main/scala/spark/util/CompletionIterator.scala new file mode 100644 index 0000000000..8139183780 --- /dev/null +++ b/core/src/main/scala/spark/util/CompletionIterator.scala @@ -0,0 +1,25 @@ +package spark.util + +/** + * Wrapper around an iterator which calls a completion method after it successfully iterates through all the elements + */ +abstract class CompletionIterator[+A, +I <: Iterator[A]](sub: I) extends Iterator[A]{ + def next = sub.next + def hasNext = { + val r = sub.hasNext + if (!r) { + completion + } + r + } + + def completion() +} + +object CompletionIterator { + def apply[A, I <: Iterator[A]](sub: I, completionFunction: => Unit) : CompletionIterator[A,I] = { + new CompletionIterator[A,I](sub) { + def completion() = completionFunction + } + } +}
\ No newline at end of file diff --git a/core/src/main/scala/spark/util/Distribution.scala b/core/src/main/scala/spark/util/Distribution.scala new file mode 100644 index 0000000000..24738b4307 --- /dev/null +++ b/core/src/main/scala/spark/util/Distribution.scala @@ -0,0 +1,65 @@ +package spark.util + +import java.io.PrintStream + +/** + * Util for getting some stats from a small sample of numeric values, with some handy summary functions. + * + * Entirely in memory, not intended as a good way to compute stats over large data sets. + * + * Assumes you are giving it a non-empty set of data + */ +class Distribution(val data: Array[Double], val startIdx: Int, val endIdx: Int) { + require(startIdx < endIdx) + def this(data: Traversable[Double]) = this(data.toArray, 0, data.size) + java.util.Arrays.sort(data, startIdx, endIdx) + val length = endIdx - startIdx + + val defaultProbabilities = Array(0,0.25,0.5,0.75,1.0) + + /** + * Get the value of the distribution at the given probabilities. Probabilities should be + * given from 0 to 1 + * @param probabilities + */ + def getQuantiles(probabilities: Traversable[Double] = defaultProbabilities) = { + probabilities.toIndexedSeq.map{p:Double => data(closestIndex(p))} + } + + private def closestIndex(p: Double) = { + math.min((p * length).toInt + startIdx, endIdx - 1) + } + + def showQuantiles(out: PrintStream = System.out) = { + out.println("min\t25%\t50%\t75%\tmax") + getQuantiles(defaultProbabilities).foreach{q => out.print(q + "\t")} + out.println + } + + def statCounter = StatCounter(data.slice(startIdx, endIdx)) + + /** + * print a summary of this distribution to the given PrintStream. + * @param out + */ + def summary(out: PrintStream = System.out) { + out.println(statCounter) + showQuantiles(out) + } +} + +object Distribution { + + def apply(data: Traversable[Double]): Option[Distribution] = { + if (data.size > 0) + Some(new Distribution(data)) + else + None + } + + def showQuantiles(out: PrintStream = System.out, quantiles: Traversable[Double]) { + out.println("min\t25%\t50%\t75%\tmax") + quantiles.foreach{q => out.print(q + "\t")} + out.println + } +}
\ No newline at end of file diff --git a/core/src/main/scala/spark/util/MetadataCleaner.scala b/core/src/main/scala/spark/util/MetadataCleaner.scala index a342d378ff..dafa906712 100644 --- a/core/src/main/scala/spark/util/MetadataCleaner.scala +++ b/core/src/main/scala/spark/util/MetadataCleaner.scala @@ -38,7 +38,7 @@ class MetadataCleaner(name: String, cleanupFunc: (Long) => Unit) extends Logging object MetadataCleaner { - def getDelaySeconds = System.getProperty("spark.cleaner.delay", "-1").toInt - def setDelaySeconds(delay: Int) { System.setProperty("spark.cleaner.delay", delay.toString) } + def getDelaySeconds = System.getProperty("spark.cleaner.ttl", "-1").toInt + def setDelaySeconds(delay: Int) { System.setProperty("spark.cleaner.ttl", delay.toString) } } diff --git a/core/src/main/scala/spark/util/NextIterator.scala b/core/src/main/scala/spark/util/NextIterator.scala new file mode 100644 index 0000000000..48b5018ddd --- /dev/null +++ b/core/src/main/scala/spark/util/NextIterator.scala @@ -0,0 +1,71 @@ +package spark.util + +/** Provides a basic/boilerplate Iterator implementation. */ +private[spark] abstract class NextIterator[U] extends Iterator[U] { + + private var gotNext = false + private var nextValue: U = _ + private var closed = false + protected var finished = false + + /** + * Method for subclasses to implement to provide the next element. + * + * If no next element is available, the subclass should set `finished` + * to `true` and may return any value (it will be ignored). + * + * This convention is required because `null` may be a valid value, + * and using `Option` seems like it might create unnecessary Some/None + * instances, given some iterators might be called in a tight loop. + * + * @return U, or set 'finished' when done + */ + protected def getNext(): U + + /** + * Method for subclasses to implement when all elements have been successfully + * iterated, and the iteration is done. + * + * <b>Note:</b> `NextIterator` cannot guarantee that `close` will be + * called because it has no control over what happens when an exception + * happens in the user code that is calling hasNext/next. + * + * Ideally you should have another try/catch, as in HadoopRDD, that + * ensures any resources are closed should iteration fail. + */ + protected def close() + + /** + * Calls the subclass-defined close method, but only once. + * + * Usually calling `close` multiple times should be fine, but historically + * there have been issues with some InputFormats throwing exceptions. + */ + def closeIfNeeded() { + if (!closed) { + close() + closed = true + } + } + + override def hasNext: Boolean = { + if (!finished) { + if (!gotNext) { + nextValue = getNext() + if (finished) { + closeIfNeeded() + } + gotNext = true + } + } + !finished + } + + override def next(): U = { + if (!hasNext) { + throw new NoSuchElementException("End of stream") + } + gotNext = false + nextValue + } +}
\ No newline at end of file diff --git a/core/src/main/scala/spark/util/StatCounter.scala b/core/src/main/scala/spark/util/StatCounter.scala index 5f80180339..2b980340b7 100644 --- a/core/src/main/scala/spark/util/StatCounter.scala +++ b/core/src/main/scala/spark/util/StatCounter.scala @@ -37,17 +37,23 @@ class StatCounter(values: TraversableOnce[Double]) extends Serializable { if (other == this) { merge(other.copy()) // Avoid overwriting fields in a weird order } else { - val delta = other.mu - mu - if (other.n * 10 < n) { - mu = mu + (delta * other.n) / (n + other.n) - } else if (n * 10 < other.n) { - mu = other.mu - (delta * n) / (n + other.n) - } else { - mu = (mu * n + other.mu * other.n) / (n + other.n) + if (n == 0) { + mu = other.mu + m2 = other.m2 + n = other.n + } else if (other.n != 0) { + val delta = other.mu - mu + if (other.n * 10 < n) { + mu = mu + (delta * other.n) / (n + other.n) + } else if (n * 10 < other.n) { + mu = other.mu - (delta * n) / (n + other.n) + } else { + mu = (mu * n + other.mu * other.n) / (n + other.n) + } + m2 += other.m2 + (delta * delta * n * other.n) / (n + other.n) + n += other.n } - m2 += other.m2 + (delta * delta * n * other.n) / (n + other.n) - n += other.n - this + this } } diff --git a/core/src/main/scala/spark/util/TimeStampedHashMap.scala b/core/src/main/scala/spark/util/TimeStampedHashMap.scala index 188f8910da..92dfaa6e6f 100644 --- a/core/src/main/scala/spark/util/TimeStampedHashMap.scala +++ b/core/src/main/scala/spark/util/TimeStampedHashMap.scala @@ -3,6 +3,7 @@ package spark.util import java.util.concurrent.ConcurrentHashMap import scala.collection.JavaConversions import scala.collection.mutable.Map +import spark.scheduler.MapStatus /** * This is a custom implementation of scala.collection.mutable.Map which stores the insertion @@ -42,6 +43,13 @@ class TimeStampedHashMap[A, B] extends Map[A, B]() with spark.Logging { this } + // Should we return previous value directly or as Option ? + def putIfAbsent(key: A, value: B): Option[B] = { + val prev = internalMap.putIfAbsent(key, (value, currentTime)) + if (prev != null) Some(prev._1) else None + } + + override def -= (key: A): this.type = { internalMap.remove(key) this diff --git a/core/src/main/scala/spark/util/Vector.scala b/core/src/main/scala/spark/util/Vector.scala index 03559751bc..835822edb2 100644 --- a/core/src/main/scala/spark/util/Vector.scala +++ b/core/src/main/scala/spark/util/Vector.scala @@ -11,12 +11,16 @@ class Vector(val elements: Array[Double]) extends Serializable { return Vector(length, i => this(i) + other(i)) } + def add(other: Vector) = this + other + def - (other: Vector): Vector = { if (length != other.length) throw new IllegalArgumentException("Vectors of different length") return Vector(length, i => this(i) - other(i)) } + def subtract(other: Vector) = this - other + def dot(other: Vector): Double = { if (length != other.length) throw new IllegalArgumentException("Vectors of different length") @@ -61,10 +65,16 @@ class Vector(val elements: Array[Double]) extends Serializable { this } + def addInPlace(other: Vector) = this +=other + def * (scale: Double): Vector = Vector(length, i => this(i) * scale) + def multiply (d: Double) = this * d + def / (d: Double): Vector = this * (1 / d) + def divide (d: Double) = this / d + def unary_- = this * -1 def sum = elements.reduceLeft(_ + _) diff --git a/core/src/main/twirl/spark/deploy/master/app_details.scala.html b/core/src/main/twirl/spark/deploy/master/app_details.scala.html index 301a7e2124..5e5e5de551 100644 --- a/core/src/main/twirl/spark/deploy/master/app_details.scala.html +++ b/core/src/main/twirl/spark/deploy/master/app_details.scala.html @@ -9,19 +9,17 @@ <li><strong>ID:</strong> @app.id</li> <li><strong>Description:</strong> @app.desc.name</li> <li><strong>User:</strong> @app.desc.user</li> - <li><strong>Cores:</strong> - @app.desc.cores - (@app.coresGranted Granted - @if(app.desc.cores == Integer.MAX_VALUE) { - + <li><strong>Cores:</strong> + @if(app.desc.maxCores == Integer.MAX_VALUE) { + Unlimited (@app.coresGranted granted) } else { - , @app.coresLeft + @app.desc.maxCores (@app.coresGranted granted, @app.coresLeft left) } - ) </li> <li><strong>Memory per Slave:</strong> @app.desc.memoryPerSlave</li> <li><strong>Submit Date:</strong> @app.submitDate</li> <li><strong>State:</strong> @app.state</li> + <li><strong><a href="@app.appUiUrl">Application Detail UI</a></strong></li> </ul> </div> </div> diff --git a/core/src/main/twirl/spark/deploy/master/executor_row.scala.html b/core/src/main/twirl/spark/deploy/master/executor_row.scala.html index d2d80fad48..21e72c7aab 100644 --- a/core/src/main/twirl/spark/deploy/master/executor_row.scala.html +++ b/core/src/main/twirl/spark/deploy/master/executor_row.scala.html @@ -3,7 +3,7 @@ <tr> <td>@executor.id</td> <td> - <a href="@executor.worker.webUiAddress">@executor.worker.id</href> + <a href="@executor.worker.webUiAddress">@executor.worker.id</a> </td> <td>@executor.cores</td> <td>@executor.memory</td> diff --git a/core/src/main/twirl/spark/deploy/master/index.scala.html b/core/src/main/twirl/spark/deploy/master/index.scala.html index ac51a39a51..b9b9f08810 100644 --- a/core/src/main/twirl/spark/deploy/master/index.scala.html +++ b/core/src/main/twirl/spark/deploy/master/index.scala.html @@ -2,7 +2,7 @@ @import spark.deploy.master._ @import spark.Utils -@spark.common.html.layout(title = "Spark Master on " + state.host) { +@spark.common.html.layout(title = "Spark Master on " + state.host + ":" + state.port) { <!-- Cluster Details --> <div class="row"> diff --git a/core/src/main/twirl/spark/deploy/master/worker_row.scala.html b/core/src/main/twirl/spark/deploy/master/worker_row.scala.html index be69e9bf02..46277ca421 100644 --- a/core/src/main/twirl/spark/deploy/master/worker_row.scala.html +++ b/core/src/main/twirl/spark/deploy/master/worker_row.scala.html @@ -4,7 +4,7 @@ <tr> <td> - <a href="@worker.webUiAddress">@worker.id</href> + <a href="@worker.webUiAddress">@worker.id</a> </td> <td>@{worker.host}:@{worker.port}</td> <td>@worker.state</td> diff --git a/core/src/main/twirl/spark/deploy/worker/index.scala.html b/core/src/main/twirl/spark/deploy/worker/index.scala.html index c39f769a73..0e66af9284 100644 --- a/core/src/main/twirl/spark/deploy/worker/index.scala.html +++ b/core/src/main/twirl/spark/deploy/worker/index.scala.html @@ -1,7 +1,7 @@ @(worker: spark.deploy.WorkerState) @import spark.Utils -@spark.common.html.layout(title = "Spark Worker on " + worker.host) { +@spark.common.html.layout(title = "Spark Worker on " + worker.host + ":" + worker.port) { <!-- Worker Details --> <div class="row"> diff --git a/core/src/main/twirl/spark/storage/worker_table.scala.html b/core/src/main/twirl/spark/storage/worker_table.scala.html index d54b8de4cc..cd72a688c1 100644 --- a/core/src/main/twirl/spark/storage/worker_table.scala.html +++ b/core/src/main/twirl/spark/storage/worker_table.scala.html @@ -12,7 +12,7 @@ <tbody> @for(status <- workersStatusList) { <tr> - <td>@(status.blockManagerId.ip + ":" + status.blockManagerId.port)</td> + <td>@(status.blockManagerId.host + ":" + status.blockManagerId.port)</td> <td> @(Utils.memoryBytesToString(status.memUsed(prefix))) (@(Utils.memoryBytesToString(status.memRemaining)) Total Available) diff --git a/core/src/test/resources/fairscheduler.xml b/core/src/test/resources/fairscheduler.xml new file mode 100644 index 0000000000..5a688b0ebb --- /dev/null +++ b/core/src/test/resources/fairscheduler.xml @@ -0,0 +1,14 @@ +<allocations> +<pool name="1"> + <minShare>2</minShare> + <weight>1</weight> + <schedulingMode>FIFO</schedulingMode> +</pool> +<pool name="2"> + <minShare>3</minShare> + <weight>1</weight> + <schedulingMode>FIFO</schedulingMode> +</pool> +<pool name="3"> +</pool> +</allocations> diff --git a/core/src/test/scala/spark/CheckpointSuite.scala b/core/src/test/scala/spark/CheckpointSuite.scala index 3e5ffa81d6..ca385972fb 100644 --- a/core/src/test/scala/spark/CheckpointSuite.scala +++ b/core/src/test/scala/spark/CheckpointSuite.scala @@ -162,6 +162,16 @@ class CheckpointSuite extends FunSuite with LocalSparkContext with Logging { rdd => new ZippedRDD(sc, rdd, rdd.map(x => x)), true, false) } + test("CheckpointRDD with zero partitions") { + val rdd = new BlockRDD[Int](sc, Array[String]()) + assert(rdd.partitions.size === 0) + assert(rdd.isCheckpointed === false) + rdd.checkpoint() + assert(rdd.count() === 0) + assert(rdd.isCheckpointed === true) + assert(rdd.partitions.size === 0) + } + /** * Test checkpointing of the final RDD generated by the given operation. By default, * this method tests whether the size of serialized RDD has reduced after checkpointing or not. diff --git a/core/src/test/scala/spark/DistributedSuite.scala b/core/src/test/scala/spark/DistributedSuite.scala index 0e2585daa4..0866fb47b3 100644 --- a/core/src/test/scala/spark/DistributedSuite.scala +++ b/core/src/test/scala/spark/DistributedSuite.scala @@ -1,9 +1,12 @@ package spark +import network.ConnectionManagerId import org.scalatest.FunSuite import org.scalatest.BeforeAndAfter +import org.scalatest.concurrent.Timeouts._ import org.scalatest.matchers.ShouldMatchers import org.scalatest.prop.Checkers +import org.scalatest.time.{Span, Millis} import org.scalacheck.Arbitrary._ import org.scalacheck.Gen import org.scalacheck.Prop._ @@ -13,9 +16,15 @@ import com.google.common.io.Files import scala.collection.mutable.ArrayBuffer import SparkContext._ -import storage.StorageLevel +import storage.{GetBlock, BlockManagerWorker, StorageLevel} -class DistributedSuite extends FunSuite with ShouldMatchers with BeforeAndAfter with LocalSparkContext { + +class NotSerializableClass +class NotSerializableExn(val notSer: NotSerializableClass) extends Throwable() {} + + +class DistributedSuite extends FunSuite with ShouldMatchers with BeforeAndAfter + with LocalSparkContext { val clusterUrl = "local-cluster[2,1,512]" @@ -24,6 +33,24 @@ class DistributedSuite extends FunSuite with ShouldMatchers with BeforeAndAfter System.clearProperty("spark.storage.memoryFraction") } + test("task throws not serializable exception") { + // Ensures that executors do not crash when an exn is not serializable. If executors crash, + // this test will hang. Correct behavior is that executors don't crash but fail tasks + // and the scheduler throws a SparkException. + + // numSlaves must be less than numPartitions + val numSlaves = 3 + val numPartitions = 10 + + sc = new SparkContext("local-cluster[%s,1,512]".format(numSlaves), "test") + val data = sc.parallelize(1 to 100, numPartitions). + map(x => throw new NotSerializableExn(new NotSerializableClass)) + intercept[SparkException] { + data.count() + } + resetSparkContext() + } + test("local-cluster format") { sc = new SparkContext("local-cluster[2,1,512]", "test") assert(sc.parallelize(1 to 2, 2).count() == 2) @@ -140,9 +167,22 @@ class DistributedSuite extends FunSuite with ShouldMatchers with BeforeAndAfter test("caching in memory and disk, serialized, replicated") { sc = new SparkContext(clusterUrl, "test") val data = sc.parallelize(1 to 1000, 10).persist(StorageLevel.MEMORY_AND_DISK_SER_2) + assert(data.count() === 1000) assert(data.count() === 1000) assert(data.count() === 1000) + + // Get all the locations of the first partition and try to fetch the partitions + // from those locations. + val blockIds = data.partitions.indices.map(index => "rdd_%d_%d".format(data.id, index)).toArray + val blockId = blockIds(0) + val blockManager = SparkEnv.get.blockManager + blockManager.master.getLocations(blockId).foreach(id => { + val bytes = BlockManagerWorker.syncGetBlock( + GetBlock(blockId), ConnectionManagerId(id.host, id.port)) + val deserialized = blockManager.dataDeserialize(blockId, bytes).asInstanceOf[Iterator[Int]].toList + assert(deserialized === (1 to 100).toList) + }) } test("compute without caching when no partitions fit in memory") { @@ -182,7 +222,6 @@ class DistributedSuite extends FunSuite with ShouldMatchers with BeforeAndAfter sc = new SparkContext(clusterUrl, "test") val data = sc.parallelize(Seq(true, true), 2) assert(data.count === 2) // force executors to start - val masterId = SparkEnv.get.blockManager.blockManagerId assert(data.map(markNodeIfIdentity).collect.size === 2) assert(data.map(failOnMarkedIdentity).collect.size === 2) } @@ -217,12 +256,69 @@ class DistributedSuite extends FunSuite with ShouldMatchers with BeforeAndAfter assert(grouped.collect.size === 1) } } + + test("recover from node failures with replication") { + import DistributedSuite.{markNodeIfIdentity, failOnMarkedIdentity} + DistributedSuite.amMaster = true + // Using more than two nodes so we don't have a symmetric communication pattern and might + // cache a partially correct list of peers. + sc = new SparkContext("local-cluster[3,1,512]", "test") + for (i <- 1 to 3) { + val data = sc.parallelize(Seq(true, false, false, false), 4) + data.persist(StorageLevel.MEMORY_ONLY_2) + + assert(data.count === 4) + assert(data.map(markNodeIfIdentity).collect.size === 4) + assert(data.map(failOnMarkedIdentity).collect.size === 4) + + // Create a new replicated RDD to make sure that cached peer information doesn't cause + // problems. + val data2 = sc.parallelize(Seq(true, true), 2).persist(StorageLevel.MEMORY_ONLY_2) + assert(data2.count === 2) + } + } + + test("unpersist RDDs") { + DistributedSuite.amMaster = true + sc = new SparkContext("local-cluster[3,1,512]", "test") + val data = sc.parallelize(Seq(true, false, false, false), 4) + data.persist(StorageLevel.MEMORY_ONLY_2) + data.count + assert(sc.persistentRdds.isEmpty === false) + data.unpersist() + assert(sc.persistentRdds.isEmpty === true) + + failAfter(Span(3000, Millis)) { + try { + while (! sc.getRDDStorageInfo.isEmpty) { + Thread.sleep(200) + } + } catch { + case _ => { Thread.sleep(10) } + // Do nothing. We might see exceptions because block manager + // is racing this thread to remove entries from the driver. + } + } + } + + test("job should fail if TaskResult exceeds Akka frame size") { + // We must use local-cluster mode since results are returned differently + // when running under LocalScheduler: + sc = new SparkContext("local-cluster[1,1,512]", "test") + val akkaFrameSize = + sc.env.actorSystem.settings.config.getBytes("akka.remote.netty.message-frame-size").toInt + val rdd = sc.parallelize(Seq(1)).map{x => new Array[Byte](akkaFrameSize)} + val exception = intercept[SparkException] { + rdd.reduce((x, y) => x) + } + exception.getMessage should endWith("result exceeded Akka frame size") + } } object DistributedSuite { // Indicates whether this JVM is marked for failure. var mark = false - + // Set by test to remember if we are in the driver program so we can assert // that we are not. var amMaster = false @@ -239,9 +335,9 @@ object DistributedSuite { // Act like an identity function, but if mark was set to true previously, fail, // crashing the entire JVM. def failOnMarkedIdentity(item: Boolean): Boolean = { - if (mark) { + if (mark) { System.exit(42) - } + } item - } + } } diff --git a/core/src/test/scala/spark/FileSuite.scala b/core/src/test/scala/spark/FileSuite.scala index 91b48c7456..e61ff7793d 100644 --- a/core/src/test/scala/spark/FileSuite.scala +++ b/core/src/test/scala/spark/FileSuite.scala @@ -7,6 +7,8 @@ import scala.io.Source import com.google.common.io.Files import org.scalatest.FunSuite import org.apache.hadoop.io._ +import org.apache.hadoop.io.compress.{DefaultCodec, CompressionCodec, GzipCodec} + import SparkContext._ @@ -26,6 +28,28 @@ class FileSuite extends FunSuite with LocalSparkContext { assert(sc.textFile(outputDir).collect().toList === List("1", "2", "3", "4")) } + test("text files (compressed)") { + sc = new SparkContext("local", "test") + val tempDir = Files.createTempDir() + val normalDir = new File(tempDir, "output_normal").getAbsolutePath + val compressedOutputDir = new File(tempDir, "output_compressed").getAbsolutePath + val codec = new DefaultCodec() + + val data = sc.parallelize("a" * 10000, 1) + data.saveAsTextFile(normalDir) + data.saveAsTextFile(compressedOutputDir, classOf[DefaultCodec]) + + val normalFile = new File(normalDir, "part-00000") + val normalContent = sc.textFile(normalDir).collect + assert(normalContent === Array.fill(10000)("a")) + + val compressedFile = new File(compressedOutputDir, "part-00000" + codec.getDefaultExtension) + val compressedContent = sc.textFile(compressedOutputDir).collect + assert(compressedContent === Array.fill(10000)("a")) + + assert(compressedFile.length < normalFile.length) + } + test("SequenceFiles") { sc = new SparkContext("local", "test") val tempDir = Files.createTempDir() @@ -37,6 +61,28 @@ class FileSuite extends FunSuite with LocalSparkContext { assert(output.map(_.toString).collect().toList === List("(1,a)", "(2,aa)", "(3,aaa)")) } + test("SequenceFile (compressed)") { + sc = new SparkContext("local", "test") + val tempDir = Files.createTempDir() + val normalDir = new File(tempDir, "output_normal").getAbsolutePath + val compressedOutputDir = new File(tempDir, "output_compressed").getAbsolutePath + val codec = new DefaultCodec() + + val data = sc.parallelize(Seq.fill(100)("abc"), 1).map(x => (x, x)) + data.saveAsSequenceFile(normalDir) + data.saveAsSequenceFile(compressedOutputDir, Some(classOf[DefaultCodec])) + + val normalFile = new File(normalDir, "part-00000") + val normalContent = sc.sequenceFile[String, String](normalDir).collect + assert(normalContent === Array.fill(100)("abc", "abc")) + + val compressedFile = new File(compressedOutputDir, "part-00000" + codec.getDefaultExtension) + val compressedContent = sc.sequenceFile[String, String](compressedOutputDir).collect + assert(compressedContent === Array.fill(100)("abc", "abc")) + + assert(compressedFile.length < normalFile.length) + } + test("SequenceFile with writable key") { sc = new SparkContext("local", "test") val tempDir = Files.createTempDir() diff --git a/core/src/test/scala/spark/JavaAPISuite.java b/core/src/test/scala/spark/JavaAPISuite.java index 9ffe7c5f99..d306124fca 100644 --- a/core/src/test/scala/spark/JavaAPISuite.java +++ b/core/src/test/scala/spark/JavaAPISuite.java @@ -8,6 +8,7 @@ import java.util.*; import scala.Tuple2; import com.google.common.base.Charsets; +import org.apache.hadoop.io.compress.DefaultCodec; import com.google.common.io.Files; import org.apache.hadoop.io.IntWritable; import org.apache.hadoop.io.Text; @@ -197,6 +198,28 @@ public class JavaAPISuite implements Serializable { } @Test + public void foldByKey() { + List<Tuple2<Integer, Integer>> pairs = Arrays.asList( + new Tuple2<Integer, Integer>(2, 1), + new Tuple2<Integer, Integer>(2, 1), + new Tuple2<Integer, Integer>(1, 1), + new Tuple2<Integer, Integer>(3, 2), + new Tuple2<Integer, Integer>(3, 1) + ); + JavaPairRDD<Integer, Integer> rdd = sc.parallelizePairs(pairs); + JavaPairRDD<Integer, Integer> sums = rdd.foldByKey(0, + new Function2<Integer, Integer, Integer>() { + @Override + public Integer call(Integer a, Integer b) { + return a + b; + } + }); + Assert.assertEquals(1, sums.lookup(1).get(0).intValue()); + Assert.assertEquals(2, sums.lookup(2).get(0).intValue()); + Assert.assertEquals(3, sums.lookup(3).get(0).intValue()); + } + + @Test public void reduceByKey() { List<Tuple2<Integer, Integer>> pairs = Arrays.asList( new Tuple2<Integer, Integer>(2, 1), @@ -423,7 +446,7 @@ public class JavaAPISuite implements Serializable { @Test public void iterator() { JavaRDD<Integer> rdd = sc.parallelize(Arrays.asList(1, 2, 3, 4, 5), 2); - TaskContext context = new TaskContext(0, 0, 0); + TaskContext context = new TaskContext(0, 0, 0, null); Assert.assertEquals(1, rdd.iterator(rdd.splits().get(0), context).next().intValue()); } @@ -452,6 +475,19 @@ public class JavaAPISuite implements Serializable { } @Test + public void textFilesCompressed() throws IOException { + File tempDir = Files.createTempDir(); + String outputDir = new File(tempDir, "output").getAbsolutePath(); + JavaRDD<Integer> rdd = sc.parallelize(Arrays.asList(1, 2, 3, 4)); + rdd.saveAsTextFile(outputDir, DefaultCodec.class); + + // Try reading it in as a text file RDD + List<String> expected = Arrays.asList("1", "2", "3", "4"); + JavaRDD<String> readRDD = sc.textFile(outputDir); + Assert.assertEquals(expected, readRDD.collect()); + } + + @Test public void sequenceFile() { File tempDir = Files.createTempDir(); String outputDir = new File(tempDir, "output").getAbsolutePath(); @@ -598,6 +634,37 @@ public class JavaAPISuite implements Serializable { } @Test + public void hadoopFileCompressed() { + File tempDir = Files.createTempDir(); + String outputDir = new File(tempDir, "output_compressed").getAbsolutePath(); + List<Tuple2<Integer, String>> pairs = Arrays.asList( + new Tuple2<Integer, String>(1, "a"), + new Tuple2<Integer, String>(2, "aa"), + new Tuple2<Integer, String>(3, "aaa") + ); + JavaPairRDD<Integer, String> rdd = sc.parallelizePairs(pairs); + + rdd.map(new PairFunction<Tuple2<Integer, String>, IntWritable, Text>() { + @Override + public Tuple2<IntWritable, Text> call(Tuple2<Integer, String> pair) { + return new Tuple2<IntWritable, Text>(new IntWritable(pair._1()), new Text(pair._2())); + } + }).saveAsHadoopFile(outputDir, IntWritable.class, Text.class, SequenceFileOutputFormat.class, + DefaultCodec.class); + + JavaPairRDD<IntWritable, Text> output = sc.hadoopFile(outputDir, + SequenceFileInputFormat.class, IntWritable.class, Text.class); + + Assert.assertEquals(pairs.toString(), output.map(new Function<Tuple2<IntWritable, Text>, + String>() { + @Override + public String call(Tuple2<IntWritable, Text> x) { + return x.toString(); + } + }).collect().toString()); + } + + @Test public void zip() { JavaRDD<Integer> rdd = sc.parallelize(Arrays.asList(1, 2, 3, 4, 5)); JavaDoubleRDD doubles = rdd.map(new DoubleFunction<Integer>() { @@ -611,6 +678,32 @@ public class JavaAPISuite implements Serializable { } @Test + public void zipPartitions() { + JavaRDD<Integer> rdd1 = sc.parallelize(Arrays.asList(1, 2, 3, 4, 5, 6), 2); + JavaRDD<String> rdd2 = sc.parallelize(Arrays.asList("1", "2", "3", "4"), 2); + FlatMapFunction2<Iterator<Integer>, Iterator<String>, Integer> sizesFn = + new FlatMapFunction2<Iterator<Integer>, Iterator<String>, Integer>() { + @Override + public Iterable<Integer> call(Iterator<Integer> i, Iterator<String> s) { + int sizeI = 0; + int sizeS = 0; + while (i.hasNext()) { + sizeI += 1; + i.next(); + } + while (s.hasNext()) { + sizeS += 1; + s.next(); + } + return Arrays.asList(sizeI, sizeS); + } + }; + + JavaRDD<Integer> sizes = rdd1.zipPartitions(sizesFn, rdd2); + Assert.assertEquals("[3, 2, 3, 2]", sizes.collect().toString()); + } + + @Test public void accumulators() { JavaRDD<Integer> rdd = sc.parallelize(Arrays.asList(1, 2, 3, 4, 5)); diff --git a/core/src/test/scala/spark/KryoSerializerSuite.scala b/core/src/test/scala/spark/KryoSerializerSuite.scala index 06d446ea24..327e2ff848 100644 --- a/core/src/test/scala/spark/KryoSerializerSuite.scala +++ b/core/src/test/scala/spark/KryoSerializerSuite.scala @@ -82,6 +82,7 @@ class KryoSerializerSuite extends FunSuite { check(mutable.HashMap(1 -> "one", 2 -> "two")) check(mutable.HashMap("one" -> 1, "two" -> 2)) check(List(Some(mutable.HashMap(1->1, 2->2)), None, Some(mutable.HashMap(3->4)))) + check(List(mutable.HashMap("one" -> 1, "two" -> 2),mutable.HashMap(1->"one",2->"two",3->"three"))) } test("custom registrator") { diff --git a/core/src/test/scala/spark/LocalSparkContext.scala b/core/src/test/scala/spark/LocalSparkContext.scala index ff00dd05dd..76d5258b02 100644 --- a/core/src/test/scala/spark/LocalSparkContext.scala +++ b/core/src/test/scala/spark/LocalSparkContext.scala @@ -27,6 +27,7 @@ object LocalSparkContext { sc.stop() // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") } /** Runs `f` by passing in `sc` and ensures that `sc` is stopped. */ @@ -38,4 +39,4 @@ object LocalSparkContext { } } -}
\ No newline at end of file +} diff --git a/core/src/test/scala/spark/MapOutputTrackerSuite.scala b/core/src/test/scala/spark/MapOutputTrackerSuite.scala index dd19442dcb..6e585e1c3a 100644 --- a/core/src/test/scala/spark/MapOutputTrackerSuite.scala +++ b/core/src/test/scala/spark/MapOutputTrackerSuite.scala @@ -8,7 +8,7 @@ import spark.storage.BlockManagerId import spark.util.AkkaUtils class MapOutputTrackerSuite extends FunSuite with LocalSparkContext { - + test("compressSize") { assert(MapOutputTracker.compressSize(0L) === 0) assert(MapOutputTracker.compressSize(1L) === 1) @@ -31,44 +31,47 @@ class MapOutputTrackerSuite extends FunSuite with LocalSparkContext { test("master start and stop") { val actorSystem = ActorSystem("test") - val tracker = new MapOutputTracker(actorSystem, true) + val tracker = new MapOutputTracker() + tracker.trackerActor = actorSystem.actorOf(Props(new MapOutputTrackerActor(tracker))) tracker.stop() } test("master register and fetch") { val actorSystem = ActorSystem("test") - val tracker = new MapOutputTracker(actorSystem, true) + val tracker = new MapOutputTracker() + tracker.trackerActor = actorSystem.actorOf(Props(new MapOutputTrackerActor(tracker))) tracker.registerShuffle(10, 2) val compressedSize1000 = MapOutputTracker.compressSize(1000L) val compressedSize10000 = MapOutputTracker.compressSize(10000L) val size1000 = MapOutputTracker.decompressSize(compressedSize1000) val size10000 = MapOutputTracker.decompressSize(compressedSize10000) - tracker.registerMapOutput(10, 0, new MapStatus(BlockManagerId("a", "hostA", 1000), + tracker.registerMapOutput(10, 0, new MapStatus(BlockManagerId("a", "hostA", 1000, 0), Array(compressedSize1000, compressedSize10000))) - tracker.registerMapOutput(10, 1, new MapStatus(BlockManagerId("b", "hostB", 1000), + tracker.registerMapOutput(10, 1, new MapStatus(BlockManagerId("b", "hostB", 1000, 0), Array(compressedSize10000, compressedSize1000))) val statuses = tracker.getServerStatuses(10, 0) - assert(statuses.toSeq === Seq((BlockManagerId("a", "hostA", 1000), size1000), - (BlockManagerId("b", "hostB", 1000), size10000))) + assert(statuses.toSeq === Seq((BlockManagerId("a", "hostA", 1000, 0), size1000), + (BlockManagerId("b", "hostB", 1000, 0), size10000))) tracker.stop() } test("master register and unregister and fetch") { val actorSystem = ActorSystem("test") - val tracker = new MapOutputTracker(actorSystem, true) + val tracker = new MapOutputTracker() + tracker.trackerActor = actorSystem.actorOf(Props(new MapOutputTrackerActor(tracker))) tracker.registerShuffle(10, 2) val compressedSize1000 = MapOutputTracker.compressSize(1000L) val compressedSize10000 = MapOutputTracker.compressSize(10000L) val size1000 = MapOutputTracker.decompressSize(compressedSize1000) val size10000 = MapOutputTracker.decompressSize(compressedSize10000) - tracker.registerMapOutput(10, 0, new MapStatus(BlockManagerId("a", "hostA", 1000), + tracker.registerMapOutput(10, 0, new MapStatus(BlockManagerId("a", "hostA", 1000, 0), Array(compressedSize1000, compressedSize1000, compressedSize1000))) - tracker.registerMapOutput(10, 1, new MapStatus(BlockManagerId("b", "hostB", 1000), + tracker.registerMapOutput(10, 1, new MapStatus(BlockManagerId("b", "hostB", 1000, 0), Array(compressedSize10000, compressedSize1000, compressedSize1000))) // As if we had two simulatenous fetch failures - tracker.unregisterMapOutput(10, 0, BlockManagerId("a", "hostA", 1000)) - tracker.unregisterMapOutput(10, 0, BlockManagerId("a", "hostA", 1000)) + tracker.unregisterMapOutput(10, 0, BlockManagerId("a", "hostA", 1000, 0)) + tracker.unregisterMapOutput(10, 0, BlockManagerId("a", "hostA", 1000, 0)) // The remaining reduce task might try to grab the output despite the shuffle failure; // this should cause it to fail, and the scheduler will ignore the failure due to the @@ -77,35 +80,40 @@ class MapOutputTrackerSuite extends FunSuite with LocalSparkContext { } test("remote fetch") { - try { - System.clearProperty("spark.driver.host") // In case some previous test had set it - val (actorSystem, boundPort) = AkkaUtils.createActorSystem("spark", "localhost", 0) - System.setProperty("spark.driver.port", boundPort.toString) - val masterTracker = new MapOutputTracker(actorSystem, true) - val slaveTracker = new MapOutputTracker(actorSystem, false) - masterTracker.registerShuffle(10, 1) - masterTracker.incrementGeneration() - slaveTracker.updateGeneration(masterTracker.getGeneration) - intercept[FetchFailedException] { slaveTracker.getServerStatuses(10, 0) } - - val compressedSize1000 = MapOutputTracker.compressSize(1000L) - val size1000 = MapOutputTracker.decompressSize(compressedSize1000) - masterTracker.registerMapOutput(10, 0, new MapStatus( - BlockManagerId("a", "hostA", 1000), Array(compressedSize1000))) - masterTracker.incrementGeneration() - slaveTracker.updateGeneration(masterTracker.getGeneration) - assert(slaveTracker.getServerStatuses(10, 0).toSeq === - Seq((BlockManagerId("a", "hostA", 1000), size1000))) - - masterTracker.unregisterMapOutput(10, 0, BlockManagerId("a", "hostA", 1000)) - masterTracker.incrementGeneration() - slaveTracker.updateGeneration(masterTracker.getGeneration) - intercept[FetchFailedException] { slaveTracker.getServerStatuses(10, 0) } - - // failure should be cached - intercept[FetchFailedException] { slaveTracker.getServerStatuses(10, 0) } - } finally { - System.clearProperty("spark.driver.port") - } + val hostname = "localhost" + val (actorSystem, boundPort) = AkkaUtils.createActorSystem("spark", hostname, 0) + System.setProperty("spark.driver.port", boundPort.toString) // Will be cleared by LocalSparkContext + System.setProperty("spark.hostPort", hostname + ":" + boundPort) + + val masterTracker = new MapOutputTracker() + masterTracker.trackerActor = actorSystem.actorOf( + Props(new MapOutputTrackerActor(masterTracker)), "MapOutputTracker") + + val (slaveSystem, _) = AkkaUtils.createActorSystem("spark-slave", hostname, 0) + val slaveTracker = new MapOutputTracker() + slaveTracker.trackerActor = slaveSystem.actorFor( + "akka://spark@localhost:" + boundPort + "/user/MapOutputTracker") + + masterTracker.registerShuffle(10, 1) + masterTracker.incrementGeneration() + slaveTracker.updateGeneration(masterTracker.getGeneration) + intercept[FetchFailedException] { slaveTracker.getServerStatuses(10, 0) } + + val compressedSize1000 = MapOutputTracker.compressSize(1000L) + val size1000 = MapOutputTracker.decompressSize(compressedSize1000) + masterTracker.registerMapOutput(10, 0, new MapStatus( + BlockManagerId("a", "hostA", 1000, 0), Array(compressedSize1000))) + masterTracker.incrementGeneration() + slaveTracker.updateGeneration(masterTracker.getGeneration) + assert(slaveTracker.getServerStatuses(10, 0).toSeq === + Seq((BlockManagerId("a", "hostA", 1000, 0), size1000))) + + masterTracker.unregisterMapOutput(10, 0, BlockManagerId("a", "hostA", 1000, 0)) + masterTracker.incrementGeneration() + slaveTracker.updateGeneration(masterTracker.getGeneration) + intercept[FetchFailedException] { slaveTracker.getServerStatuses(10, 0) } + + // failure should be cached + intercept[FetchFailedException] { slaveTracker.getServerStatuses(10, 0) } } } diff --git a/core/src/test/scala/spark/PartitioningSuite.scala b/core/src/test/scala/spark/PartitioningSuite.scala index af1107cd19..16f93e71a3 100644 --- a/core/src/test/scala/spark/PartitioningSuite.scala +++ b/core/src/test/scala/spark/PartitioningSuite.scala @@ -1,10 +1,10 @@ package spark import org.scalatest.FunSuite - import scala.collection.mutable.ArrayBuffer - import SparkContext._ +import spark.util.StatCounter +import scala.math.abs class PartitioningSuite extends FunSuite with LocalSparkContext { @@ -84,10 +84,10 @@ class PartitioningSuite extends FunSuite with LocalSparkContext { assert(grouped4.groupByKey(3).partitioner != grouped4.partitioner) assert(grouped4.groupByKey(4).partitioner === grouped4.partitioner) - assert(grouped2.join(grouped4).partitioner === grouped2.partitioner) - assert(grouped2.leftOuterJoin(grouped4).partitioner === grouped2.partitioner) - assert(grouped2.rightOuterJoin(grouped4).partitioner === grouped2.partitioner) - assert(grouped2.cogroup(grouped4).partitioner === grouped2.partitioner) + assert(grouped2.join(grouped4).partitioner === grouped4.partitioner) + assert(grouped2.leftOuterJoin(grouped4).partitioner === grouped4.partitioner) + assert(grouped2.rightOuterJoin(grouped4).partitioner === grouped4.partitioner) + assert(grouped2.cogroup(grouped4).partitioner === grouped4.partitioner) assert(grouped2.join(reduced2).partitioner === grouped2.partitioner) assert(grouped2.leftOuterJoin(reduced2).partitioner === grouped2.partitioner) @@ -120,4 +120,21 @@ class PartitioningSuite extends FunSuite with LocalSparkContext { assert(intercept[SparkException]{ arrPairs.reduceByKeyLocally(_ + _) }.getMessage.contains("array")) assert(intercept[SparkException]{ arrPairs.reduceByKey(_ + _) }.getMessage.contains("array")) } + + test("Zero-length partitions should be correctly handled") { + // Create RDD with some consecutive empty partitions (including the "first" one) + sc = new SparkContext("local", "test") + val rdd: RDD[Double] = sc + .parallelize(Array(-1.0, -1.0, -1.0, -1.0, 2.0, 4.0, -1.0, -1.0), 8) + .filter(_ >= 0.0) + + // Run the partitions, including the consecutive empty ones, through StatCounter + val stats: StatCounter = rdd.stats(); + assert(abs(6.0 - stats.sum) < 0.01); + assert(abs(6.0/2 - rdd.mean) < 0.01); + assert(abs(1.0 - rdd.variance) < 0.01); + assert(abs(1.0 - rdd.stdev) < 0.01); + + // Add other tests here for classes that should be able to handle empty partitions correctly + } } diff --git a/core/src/test/scala/spark/PipedRDDSuite.scala b/core/src/test/scala/spark/PipedRDDSuite.scala index a6344edf8f..ed075f93ec 100644 --- a/core/src/test/scala/spark/PipedRDDSuite.scala +++ b/core/src/test/scala/spark/PipedRDDSuite.scala @@ -19,6 +19,45 @@ class PipedRDDSuite extends FunSuite with LocalSparkContext { assert(c(3) === "4") } + test("advanced pipe") { + sc = new SparkContext("local", "test") + val nums = sc.makeRDD(Array(1, 2, 3, 4), 2) + val bl = sc.broadcast(List("0")) + + val piped = nums.pipe(Seq("cat"), + Map[String, String](), + (f: String => Unit) => {bl.value.map(f(_));f("\u0001")}, + (i:Int, f: String=> Unit) => f(i + "_")) + + val c = piped.collect() + + assert(c.size === 8) + assert(c(0) === "0") + assert(c(1) === "\u0001") + assert(c(2) === "1_") + assert(c(3) === "2_") + assert(c(4) === "0") + assert(c(5) === "\u0001") + assert(c(6) === "3_") + assert(c(7) === "4_") + + val nums1 = sc.makeRDD(Array("a\t1", "b\t2", "a\t3", "b\t4"), 2) + val d = nums1.groupBy(str=>str.split("\t")(0)). + pipe(Seq("cat"), + Map[String, String](), + (f: String => Unit) => {bl.value.map(f(_));f("\u0001")}, + (i:Tuple2[String, Seq[String]], f: String=> Unit) => {for (e <- i._2){ f(e + "_")}}).collect() + assert(d.size === 8) + assert(d(0) === "0") + assert(d(1) === "\u0001") + assert(d(2) === "b\t2_") + assert(d(3) === "b\t4_") + assert(d(4) === "0") + assert(d(5) === "\u0001") + assert(d(6) === "a\t1_") + assert(d(7) === "a\t3_") + } + test("pipe with env variable") { sc = new SparkContext("local", "test") val nums = sc.makeRDD(Array(1, 2, 3, 4), 2) diff --git a/core/src/test/scala/spark/RDDSuite.scala b/core/src/test/scala/spark/RDDSuite.scala index 9739ba869b..67f3332d44 100644 --- a/core/src/test/scala/spark/RDDSuite.scala +++ b/core/src/test/scala/spark/RDDSuite.scala @@ -2,8 +2,10 @@ package spark import scala.collection.mutable.HashMap import org.scalatest.FunSuite +import org.scalatest.concurrent.Timeouts._ +import org.scalatest.time.{Span, Millis} import spark.SparkContext._ -import spark.rdd.{CoalescedRDD, PartitionPruningRDD} +import spark.rdd.{CoalescedRDD, CoGroupedRDD, EmptyRDD, PartitionPruningRDD, ShuffledRDD} class RDDSuite extends FunSuite with LocalSparkContext { @@ -100,6 +102,28 @@ class RDDSuite extends FunSuite with LocalSparkContext { assert(rdd.collect().toList === List(1, 2, 3, 4)) } + test("unpersist RDD") { + sc = new SparkContext("local", "test") + val rdd = sc.makeRDD(Array(1, 2, 3, 4), 2).cache() + rdd.count + assert(sc.persistentRdds.isEmpty === false) + rdd.unpersist() + assert(sc.persistentRdds.isEmpty === true) + + failAfter(Span(3000, Millis)) { + try { + while (! sc.getRDDStorageInfo.isEmpty) { + Thread.sleep(200) + } + } catch { + case _ => { Thread.sleep(10) } + // Do nothing. We might see exceptions because block manager + // is racing this thread to remove entries from the driver. + } + } + assert(sc.getRDDStorageInfo.isEmpty === true) + } + test("caching with failures") { sc = new SparkContext("local", "test") val onlySplit = new Partition { override def index: Int = 0 } @@ -123,6 +147,56 @@ class RDDSuite extends FunSuite with LocalSparkContext { assert(rdd.collect().toList === List(1, 2, 3, 4)) } + test("empty RDD") { + sc = new SparkContext("local", "test") + val empty = new EmptyRDD[Int](sc) + assert(empty.count === 0) + assert(empty.collect().size === 0) + + val thrown = intercept[UnsupportedOperationException]{ + empty.reduce(_+_) + } + assert(thrown.getMessage.contains("empty")) + + val emptyKv = new EmptyRDD[(Int, Int)](sc) + val rdd = sc.parallelize(1 to 2, 2).map(x => (x, x)) + assert(rdd.join(emptyKv).collect().size === 0) + assert(rdd.rightOuterJoin(emptyKv).collect().size === 0) + assert(rdd.leftOuterJoin(emptyKv).collect().size === 2) + assert(rdd.cogroup(emptyKv).collect().size === 2) + assert(rdd.union(emptyKv).collect().size === 2) + } + + test("cogrouped RDDs") { + sc = new SparkContext("local", "test") + val rdd1 = sc.makeRDD(Array((1, "one"), (1, "another one"), (2, "two"), (3, "three")), 2) + val rdd2 = sc.makeRDD(Array((1, "one1"), (1, "another one1"), (2, "two1")), 2) + + // Use cogroup function + val cogrouped = rdd1.cogroup(rdd2).collectAsMap() + assert(cogrouped(1) === (Seq("one", "another one"), Seq("one1", "another one1"))) + assert(cogrouped(2) === (Seq("two"), Seq("two1"))) + assert(cogrouped(3) === (Seq("three"), Seq())) + + // Construct CoGroupedRDD directly, with map side combine enabled + val cogrouped1 = new CoGroupedRDD[Int]( + Seq(rdd1.asInstanceOf[RDD[(Int, Any)]], rdd2.asInstanceOf[RDD[(Int, Any)]]), + new HashPartitioner(3), + true).collectAsMap() + assert(cogrouped1(1).toSeq === Seq(Seq("one", "another one"), Seq("one1", "another one1"))) + assert(cogrouped1(2).toSeq === Seq(Seq("two"), Seq("two1"))) + assert(cogrouped1(3).toSeq === Seq(Seq("three"), Seq())) + + // Construct CoGroupedRDD directly, with map side combine disabled + val cogrouped2 = new CoGroupedRDD[Int]( + Seq(rdd1.asInstanceOf[RDD[(Int, Any)]], rdd2.asInstanceOf[RDD[(Int, Any)]]), + new HashPartitioner(3), + false).collectAsMap() + assert(cogrouped2(1).toSeq === Seq(Seq("one", "another one"), Seq("one1", "another one1"))) + assert(cogrouped2(2).toSeq === Seq(Seq("two"), Seq("two1"))) + assert(cogrouped2(3).toSeq === Seq(Seq("three"), Seq())) + } + test("coalesced RDDs") { sc = new SparkContext("local", "test") val data = sc.parallelize(1 to 10, 10) @@ -154,6 +228,11 @@ class RDDSuite extends FunSuite with LocalSparkContext { assert(coalesced4.collect().toList === (1 to 10).toList) assert(coalesced4.glom().collect().map(_.toList).toList === (1 to 10).map(x => List(x)).toList) + + // we can optionally shuffle to keep the upstream parallel + val coalesced5 = data.coalesce(1, shuffle = true) + assert(coalesced5.dependencies.head.rdd.dependencies.head.rdd.asInstanceOf[ShuffledRDD[_, _]] != + null) } test("zipped RDDs") { @@ -178,4 +257,83 @@ class RDDSuite extends FunSuite with LocalSparkContext { assert(prunedData.size === 1) assert(prunedData(0) === 10) } + + test("mapWith") { + import java.util.Random + sc = new SparkContext("local", "test") + val ones = sc.makeRDD(Array(1, 1, 1, 1, 1, 1), 2) + val randoms = ones.mapWith( + (index: Int) => new Random(index + 42)) + {(t: Int, prng: Random) => prng.nextDouble * t}.collect() + val prn42_3 = { + val prng42 = new Random(42) + prng42.nextDouble(); prng42.nextDouble(); prng42.nextDouble() + } + val prn43_3 = { + val prng43 = new Random(43) + prng43.nextDouble(); prng43.nextDouble(); prng43.nextDouble() + } + assert(randoms(2) === prn42_3) + assert(randoms(5) === prn43_3) + } + + test("flatMapWith") { + import java.util.Random + sc = new SparkContext("local", "test") + val ones = sc.makeRDD(Array(1, 1, 1, 1, 1, 1), 2) + val randoms = ones.flatMapWith( + (index: Int) => new Random(index + 42)) + {(t: Int, prng: Random) => + val random = prng.nextDouble() + Seq(random * t, random * t * 10)}. + collect() + val prn42_3 = { + val prng42 = new Random(42) + prng42.nextDouble(); prng42.nextDouble(); prng42.nextDouble() + } + val prn43_3 = { + val prng43 = new Random(43) + prng43.nextDouble(); prng43.nextDouble(); prng43.nextDouble() + } + assert(randoms(5) === prn42_3 * 10) + assert(randoms(11) === prn43_3 * 10) + } + + test("filterWith") { + import java.util.Random + sc = new SparkContext("local", "test") + val ints = sc.makeRDD(Array(1, 2, 3, 4, 5, 6), 2) + val sample = ints.filterWith( + (index: Int) => new Random(index + 42)) + {(t: Int, prng: Random) => prng.nextInt(3) == 0}. + collect() + val checkSample = { + val prng42 = new Random(42) + val prng43 = new Random(43) + Array(1, 2, 3, 4, 5, 6).filter{i => + if (i < 4) 0 == prng42.nextInt(3) + else 0 == prng43.nextInt(3)} + } + assert(sample.size === checkSample.size) + for (i <- 0 until sample.size) assert(sample(i) === checkSample(i)) + } + + test("top with predefined ordering") { + sc = new SparkContext("local", "test") + val nums = Array.range(1, 100000) + val ints = sc.makeRDD(scala.util.Random.shuffle(nums), 2) + val topK = ints.top(5) + assert(topK.size === 5) + assert(topK.sorted === nums.sorted.takeRight(5)) + } + + test("top with custom ordering") { + sc = new SparkContext("local", "test") + val words = Vector("a", "b", "c", "d") + implicit val ord = implicitly[Ordering[String]].reverse + val rdd = sc.makeRDD(words, 2) + val topK = rdd.top(2) + assert(topK.size === 2) + assert(topK.sorted === Array("b", "a")) + } } diff --git a/core/src/test/scala/spark/ShuffleNettySuite.scala b/core/src/test/scala/spark/ShuffleNettySuite.scala new file mode 100644 index 0000000000..bfaffa953e --- /dev/null +++ b/core/src/test/scala/spark/ShuffleNettySuite.scala @@ -0,0 +1,17 @@ +package spark + +import org.scalatest.BeforeAndAfterAll + + +class ShuffleNettySuite extends ShuffleSuite with BeforeAndAfterAll { + + // This test suite should run all tests in ShuffleSuite with Netty shuffle mode. + + override def beforeAll(configMap: Map[String, Any]) { + System.setProperty("spark.shuffle.use.netty", "true") + } + + override def afterAll(configMap: Map[String, Any]) { + System.setProperty("spark.shuffle.use.netty", "false") + } +} diff --git a/core/src/test/scala/spark/ShuffleSuite.scala b/core/src/test/scala/spark/ShuffleSuite.scala index 92c3f67416..0c1ec29f96 100644 --- a/core/src/test/scala/spark/ShuffleSuite.scala +++ b/core/src/test/scala/spark/ShuffleSuite.scala @@ -99,7 +99,7 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with LocalSparkContext { val sums = pairs.reduceByKey(_+_, 10).collect() assert(sums.toSet === Set((1, 7), (2, 1))) } - + test("reduceByKey with partitioner") { sc = new SparkContext("local", "test") val p = new Partitioner() { @@ -234,11 +234,194 @@ class ShuffleSuite extends FunSuite with ShouldMatchers with LocalSparkContext { assert(rdd.keys.collect().toList === List(1, 2)) assert(rdd.values.collect().toList === List("a", "b")) } + + test("default partitioner uses partition size") { + sc = new SparkContext("local", "test") + // specify 2000 partitions + val a = sc.makeRDD(Array(1, 2, 3, 4), 2000) + // do a map, which loses the partitioner + val b = a.map(a => (a, (a * 2).toString)) + // then a group by, and see we didn't revert to 2 partitions + val c = b.groupByKey() + assert(c.partitions.size === 2000) + } + + test("default partitioner uses largest partitioner") { + sc = new SparkContext("local", "test") + val a = sc.makeRDD(Array((1, "a"), (2, "b")), 2) + val b = sc.makeRDD(Array((1, "a"), (2, "b")), 2000) + val c = a.join(b) + assert(c.partitions.size === 2000) + } + + test("subtract") { + sc = new SparkContext("local", "test") + val a = sc.parallelize(Array(1, 2, 3), 2) + val b = sc.parallelize(Array(2, 3, 4), 4) + val c = a.subtract(b) + assert(c.collect().toSet === Set(1)) + assert(c.partitions.size === a.partitions.size) + } + + test("subtract with narrow dependency") { + sc = new SparkContext("local", "test") + // use a deterministic partitioner + val p = new Partitioner() { + def numPartitions = 5 + def getPartition(key: Any) = key.asInstanceOf[Int] + } + // partitionBy so we have a narrow dependency + val a = sc.parallelize(Array((1, "a"), (2, "b"), (3, "c"))).partitionBy(p) + // more partitions/no partitioner so a shuffle dependency + val b = sc.parallelize(Array((2, "b"), (3, "cc"), (4, "d")), 4) + val c = a.subtract(b) + assert(c.collect().toSet === Set((1, "a"), (3, "c"))) + // Ideally we could keep the original partitioner... + assert(c.partitioner === None) + } + + test("subtractByKey") { + sc = new SparkContext("local", "test") + val a = sc.parallelize(Array((1, "a"), (1, "a"), (2, "b"), (3, "c")), 2) + val b = sc.parallelize(Array((2, 20), (3, 30), (4, 40)), 4) + val c = a.subtractByKey(b) + assert(c.collect().toSet === Set((1, "a"), (1, "a"))) + assert(c.partitions.size === a.partitions.size) + } + + test("subtractByKey with narrow dependency") { + sc = new SparkContext("local", "test") + // use a deterministic partitioner + val p = new Partitioner() { + def numPartitions = 5 + def getPartition(key: Any) = key.asInstanceOf[Int] + } + // partitionBy so we have a narrow dependency + val a = sc.parallelize(Array((1, "a"), (1, "a"), (2, "b"), (3, "c"))).partitionBy(p) + // more partitions/no partitioner so a shuffle dependency + val b = sc.parallelize(Array((2, "b"), (3, "cc"), (4, "d")), 4) + val c = a.subtractByKey(b) + assert(c.collect().toSet === Set((1, "a"), (1, "a"))) + assert(c.partitioner.get === p) + } + + test("shuffle non-zero block size") { + sc = new SparkContext("local-cluster[2,1,512]", "test") + val NUM_BLOCKS = 3 + + val a = sc.parallelize(1 to 10, 2) + val b = a.map { x => + (x, new ShuffleSuite.NonJavaSerializableClass(x * 2)) + } + // If the Kryo serializer is not used correctly, the shuffle would fail because the + // default Java serializer cannot handle the non serializable class. + val c = new ShuffledRDD(b, new HashPartitioner(NUM_BLOCKS), + classOf[spark.KryoSerializer].getName) + val shuffleId = c.dependencies.head.asInstanceOf[ShuffleDependency[Int, Int]].shuffleId + + assert(c.count === 10) + + // All blocks must have non-zero size + (0 until NUM_BLOCKS).foreach { id => + val statuses = SparkEnv.get.mapOutputTracker.getServerStatuses(shuffleId, id) + assert(statuses.forall(s => s._2 > 0)) + } + } + + test("shuffle serializer") { + // Use a local cluster with 2 processes to make sure there are both local and remote blocks + sc = new SparkContext("local-cluster[2,1,512]", "test") + val a = sc.parallelize(1 to 10, 2) + val b = a.map { x => + (x, new ShuffleSuite.NonJavaSerializableClass(x * 2)) + } + // If the Kryo serializer is not used correctly, the shuffle would fail because the + // default Java serializer cannot handle the non serializable class. + val c = new ShuffledRDD(b, new HashPartitioner(3), classOf[spark.KryoSerializer].getName) + assert(c.count === 10) + } + + test("zero sized blocks") { + // Use a local cluster with 2 processes to make sure there are both local and remote blocks + sc = new SparkContext("local-cluster[2,1,512]", "test") + + // 10 partitions from 4 keys + val NUM_BLOCKS = 10 + val a = sc.parallelize(1 to 4, NUM_BLOCKS) + val b = a.map(x => (x, x*2)) + + // NOTE: The default Java serializer doesn't create zero-sized blocks. + // So, use Kryo + val c = new ShuffledRDD(b, new HashPartitioner(10), classOf[spark.KryoSerializer].getName) + + val shuffleId = c.dependencies.head.asInstanceOf[ShuffleDependency[Int, Int]].shuffleId + assert(c.count === 4) + + val blockSizes = (0 until NUM_BLOCKS).flatMap { id => + val statuses = SparkEnv.get.mapOutputTracker.getServerStatuses(shuffleId, id) + statuses.map(x => x._2) + } + val nonEmptyBlocks = blockSizes.filter(x => x > 0) + + // We should have at most 4 non-zero sized partitions + assert(nonEmptyBlocks.size <= 4) + } + + test("zero sized blocks without kryo") { + // Use a local cluster with 2 processes to make sure there are both local and remote blocks + sc = new SparkContext("local-cluster[2,1,512]", "test") + + // 10 partitions from 4 keys + val NUM_BLOCKS = 10 + val a = sc.parallelize(1 to 4, NUM_BLOCKS) + val b = a.map(x => (x, x*2)) + + // NOTE: The default Java serializer should create zero-sized blocks + val c = new ShuffledRDD(b, new HashPartitioner(10)) + + val shuffleId = c.dependencies.head.asInstanceOf[ShuffleDependency[Int, Int]].shuffleId + assert(c.count === 4) + + val blockSizes = (0 until NUM_BLOCKS).flatMap { id => + val statuses = SparkEnv.get.mapOutputTracker.getServerStatuses(shuffleId, id) + statuses.map(x => x._2) + } + val nonEmptyBlocks = blockSizes.filter(x => x > 0) + + // We should have at most 4 non-zero sized partitions + assert(nonEmptyBlocks.size <= 4) + } + + test("foldByKey") { + sc = new SparkContext("local", "test") + val pairs = sc.parallelize(Array((1, 1), (1, 2), (1, 3), (1, 1), (2, 1))) + val sums = pairs.foldByKey(0)(_+_).collect() + assert(sums.toSet === Set((1, 7), (2, 1))) + } + + test("foldByKey with mutable result type") { + sc = new SparkContext("local", "test") + val pairs = sc.parallelize(Array((1, 1), (1, 2), (1, 3), (1, 1), (2, 1))) + val bufs = pairs.mapValues(v => ArrayBuffer(v)).cache() + // Fold the values using in-place mutation + val sums = bufs.foldByKey(new ArrayBuffer[Int])(_ ++= _).collect() + assert(sums.toSet === Set((1, ArrayBuffer(1, 2, 3, 1)), (2, ArrayBuffer(1)))) + // Check that the mutable objects in the original RDD were not changed + assert(bufs.collect().toSet === Set( + (1, ArrayBuffer(1)), + (1, ArrayBuffer(2)), + (1, ArrayBuffer(3)), + (1, ArrayBuffer(1)), + (2, ArrayBuffer(1)))) + } } object ShuffleSuite { + def mergeCombineException(x: Int, y: Int): Int = { throw new SparkException("Exception for map-side combine.") x + y } + + class NonJavaSerializableClass(val value: Int) } diff --git a/core/src/test/scala/spark/ZippedPartitionsSuite.scala b/core/src/test/scala/spark/ZippedPartitionsSuite.scala new file mode 100644 index 0000000000..5f60aa75d7 --- /dev/null +++ b/core/src/test/scala/spark/ZippedPartitionsSuite.scala @@ -0,0 +1,34 @@ +package spark + +import scala.collection.immutable.NumericRange + +import org.scalatest.FunSuite +import org.scalatest.prop.Checkers +import org.scalacheck.Arbitrary._ +import org.scalacheck.Gen +import org.scalacheck.Prop._ + +import SparkContext._ + + +object ZippedPartitionsSuite { + def procZippedData(i: Iterator[Int], s: Iterator[String], d: Iterator[Double]) : Iterator[Int] = { + Iterator(i.toArray.size, s.toArray.size, d.toArray.size) + } +} + +class ZippedPartitionsSuite extends FunSuite with LocalSparkContext { + test("print sizes") { + sc = new SparkContext("local", "test") + val data1 = sc.makeRDD(Array(1, 2, 3, 4), 2) + val data2 = sc.makeRDD(Array("1", "2", "3", "4", "5", "6"), 2) + val data3 = sc.makeRDD(Array(1.0, 2.0), 2) + + val zippedRDD = data1.zipPartitions(ZippedPartitionsSuite.procZippedData, data2, data3) + + val obtainedSizes = zippedRDD.collect() + val expectedSizes = Array(2, 3, 1, 2, 3, 1) + assert(obtainedSizes.size == 6) + assert(obtainedSizes.zip(expectedSizes).forall(x => x._1 == x._2)) + } +} diff --git a/core/src/test/scala/spark/rdd/JdbcRDDSuite.scala b/core/src/test/scala/spark/rdd/JdbcRDDSuite.scala new file mode 100644 index 0000000000..6afb0fa9bc --- /dev/null +++ b/core/src/test/scala/spark/rdd/JdbcRDDSuite.scala @@ -0,0 +1,56 @@ +package spark + +import org.scalatest.{ BeforeAndAfter, FunSuite } +import spark.SparkContext._ +import spark.rdd.JdbcRDD +import java.sql._ + +class JdbcRDDSuite extends FunSuite with BeforeAndAfter with LocalSparkContext { + + before { + Class.forName("org.apache.derby.jdbc.EmbeddedDriver") + val conn = DriverManager.getConnection("jdbc:derby:target/JdbcRDDSuiteDb;create=true") + try { + val create = conn.createStatement + create.execute(""" + CREATE TABLE FOO( + ID INTEGER NOT NULL GENERATED ALWAYS AS IDENTITY (START WITH 1, INCREMENT BY 1), + DATA INTEGER + )""") + create.close + val insert = conn.prepareStatement("INSERT INTO FOO(DATA) VALUES(?)") + (1 to 100).foreach { i => + insert.setInt(1, i * 2) + insert.executeUpdate + } + insert.close + } catch { + case e: SQLException if e.getSQLState == "X0Y32" => + // table exists + } finally { + conn.close + } + } + + test("basic functionality") { + sc = new SparkContext("local", "test") + val rdd = new JdbcRDD( + sc, + () => { DriverManager.getConnection("jdbc:derby:target/JdbcRDDSuiteDb") }, + "SELECT DATA FROM FOO WHERE ? <= ID AND ID <= ?", + 1, 100, 3, + (r: ResultSet) => { r.getInt(1) } ).cache + + assert(rdd.count === 100) + assert(rdd.reduce(_+_) === 10100) + } + + after { + try { + DriverManager.getConnection("jdbc:derby:;shutdown=true") + } catch { + case se: SQLException if se.getSQLState == "XJ015" => + // normal shutdown + } + } +} diff --git a/core/src/test/scala/spark/scheduler/ClusterSchedulerSuite.scala b/core/src/test/scala/spark/scheduler/ClusterSchedulerSuite.scala new file mode 100644 index 0000000000..8e1ad27e14 --- /dev/null +++ b/core/src/test/scala/spark/scheduler/ClusterSchedulerSuite.scala @@ -0,0 +1,250 @@ +package spark.scheduler + +import org.scalatest.FunSuite +import org.scalatest.BeforeAndAfter + +import spark._ +import spark.scheduler._ +import spark.scheduler.cluster._ +import scala.collection.mutable.ArrayBuffer + +import java.util.Properties + +class DummyTaskSetManager( + initPriority: Int, + initStageId: Int, + initNumTasks: Int, + clusterScheduler: ClusterScheduler, + taskSet: TaskSet) + extends ClusterTaskSetManager(clusterScheduler,taskSet) { + + parent = null + weight = 1 + minShare = 2 + runningTasks = 0 + priority = initPriority + stageId = initStageId + name = "TaskSet_"+stageId + override val numTasks = initNumTasks + tasksFinished = 0 + + override def increaseRunningTasks(taskNum: Int) { + runningTasks += taskNum + if (parent != null) { + parent.increaseRunningTasks(taskNum) + } + } + + override def decreaseRunningTasks(taskNum: Int) { + runningTasks -= taskNum + if (parent != null) { + parent.decreaseRunningTasks(taskNum) + } + } + + override def addSchedulable(schedulable: Schedulable) { + } + + override def removeSchedulable(schedulable: Schedulable) { + } + + override def getSchedulableByName(name: String): Schedulable = { + return null + } + + override def executorLost(executorId: String, host: String): Unit = { + } + + override def slaveOffer(execId: String, host: String, avaiableCpus: Double, overrideLocality: TaskLocality.TaskLocality = null): Option[TaskDescription] = { + if (tasksFinished + runningTasks < numTasks) { + increaseRunningTasks(1) + return Some(new TaskDescription(0, execId, "task 0:0", null)) + } + return None + } + + override def checkSpeculatableTasks(): Boolean = { + return true + } + + def taskFinished() { + decreaseRunningTasks(1) + tasksFinished +=1 + if (tasksFinished == numTasks) { + parent.removeSchedulable(this) + } + } + + def abort() { + decreaseRunningTasks(runningTasks) + parent.removeSchedulable(this) + } +} + +class DummyTask(stageId: Int) extends Task[Int](stageId) +{ + def run(attemptId: Long): Int = { + return 0 + } +} + +class ClusterSchedulerSuite extends FunSuite with LocalSparkContext with Logging { + + def createDummyTaskSetManager(priority: Int, stage: Int, numTasks: Int, cs: ClusterScheduler, taskSet: TaskSet): DummyTaskSetManager = { + new DummyTaskSetManager(priority, stage, numTasks, cs , taskSet) + } + + def resourceOffer(rootPool: Pool): Int = { + val taskSetQueue = rootPool.getSortedTaskSetQueue() + /* Just for Test*/ + for (manager <- taskSetQueue) { + logInfo("parentName:%s, parent running tasks:%d, name:%s,runningTasks:%d".format(manager.parent.name, manager.parent.runningTasks, manager.name, manager.runningTasks)) + } + for (taskSet <- taskSetQueue) { + taskSet.slaveOffer("execId_1", "hostname_1", 1) match { + case Some(task) => + return taskSet.stageId + case None => {} + } + } + -1 + } + + def checkTaskSetId(rootPool: Pool, expectedTaskSetId: Int) { + assert(resourceOffer(rootPool) === expectedTaskSetId) + } + + test("FIFO Scheduler Test") { + sc = new SparkContext("local", "ClusterSchedulerSuite") + val clusterScheduler = new ClusterScheduler(sc) + var tasks = ArrayBuffer[Task[_]]() + val task = new DummyTask(0) + tasks += task + val taskSet = new TaskSet(tasks.toArray,0,0,0,null) + + val rootPool = new Pool("", SchedulingMode.FIFO, 0, 0) + val schedulableBuilder = new FIFOSchedulableBuilder(rootPool) + schedulableBuilder.buildPools() + + val taskSetManager0 = createDummyTaskSetManager(0, 0, 2, clusterScheduler, taskSet) + val taskSetManager1 = createDummyTaskSetManager(0, 1, 2, clusterScheduler, taskSet) + val taskSetManager2 = createDummyTaskSetManager(0, 2, 2, clusterScheduler, taskSet) + schedulableBuilder.addTaskSetManager(taskSetManager0, null) + schedulableBuilder.addTaskSetManager(taskSetManager1, null) + schedulableBuilder.addTaskSetManager(taskSetManager2, null) + + checkTaskSetId(rootPool, 0) + resourceOffer(rootPool) + checkTaskSetId(rootPool, 1) + resourceOffer(rootPool) + taskSetManager1.abort() + checkTaskSetId(rootPool, 2) + } + + test("Fair Scheduler Test") { + sc = new SparkContext("local", "ClusterSchedulerSuite") + val clusterScheduler = new ClusterScheduler(sc) + var tasks = ArrayBuffer[Task[_]]() + val task = new DummyTask(0) + tasks += task + val taskSet = new TaskSet(tasks.toArray,0,0,0,null) + + val xmlPath = getClass.getClassLoader.getResource("fairscheduler.xml").getFile() + System.setProperty("spark.fairscheduler.allocation.file", xmlPath) + val rootPool = new Pool("", SchedulingMode.FAIR, 0, 0) + val schedulableBuilder = new FairSchedulableBuilder(rootPool) + schedulableBuilder.buildPools() + + assert(rootPool.getSchedulableByName("default") != null) + assert(rootPool.getSchedulableByName("1") != null) + assert(rootPool.getSchedulableByName("2") != null) + assert(rootPool.getSchedulableByName("3") != null) + assert(rootPool.getSchedulableByName("1").minShare === 2) + assert(rootPool.getSchedulableByName("1").weight === 1) + assert(rootPool.getSchedulableByName("2").minShare === 3) + assert(rootPool.getSchedulableByName("2").weight === 1) + assert(rootPool.getSchedulableByName("3").minShare === 2) + assert(rootPool.getSchedulableByName("3").weight === 1) + + val properties1 = new Properties() + properties1.setProperty("spark.scheduler.cluster.fair.pool","1") + val properties2 = new Properties() + properties2.setProperty("spark.scheduler.cluster.fair.pool","2") + + val taskSetManager10 = createDummyTaskSetManager(1, 0, 1, clusterScheduler, taskSet) + val taskSetManager11 = createDummyTaskSetManager(1, 1, 1, clusterScheduler, taskSet) + val taskSetManager12 = createDummyTaskSetManager(1, 2, 2, clusterScheduler, taskSet) + schedulableBuilder.addTaskSetManager(taskSetManager10, properties1) + schedulableBuilder.addTaskSetManager(taskSetManager11, properties1) + schedulableBuilder.addTaskSetManager(taskSetManager12, properties1) + + val taskSetManager23 = createDummyTaskSetManager(2, 3, 2, clusterScheduler, taskSet) + val taskSetManager24 = createDummyTaskSetManager(2, 4, 2, clusterScheduler, taskSet) + schedulableBuilder.addTaskSetManager(taskSetManager23, properties2) + schedulableBuilder.addTaskSetManager(taskSetManager24, properties2) + + checkTaskSetId(rootPool, 0) + checkTaskSetId(rootPool, 3) + checkTaskSetId(rootPool, 3) + checkTaskSetId(rootPool, 1) + checkTaskSetId(rootPool, 4) + checkTaskSetId(rootPool, 2) + checkTaskSetId(rootPool, 2) + checkTaskSetId(rootPool, 4) + + taskSetManager12.taskFinished() + assert(rootPool.getSchedulableByName("1").runningTasks === 3) + taskSetManager24.abort() + assert(rootPool.getSchedulableByName("2").runningTasks === 2) + } + + test("Nested Pool Test") { + sc = new SparkContext("local", "ClusterSchedulerSuite") + val clusterScheduler = new ClusterScheduler(sc) + var tasks = ArrayBuffer[Task[_]]() + val task = new DummyTask(0) + tasks += task + val taskSet = new TaskSet(tasks.toArray,0,0,0,null) + + val rootPool = new Pool("", SchedulingMode.FAIR, 0, 0) + val pool0 = new Pool("0", SchedulingMode.FAIR, 3, 1) + val pool1 = new Pool("1", SchedulingMode.FAIR, 4, 1) + rootPool.addSchedulable(pool0) + rootPool.addSchedulable(pool1) + + val pool00 = new Pool("00", SchedulingMode.FAIR, 2, 2) + val pool01 = new Pool("01", SchedulingMode.FAIR, 1, 1) + pool0.addSchedulable(pool00) + pool0.addSchedulable(pool01) + + val pool10 = new Pool("10", SchedulingMode.FAIR, 2, 2) + val pool11 = new Pool("11", SchedulingMode.FAIR, 2, 1) + pool1.addSchedulable(pool10) + pool1.addSchedulable(pool11) + + val taskSetManager000 = createDummyTaskSetManager(0, 0, 5, clusterScheduler, taskSet) + val taskSetManager001 = createDummyTaskSetManager(0, 1, 5, clusterScheduler, taskSet) + pool00.addSchedulable(taskSetManager000) + pool00.addSchedulable(taskSetManager001) + + val taskSetManager010 = createDummyTaskSetManager(1, 2, 5, clusterScheduler, taskSet) + val taskSetManager011 = createDummyTaskSetManager(1, 3, 5, clusterScheduler, taskSet) + pool01.addSchedulable(taskSetManager010) + pool01.addSchedulable(taskSetManager011) + + val taskSetManager100 = createDummyTaskSetManager(2, 4, 5, clusterScheduler, taskSet) + val taskSetManager101 = createDummyTaskSetManager(2, 5, 5, clusterScheduler, taskSet) + pool10.addSchedulable(taskSetManager100) + pool10.addSchedulable(taskSetManager101) + + val taskSetManager110 = createDummyTaskSetManager(3, 6, 5, clusterScheduler, taskSet) + val taskSetManager111 = createDummyTaskSetManager(3, 7, 5, clusterScheduler, taskSet) + pool11.addSchedulable(taskSetManager110) + pool11.addSchedulable(taskSetManager111) + + checkTaskSetId(rootPool, 0) + checkTaskSetId(rootPool, 4) + checkTaskSetId(rootPool, 6) + checkTaskSetId(rootPool, 2) + } +} diff --git a/core/src/test/scala/spark/scheduler/DAGSchedulerSuite.scala b/core/src/test/scala/spark/scheduler/DAGSchedulerSuite.scala index 8de490eb86..30e6fef950 100644 --- a/core/src/test/scala/spark/scheduler/DAGSchedulerSuite.scala +++ b/core/src/test/scala/spark/scheduler/DAGSchedulerSuite.scala @@ -4,16 +4,8 @@ import scala.collection.mutable.{Map, HashMap} import org.scalatest.FunSuite import org.scalatest.BeforeAndAfter -import org.scalatest.concurrent.TimeLimitedTests -import org.scalatest.mock.EasyMockSugar -import org.scalatest.time.{Span, Seconds} -import org.easymock.EasyMock._ -import org.easymock.Capture -import org.easymock.EasyMock -import org.easymock.{IAnswer, IArgumentMatcher} - -import akka.actor.ActorSystem +import spark.LocalSparkContext import spark.storage.BlockManager import spark.storage.BlockManagerId @@ -42,27 +34,24 @@ import spark.{FetchFailed, Success} * DAGScheduler#submitWaitingStages (via test utility functions like runEvent or respondToTaskSet) * and capturing the resulting TaskSets from the mock TaskScheduler. */ -class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar with TimeLimitedTests { - - // impose a time limit on this test in case we don't let the job finish, in which case - // JobWaiter#getResult will hang. - override val timeLimit = Span(5, Seconds) +class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkContext { + + /** Set of TaskSets the DAGScheduler has requested executed. */ + val taskSets = scala.collection.mutable.Buffer[TaskSet]() + val taskScheduler = new TaskScheduler() { + override def start() = {} + override def stop() = {} + override def submitTasks(taskSet: TaskSet) = { + // normally done by TaskSetManager + taskSet.tasks.foreach(_.generation = mapOutputTracker.getGeneration) + taskSets += taskSet + } + override def setListener(listener: TaskSchedulerListener) = {} + override def defaultParallelism() = 2 + } - val sc: SparkContext = new SparkContext("local", "DAGSchedulerSuite") - var scheduler: DAGScheduler = null - val taskScheduler = mock[TaskScheduler] - val blockManagerMaster = mock[BlockManagerMaster] var mapOutputTracker: MapOutputTracker = null - var schedulerThread: Thread = null - var schedulerException: Throwable = null - - /** - * Set of EasyMock argument matchers that match a TaskSet for a given RDD. - * We cache these so we do not create duplicate matchers for the same RDD. - * This allows us to easily setup a sequence of expectations for task sets for - * that RDD. - */ - val taskSetMatchers = new HashMap[MyRDD, IArgumentMatcher] + var scheduler: DAGScheduler = null /** * Set of cache locations to return from our mock BlockManagerMaster. @@ -70,68 +59,50 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar * list of cache locations silently. */ val cacheLocations = new HashMap[(Int, Int), Seq[BlockManagerId]] - - /** - * JobWaiter for the last JobSubmitted event we pushed. To keep tests (most of which - * will only submit one job) from needing to explicitly track it. - */ - var lastJobWaiter: JobWaiter[Int] = null - - /** - * Array into which we are accumulating the results from the last job asynchronously. - */ - var lastJobResult: Array[Int] = null - - /** - * Tell EasyMockSugar what mock objects we want to be configured by expecting {...} - * and whenExecuting {...} */ - implicit val mocks = MockObjects(taskScheduler, blockManagerMaster) - - /** - * Utility function to reset mocks and set expectations on them. EasyMock wants mock objects - * to be reset after each time their expectations are set, and we tend to check mock object - * calls over a single call to DAGScheduler. - * - * We also set a default expectation here that blockManagerMaster.getLocations can be called - * and will return values from cacheLocations. - */ - def resetExpecting(f: => Unit) { - reset(taskScheduler) - reset(blockManagerMaster) - expecting { - expectGetLocations() - f + // stub out BlockManagerMaster.getLocations to use our cacheLocations + val blockManagerMaster = new BlockManagerMaster(null) { + override def getLocations(blockIds: Array[String]): Seq[Seq[BlockManagerId]] = { + blockIds.map { name => + val pieces = name.split("_") + if (pieces(0) == "rdd") { + val key = pieces(1).toInt -> pieces(2).toInt + cacheLocations.getOrElse(key, Seq()) + } else { + Seq() + } + }.toSeq + } + override def removeExecutor(execId: String) { + // don't need to propagate to the driver, which we don't have + } } + + /** The list of results that DAGScheduler has collected. */ + val results = new HashMap[Int, Any]() + var failure: Exception = _ + val listener = new JobListener() { + override def taskSucceeded(index: Int, result: Any) = results.put(index, result) + override def jobFailed(exception: Exception) = { failure = exception } } before { - taskSetMatchers.clear() + sc = new SparkContext("local", "DAGSchedulerSuite") + taskSets.clear() cacheLocations.clear() - val actorSystem = ActorSystem("test") - mapOutputTracker = new MapOutputTracker(actorSystem, true) - resetExpecting { - taskScheduler.setListener(anyObject()) - } - whenExecuting { - scheduler = new DAGScheduler(taskScheduler, mapOutputTracker, blockManagerMaster, null) + results.clear() + mapOutputTracker = new MapOutputTracker() + scheduler = new DAGScheduler(taskScheduler, mapOutputTracker, blockManagerMaster, null) { + override def runLocally(job: ActiveJob) { + // don't bother with the thread while unit testing + runLocallyWithinThread(job) + } } } after { - assert(scheduler.processEvent(StopDAGScheduler)) - resetExpecting { - taskScheduler.stop() - } - whenExecuting { - scheduler.stop() - } - sc.stop() - System.clearProperty("spark.master.port") + scheduler.stop() } - def makeBlockManagerId(host: String): BlockManagerId = - BlockManagerId("exec-" + host, host, 12345) - /** * Type of RDD we use for testing. Note that we should never call the real RDD compute methods. * This is a pair RDD type so it can always be used in ShuffleDependencies. @@ -143,7 +114,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar * preferredLocations (if any) that are passed to them. They are deliberately not executable * so we can test that DAGScheduler does not try to execute RDDs locally. */ - def makeRdd( + private def makeRdd( numPartitions: Int, dependencies: List[Dependency[_]], locations: Seq[Seq[String]] = Nil @@ -165,192 +136,66 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar } /** - * EasyMock matcher method. For use as an argument matcher for a TaskSet whose first task - * is from a particular RDD. - */ - def taskSetForRdd(rdd: MyRDD): TaskSet = { - val matcher = taskSetMatchers.getOrElseUpdate(rdd, - new IArgumentMatcher { - override def matches(actual: Any): Boolean = { - val taskSet = actual.asInstanceOf[TaskSet] - taskSet.tasks(0) match { - case rt: ResultTask[_, _] => rt.rdd.id == rdd.id - case smt: ShuffleMapTask => smt.rdd.id == rdd.id - case _ => false - } - } - override def appendTo(buf: StringBuffer) { - buf.append("taskSetForRdd(" + rdd + ")") - } - }) - EasyMock.reportMatcher(matcher) - return null - } - - /** - * Setup an EasyMock expectation to repsond to blockManagerMaster.getLocations() called from - * cacheLocations. - */ - def expectGetLocations(): Unit = { - EasyMock.expect(blockManagerMaster.getLocations(anyObject().asInstanceOf[Array[String]])). - andAnswer(new IAnswer[Seq[Seq[BlockManagerId]]] { - override def answer(): Seq[Seq[BlockManagerId]] = { - val blocks = getCurrentArguments()(0).asInstanceOf[Array[String]] - return blocks.map { name => - val pieces = name.split("_") - if (pieces(0) == "rdd") { - val key = pieces(1).toInt -> pieces(2).toInt - if (cacheLocations.contains(key)) { - cacheLocations(key) - } else { - Seq[BlockManagerId]() - } - } else { - Seq[BlockManagerId]() - } - }.toSeq - } - }).anyTimes() - } - - /** * Process the supplied event as if it were the top of the DAGScheduler event queue, expecting * the scheduler not to exit. * * After processing the event, submit waiting stages as is done on most iterations of the * DAGScheduler event loop. */ - def runEvent(event: DAGSchedulerEvent) { + private def runEvent(event: DAGSchedulerEvent) { assert(!scheduler.processEvent(event)) scheduler.submitWaitingStages() } /** - * Expect a TaskSet for the specified RDD to be submitted to the TaskScheduler. Should be - * called from a resetExpecting { ... } block. - * - * Returns a easymock Capture that will contain the task set after the stage is submitted. - * Most tests should use interceptStage() instead of this directly. - */ - def expectStage(rdd: MyRDD): Capture[TaskSet] = { - val taskSetCapture = new Capture[TaskSet] - taskScheduler.submitTasks(and(capture(taskSetCapture), taskSetForRdd(rdd))) - return taskSetCapture - } - - /** - * Expect the supplied code snippet to submit a stage for the specified RDD. - * Return the resulting TaskSet. First marks all the tasks are belonging to the - * current MapOutputTracker generation. + * When we submit dummy Jobs, this is the compute function we supply. Except in a local test + * below, we do not expect this function to ever be executed; instead, we will return results + * directly through CompletionEvents. */ - def interceptStage(rdd: MyRDD)(f: => Unit): TaskSet = { - var capture: Capture[TaskSet] = null - resetExpecting { - capture = expectStage(rdd) - } - whenExecuting { - f - } - val taskSet = capture.getValue - for (task <- taskSet.tasks) { - task.generation = mapOutputTracker.getGeneration - } - return taskSet - } + private val jobComputeFunc = (context: TaskContext, it: Iterator[(_)]) => + it.next.asInstanceOf[Tuple2[_, _]]._1 - /** - * Send the given CompletionEvent messages for the tasks in the TaskSet. - */ - def respondToTaskSet(taskSet: TaskSet, results: Seq[(TaskEndReason, Any)]) { + /** Send the given CompletionEvent messages for the tasks in the TaskSet. */ + private def complete(taskSet: TaskSet, results: Seq[(TaskEndReason, Any)]) { assert(taskSet.tasks.size >= results.size) for ((result, i) <- results.zipWithIndex) { if (i < taskSet.tasks.size) { - runEvent(CompletionEvent(taskSet.tasks(i), result._1, result._2, Map[Long, Any]())) + runEvent(CompletionEvent(taskSet.tasks(i), result._1, result._2, Map[Long, Any](), null, null)) } } } - /** - * Assert that the supplied TaskSet has exactly the given preferredLocations. - */ - def expectTaskSetLocations(taskSet: TaskSet, locations: Seq[Seq[String]]) { - assert(locations.size === taskSet.tasks.size) - for ((expectLocs, taskLocs) <- - taskSet.tasks.map(_.preferredLocations).zip(locations)) { - assert(expectLocs === taskLocs) - } - } - - /** - * When we submit dummy Jobs, this is the compute function we supply. Except in a local test - * below, we do not expect this function to ever be executed; instead, we will return results - * directly through CompletionEvents. - */ - def jobComputeFunc(context: TaskContext, it: Iterator[(Int, Int)]): Int = - it.next._1.asInstanceOf[Int] - - - /** - * Start a job to compute the given RDD. Returns the JobWaiter that will - * collect the result of the job via callbacks from DAGScheduler. - */ - def submitRdd(rdd: MyRDD, allowLocal: Boolean = false): (JobWaiter[Int], Array[Int]) = { - val resultArray = new Array[Int](rdd.partitions.size) - val (toSubmit, waiter) = scheduler.prepareJob[(Int, Int), Int]( - rdd, - jobComputeFunc, - (0 to (rdd.partitions.size - 1)), - "test-site", - allowLocal, - (i: Int, value: Int) => resultArray(i) = value - ) - lastJobWaiter = waiter - lastJobResult = resultArray - runEvent(toSubmit) - return (waiter, resultArray) - } - - /** - * Assert that a job we started has failed. - */ - def expectJobException(waiter: JobWaiter[Int] = lastJobWaiter) { - waiter.awaitResult() match { - case JobSucceeded => fail() - case JobFailed(_) => return - } + /** Sends the rdd to the scheduler for scheduling. */ + private def submit( + rdd: RDD[_], + partitions: Array[Int], + func: (TaskContext, Iterator[_]) => _ = jobComputeFunc, + allowLocal: Boolean = false, + listener: JobListener = listener) { + runEvent(JobSubmitted(rdd, func, partitions, allowLocal, null, listener)) } - /** - * Assert that a job we started has succeeded and has the given result. - */ - def expectJobResult(expected: Array[Int], waiter: JobWaiter[Int] = lastJobWaiter, - result: Array[Int] = lastJobResult) { - waiter.awaitResult match { - case JobSucceeded => - assert(expected === result) - case JobFailed(_) => - fail() - } + /** Sends TaskSetFailed to the scheduler. */ + private def failed(taskSet: TaskSet, message: String) { + runEvent(TaskSetFailed(taskSet, message)) } - def makeMapStatus(host: String, reduces: Int): MapStatus = - new MapStatus(makeBlockManagerId(host), Array.fill[Byte](reduces)(2)) - test("zero split job") { val rdd = makeRdd(0, Nil) var numResults = 0 - def accumulateResult(partition: Int, value: Int) { - numResults += 1 + val fakeListener = new JobListener() { + override def taskSucceeded(partition: Int, value: Any) = numResults += 1 + override def jobFailed(exception: Exception) = throw exception } - scheduler.runJob(rdd, jobComputeFunc, Seq(), "test-site", false, accumulateResult) + submit(rdd, Array(), listener = fakeListener) assert(numResults === 0) } test("run trivial job") { val rdd = makeRdd(1, Nil) - val taskSet = interceptStage(rdd) { submitRdd(rdd) } - respondToTaskSet(taskSet, List( (Success, 42) )) - expectJobResult(Array(42)) + submit(rdd, Array(0)) + complete(taskSets(0), List((Success, 42))) + assert(results === Map(0 -> 42)) } test("local job") { @@ -361,16 +206,16 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar override def getPreferredLocations(split: Partition) = Nil override def toString = "DAGSchedulerSuite Local RDD" } - submitRdd(rdd, true) - expectJobResult(Array(42)) + runEvent(JobSubmitted(rdd, jobComputeFunc, Array(0), true, null, listener)) + assert(results === Map(0 -> 42)) } test("run trivial job w/ dependency") { val baseRdd = makeRdd(1, Nil) val finalRdd = makeRdd(1, List(new OneToOneDependency(baseRdd))) - val taskSet = interceptStage(finalRdd) { submitRdd(finalRdd) } - respondToTaskSet(taskSet, List( (Success, 42) )) - expectJobResult(Array(42)) + submit(finalRdd, Array(0)) + complete(taskSets(0), Seq((Success, 42))) + assert(results === Map(0 -> 42)) } test("cache location preferences w/ dependency") { @@ -378,17 +223,17 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar val finalRdd = makeRdd(1, List(new OneToOneDependency(baseRdd))) cacheLocations(baseRdd.id -> 0) = Seq(makeBlockManagerId("hostA"), makeBlockManagerId("hostB")) - val taskSet = interceptStage(finalRdd) { submitRdd(finalRdd) } - expectTaskSetLocations(taskSet, List(Seq("hostA", "hostB"))) - respondToTaskSet(taskSet, List( (Success, 42) )) - expectJobResult(Array(42)) + submit(finalRdd, Array(0)) + val taskSet = taskSets(0) + assertLocations(taskSet, Seq(Seq("hostA", "hostB"))) + complete(taskSet, Seq((Success, 42))) + assert(results === Map(0 -> 42)) } test("trivial job failure") { - val rdd = makeRdd(1, Nil) - val taskSet = interceptStage(rdd) { submitRdd(rdd) } - runEvent(TaskSetFailed(taskSet, "test failure")) - expectJobException() + submit(makeRdd(1, Nil), Array(0)) + failed(taskSets(0), "some failure") + assert(failure.getMessage === "Job failed: some failure") } test("run trivial shuffle") { @@ -396,18 +241,14 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar val shuffleDep = new ShuffleDependency(shuffleMapRdd, null) val shuffleId = shuffleDep.shuffleId val reduceRdd = makeRdd(1, List(shuffleDep)) - - val firstStage = interceptStage(shuffleMapRdd) { submitRdd(reduceRdd) } - val secondStage = interceptStage(reduceRdd) { - respondToTaskSet(firstStage, List( + submit(reduceRdd, Array(0)) + complete(taskSets(0), Seq( (Success, makeMapStatus("hostA", 1)), - (Success, makeMapStatus("hostB", 1)) - )) - } + (Success, makeMapStatus("hostB", 1)))) assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1) === Array(makeBlockManagerId("hostA"), makeBlockManagerId("hostB"))) - respondToTaskSet(secondStage, List( (Success, 42) )) - expectJobResult(Array(42)) + complete(taskSets(1), Seq((Success, 42))) + assert(results === Map(0 -> 42)) } test("run trivial shuffle with fetch failure") { @@ -415,33 +256,24 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar val shuffleDep = new ShuffleDependency(shuffleMapRdd, null) val shuffleId = shuffleDep.shuffleId val reduceRdd = makeRdd(2, List(shuffleDep)) - - val firstStage = interceptStage(shuffleMapRdd) { submitRdd(reduceRdd) } - val secondStage = interceptStage(reduceRdd) { - respondToTaskSet(firstStage, List( + submit(reduceRdd, Array(0, 1)) + complete(taskSets(0), Seq( (Success, makeMapStatus("hostA", 1)), - (Success, makeMapStatus("hostB", 1)) - )) - } - resetExpecting { - blockManagerMaster.removeExecutor("exec-hostA") - } - whenExecuting { - respondToTaskSet(secondStage, List( + (Success, makeMapStatus("hostB", 1)))) + // the 2nd ResultTask failed + complete(taskSets(1), Seq( (Success, 42), - (FetchFailed(makeBlockManagerId("hostA"), shuffleId, 0, 0), null) - )) - } - val thirdStage = interceptStage(shuffleMapRdd) { - scheduler.resubmitFailedStages() - } - val fourthStage = interceptStage(reduceRdd) { - respondToTaskSet(thirdStage, List( (Success, makeMapStatus("hostA", 1)) )) - } - assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1) === - Array(makeBlockManagerId("hostA"), makeBlockManagerId("hostB"))) - respondToTaskSet(fourthStage, List( (Success, 43) )) - expectJobResult(Array(42, 43)) + (FetchFailed(makeBlockManagerId("hostA"), shuffleId, 0, 0), null))) + // this will get called + // blockManagerMaster.removeExecutor("exec-hostA") + // ask the scheduler to try it again + scheduler.resubmitFailedStages() + // have the 2nd attempt pass + complete(taskSets(2), Seq((Success, makeMapStatus("hostA", 1)))) + // we can see both result blocks now + assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1.host) === Array("hostA", "hostB")) + complete(taskSets(3), Seq((Success, 43))) + assert(results === Map(0 -> 42, 1 -> 43)) } test("ignore late map task completions") { @@ -449,33 +281,27 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar val shuffleDep = new ShuffleDependency(shuffleMapRdd, null) val shuffleId = shuffleDep.shuffleId val reduceRdd = makeRdd(2, List(shuffleDep)) - - val taskSet = interceptStage(shuffleMapRdd) { submitRdd(reduceRdd) } + submit(reduceRdd, Array(0, 1)) + // pretend we were told hostA went away val oldGeneration = mapOutputTracker.getGeneration - resetExpecting { - blockManagerMaster.removeExecutor("exec-hostA") - } - whenExecuting { - runEvent(ExecutorLost("exec-hostA")) - } + runEvent(ExecutorLost("exec-hostA")) val newGeneration = mapOutputTracker.getGeneration assert(newGeneration > oldGeneration) val noAccum = Map[Long, Any]() - // We rely on the event queue being ordered and increasing the generation number by 1 + val taskSet = taskSets(0) // should be ignored for being too old - runEvent(CompletionEvent(taskSet.tasks(0), Success, makeMapStatus("hostA", 1), noAccum)) + runEvent(CompletionEvent(taskSet.tasks(0), Success, makeMapStatus("hostA", 1), noAccum, null, null)) // should work because it's a non-failed host - runEvent(CompletionEvent(taskSet.tasks(0), Success, makeMapStatus("hostB", 1), noAccum)) + runEvent(CompletionEvent(taskSet.tasks(0), Success, makeMapStatus("hostB", 1), noAccum, null, null)) // should be ignored for being too old - runEvent(CompletionEvent(taskSet.tasks(0), Success, makeMapStatus("hostA", 1), noAccum)) + runEvent(CompletionEvent(taskSet.tasks(0), Success, makeMapStatus("hostA", 1), noAccum, null, null)) + // should work because it's a new generation taskSet.tasks(1).generation = newGeneration - val secondStage = interceptStage(reduceRdd) { - runEvent(CompletionEvent(taskSet.tasks(1), Success, makeMapStatus("hostA", 1), noAccum)) - } + runEvent(CompletionEvent(taskSet.tasks(1), Success, makeMapStatus("hostA", 1), noAccum, null, null)) assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1) === Array(makeBlockManagerId("hostB"), makeBlockManagerId("hostA"))) - respondToTaskSet(secondStage, List( (Success, 42), (Success, 43) )) - expectJobResult(Array(42, 43)) + complete(taskSets(1), Seq((Success, 42), (Success, 43))) + assert(results === Map(0 -> 42, 1 -> 43)) } test("run trivial shuffle with out-of-band failure and retry") { @@ -483,76 +309,49 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar val shuffleDep = new ShuffleDependency(shuffleMapRdd, null) val shuffleId = shuffleDep.shuffleId val reduceRdd = makeRdd(1, List(shuffleDep)) - - val firstStage = interceptStage(shuffleMapRdd) { submitRdd(reduceRdd) } - resetExpecting { - blockManagerMaster.removeExecutor("exec-hostA") - } - whenExecuting { - runEvent(ExecutorLost("exec-hostA")) - } + submit(reduceRdd, Array(0)) + // blockManagerMaster.removeExecutor("exec-hostA") + // pretend we were told hostA went away + runEvent(ExecutorLost("exec-hostA")) // DAGScheduler will immediately resubmit the stage after it appears to have no pending tasks // rather than marking it is as failed and waiting. - val secondStage = interceptStage(shuffleMapRdd) { - respondToTaskSet(firstStage, List( + complete(taskSets(0), Seq( (Success, makeMapStatus("hostA", 1)), - (Success, makeMapStatus("hostB", 1)) - )) - } - val thirdStage = interceptStage(reduceRdd) { - respondToTaskSet(secondStage, List( - (Success, makeMapStatus("hostC", 1)) - )) - } - assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1) === - Array(makeBlockManagerId("hostC"), makeBlockManagerId("hostB"))) - respondToTaskSet(thirdStage, List( (Success, 42) )) - expectJobResult(Array(42)) - } - - test("recursive shuffle failures") { + (Success, makeMapStatus("hostB", 1)))) + // have hostC complete the resubmitted task + complete(taskSets(1), Seq((Success, makeMapStatus("hostC", 1)))) + assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1) === + Array(makeBlockManagerId("hostC"), makeBlockManagerId("hostB"))) + complete(taskSets(2), Seq((Success, 42))) + assert(results === Map(0 -> 42)) + } + + test("recursive shuffle failures") { val shuffleOneRdd = makeRdd(2, Nil) val shuffleDepOne = new ShuffleDependency(shuffleOneRdd, null) val shuffleTwoRdd = makeRdd(2, List(shuffleDepOne)) val shuffleDepTwo = new ShuffleDependency(shuffleTwoRdd, null) val finalRdd = makeRdd(1, List(shuffleDepTwo)) - - val firstStage = interceptStage(shuffleOneRdd) { submitRdd(finalRdd) } - val secondStage = interceptStage(shuffleTwoRdd) { - respondToTaskSet(firstStage, List( + submit(finalRdd, Array(0)) + // have the first stage complete normally + complete(taskSets(0), Seq( (Success, makeMapStatus("hostA", 2)), - (Success, makeMapStatus("hostB", 2)) - )) - } - val thirdStage = interceptStage(finalRdd) { - respondToTaskSet(secondStage, List( + (Success, makeMapStatus("hostB", 2)))) + // have the second stage complete normally + complete(taskSets(1), Seq( (Success, makeMapStatus("hostA", 1)), - (Success, makeMapStatus("hostC", 1)) - )) - } - resetExpecting { - blockManagerMaster.removeExecutor("exec-hostA") - } - whenExecuting { - respondToTaskSet(thirdStage, List( - (FetchFailed(makeBlockManagerId("hostA"), shuffleDepTwo.shuffleId, 0, 0), null) - )) - } - val recomputeOne = interceptStage(shuffleOneRdd) { - scheduler.resubmitFailedStages() - } - val recomputeTwo = interceptStage(shuffleTwoRdd) { - respondToTaskSet(recomputeOne, List( - (Success, makeMapStatus("hostA", 2)) - )) - } - val finalStage = interceptStage(finalRdd) { - respondToTaskSet(recomputeTwo, List( - (Success, makeMapStatus("hostA", 1)) - )) - } - respondToTaskSet(finalStage, List( (Success, 42) )) - expectJobResult(Array(42)) + (Success, makeMapStatus("hostC", 1)))) + // fail the third stage because hostA went down + complete(taskSets(2), Seq( + (FetchFailed(makeBlockManagerId("hostA"), shuffleDepTwo.shuffleId, 0, 0), null))) + // TODO assert this: + // blockManagerMaster.removeExecutor("exec-hostA") + // have DAGScheduler try again + scheduler.resubmitFailedStages() + complete(taskSets(3), Seq((Success, makeMapStatus("hostA", 2)))) + complete(taskSets(4), Seq((Success, makeMapStatus("hostA", 1)))) + complete(taskSets(5), Seq((Success, 42))) + assert(results === Map(0 -> 42)) } test("cached post-shuffle") { @@ -561,103 +360,44 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with EasyMockSugar val shuffleTwoRdd = makeRdd(2, List(shuffleDepOne)) val shuffleDepTwo = new ShuffleDependency(shuffleTwoRdd, null) val finalRdd = makeRdd(1, List(shuffleDepTwo)) - - val firstShuffleStage = interceptStage(shuffleOneRdd) { submitRdd(finalRdd) } + submit(finalRdd, Array(0)) cacheLocations(shuffleTwoRdd.id -> 0) = Seq(makeBlockManagerId("hostD")) cacheLocations(shuffleTwoRdd.id -> 1) = Seq(makeBlockManagerId("hostC")) - val secondShuffleStage = interceptStage(shuffleTwoRdd) { - respondToTaskSet(firstShuffleStage, List( + // complete stage 2 + complete(taskSets(0), Seq( (Success, makeMapStatus("hostA", 2)), - (Success, makeMapStatus("hostB", 2)) - )) - } - val reduceStage = interceptStage(finalRdd) { - respondToTaskSet(secondShuffleStage, List( + (Success, makeMapStatus("hostB", 2)))) + // complete stage 1 + complete(taskSets(1), Seq( (Success, makeMapStatus("hostA", 1)), - (Success, makeMapStatus("hostB", 1)) - )) - } - resetExpecting { - blockManagerMaster.removeExecutor("exec-hostA") - } - whenExecuting { - respondToTaskSet(reduceStage, List( - (FetchFailed(makeBlockManagerId("hostA"), shuffleDepTwo.shuffleId, 0, 0), null) - )) - } + (Success, makeMapStatus("hostB", 1)))) + // pretend stage 0 failed because hostA went down + complete(taskSets(2), Seq( + (FetchFailed(makeBlockManagerId("hostA"), shuffleDepTwo.shuffleId, 0, 0), null))) + // TODO assert this: + // blockManagerMaster.removeExecutor("exec-hostA") // DAGScheduler should notice the cached copy of the second shuffle and try to get it rerun. - val recomputeTwo = interceptStage(shuffleTwoRdd) { - scheduler.resubmitFailedStages() - } - expectTaskSetLocations(recomputeTwo, Seq(Seq("hostD"))) - val finalRetry = interceptStage(finalRdd) { - respondToTaskSet(recomputeTwo, List( - (Success, makeMapStatus("hostD", 1)) - )) - } - respondToTaskSet(finalRetry, List( (Success, 42) )) - expectJobResult(Array(42)) + scheduler.resubmitFailedStages() + assertLocations(taskSets(3), Seq(Seq("hostD"))) + // allow hostD to recover + complete(taskSets(3), Seq((Success, makeMapStatus("hostD", 1)))) + complete(taskSets(4), Seq((Success, 42))) + assert(results === Map(0 -> 42)) } - test("cached post-shuffle but fails") { - val shuffleOneRdd = makeRdd(2, Nil) - val shuffleDepOne = new ShuffleDependency(shuffleOneRdd, null) - val shuffleTwoRdd = makeRdd(2, List(shuffleDepOne)) - val shuffleDepTwo = new ShuffleDependency(shuffleTwoRdd, null) - val finalRdd = makeRdd(1, List(shuffleDepTwo)) - - val firstShuffleStage = interceptStage(shuffleOneRdd) { submitRdd(finalRdd) } - cacheLocations(shuffleTwoRdd.id -> 0) = Seq(makeBlockManagerId("hostD")) - cacheLocations(shuffleTwoRdd.id -> 1) = Seq(makeBlockManagerId("hostC")) - val secondShuffleStage = interceptStage(shuffleTwoRdd) { - respondToTaskSet(firstShuffleStage, List( - (Success, makeMapStatus("hostA", 2)), - (Success, makeMapStatus("hostB", 2)) - )) - } - val reduceStage = interceptStage(finalRdd) { - respondToTaskSet(secondShuffleStage, List( - (Success, makeMapStatus("hostA", 1)), - (Success, makeMapStatus("hostB", 1)) - )) - } - resetExpecting { - blockManagerMaster.removeExecutor("exec-hostA") - } - whenExecuting { - respondToTaskSet(reduceStage, List( - (FetchFailed(makeBlockManagerId("hostA"), shuffleDepTwo.shuffleId, 0, 0), null) - )) - } - val recomputeTwoCached = interceptStage(shuffleTwoRdd) { - scheduler.resubmitFailedStages() - } - expectTaskSetLocations(recomputeTwoCached, Seq(Seq("hostD"))) - intercept[FetchFailedException]{ - mapOutputTracker.getServerStatuses(shuffleDepOne.shuffleId, 0) + /** Assert that the supplied TaskSet has exactly the given preferredLocations. Note, converts taskSet's locations to host only. */ + private def assertLocations(taskSet: TaskSet, locations: Seq[Seq[String]]) { + assert(locations.size === taskSet.tasks.size) + for ((expectLocs, taskLocs) <- + taskSet.tasks.map(_.preferredLocations).zip(locations)) { + assert(expectLocs.map(loc => spark.Utils.parseHostPort(loc)._1) === taskLocs) } + } - // Simulate the shuffle input data failing to be cached. - cacheLocations.remove(shuffleTwoRdd.id -> 0) - respondToTaskSet(recomputeTwoCached, List( - (FetchFailed(null, shuffleDepOne.shuffleId, 0, 0), null) - )) + private def makeMapStatus(host: String, reduces: Int): MapStatus = + new MapStatus(makeBlockManagerId(host), Array.fill[Byte](reduces)(2)) - // After the fetch failure, DAGScheduler should recheck the cache and decide to resubmit - // everything. - val recomputeOne = interceptStage(shuffleOneRdd) { - scheduler.resubmitFailedStages() - } - // We use hostA here to make sure DAGScheduler doesn't think it's still dead. - val recomputeTwoUncached = interceptStage(shuffleTwoRdd) { - respondToTaskSet(recomputeOne, List( (Success, makeMapStatus("hostA", 1)) )) - } - expectTaskSetLocations(recomputeTwoUncached, Seq(Seq[String]())) - val finalRetry = interceptStage(finalRdd) { - respondToTaskSet(recomputeTwoUncached, List( (Success, makeMapStatus("hostA", 1)) )) + private def makeBlockManagerId(host: String): BlockManagerId = + BlockManagerId("exec-" + host, host, 12345, 0) - } - respondToTaskSet(finalRetry, List( (Success, 42) )) - expectJobResult(Array(42)) - } } diff --git a/core/src/test/scala/spark/scheduler/JobLoggerSuite.scala b/core/src/test/scala/spark/scheduler/JobLoggerSuite.scala new file mode 100644 index 0000000000..4000c4d520 --- /dev/null +++ b/core/src/test/scala/spark/scheduler/JobLoggerSuite.scala @@ -0,0 +1,105 @@ +package spark.scheduler + +import java.util.Properties +import java.util.concurrent.LinkedBlockingQueue +import org.scalatest.FunSuite +import org.scalatest.matchers.ShouldMatchers +import scala.collection.mutable +import spark._ +import spark.SparkContext._ + + +class JobLoggerSuite extends FunSuite with LocalSparkContext with ShouldMatchers { + + test("inner method") { + sc = new SparkContext("local", "joblogger") + val joblogger = new JobLogger { + def createLogWriterTest(jobID: Int) = createLogWriter(jobID) + def closeLogWriterTest(jobID: Int) = closeLogWriter(jobID) + def getRddNameTest(rdd: RDD[_]) = getRddName(rdd) + def buildJobDepTest(jobID: Int, stage: Stage) = buildJobDep(jobID, stage) + } + type MyRDD = RDD[(Int, Int)] + def makeRdd( + numPartitions: Int, + dependencies: List[Dependency[_]] + ): MyRDD = { + val maxPartition = numPartitions - 1 + return new MyRDD(sc, dependencies) { + override def compute(split: Partition, context: TaskContext): Iterator[(Int, Int)] = + throw new RuntimeException("should not be reached") + override def getPartitions = (0 to maxPartition).map(i => new Partition { + override def index = i + }).toArray + } + } + val jobID = 5 + val parentRdd = makeRdd(4, Nil) + val shuffleDep = new ShuffleDependency(parentRdd, null) + val rootRdd = makeRdd(4, List(shuffleDep)) + val shuffleMapStage = new Stage(1, parentRdd, Some(shuffleDep), Nil, jobID) + val rootStage = new Stage(0, rootRdd, None, List(shuffleMapStage), jobID) + + joblogger.onStageSubmitted(SparkListenerStageSubmitted(rootStage, 4)) + joblogger.getEventQueue.size should be (1) + joblogger.getRddNameTest(parentRdd) should be (parentRdd.getClass.getName) + parentRdd.setName("MyRDD") + joblogger.getRddNameTest(parentRdd) should be ("MyRDD") + joblogger.createLogWriterTest(jobID) + joblogger.getJobIDtoPrintWriter.size should be (1) + joblogger.buildJobDepTest(jobID, rootStage) + joblogger.getJobIDToStages.get(jobID).get.size should be (2) + joblogger.getStageIDToJobID.get(0) should be (Some(jobID)) + joblogger.getStageIDToJobID.get(1) should be (Some(jobID)) + joblogger.closeLogWriterTest(jobID) + joblogger.getStageIDToJobID.size should be (0) + joblogger.getJobIDToStages.size should be (0) + joblogger.getJobIDtoPrintWriter.size should be (0) + } + + test("inner variables") { + sc = new SparkContext("local[4]", "joblogger") + val joblogger = new JobLogger { + override protected def closeLogWriter(jobID: Int) = + getJobIDtoPrintWriter.get(jobID).foreach { fileWriter => + fileWriter.close() + } + } + sc.addSparkListener(joblogger) + val rdd = sc.parallelize(1 to 1e2.toInt, 4).map{ i => (i % 12, 2 * i) } + rdd.reduceByKey(_+_).collect() + + joblogger.getLogDir should be ("/tmp/spark") + joblogger.getJobIDtoPrintWriter.size should be (1) + joblogger.getStageIDToJobID.size should be (2) + joblogger.getStageIDToJobID.get(0) should be (Some(0)) + joblogger.getStageIDToJobID.get(1) should be (Some(0)) + joblogger.getJobIDToStages.size should be (1) + } + + + test("interface functions") { + sc = new SparkContext("local[4]", "joblogger") + val joblogger = new JobLogger { + var onTaskEndCount = 0 + var onJobEndCount = 0 + var onJobStartCount = 0 + var onStageCompletedCount = 0 + var onStageSubmittedCount = 0 + override def onTaskEnd(taskEnd: SparkListenerTaskEnd) = onTaskEndCount += 1 + override def onJobEnd(jobEnd: SparkListenerJobEnd) = onJobEndCount += 1 + override def onJobStart(jobStart: SparkListenerJobStart) = onJobStartCount += 1 + override def onStageCompleted(stageCompleted: StageCompleted) = onStageCompletedCount += 1 + override def onStageSubmitted(stageSubmitted: SparkListenerStageSubmitted) = onStageSubmittedCount += 1 + } + sc.addSparkListener(joblogger) + val rdd = sc.parallelize(1 to 1e2.toInt, 4).map{ i => (i % 12, 2 * i) } + rdd.reduceByKey(_+_).collect() + + joblogger.onJobStartCount should be (1) + joblogger.onJobEndCount should be (1) + joblogger.onTaskEndCount should be (8) + joblogger.onStageSubmittedCount should be (2) + joblogger.onStageCompletedCount should be (2) + } +} diff --git a/core/src/test/scala/spark/scheduler/LocalSchedulerSuite.scala b/core/src/test/scala/spark/scheduler/LocalSchedulerSuite.scala new file mode 100644 index 0000000000..8bd813fd14 --- /dev/null +++ b/core/src/test/scala/spark/scheduler/LocalSchedulerSuite.scala @@ -0,0 +1,206 @@ +package spark.scheduler + +import org.scalatest.FunSuite +import org.scalatest.BeforeAndAfter + +import spark._ +import spark.scheduler._ +import spark.scheduler.cluster._ +import scala.collection.mutable.ArrayBuffer +import scala.collection.mutable.{ConcurrentMap, HashMap} +import java.util.concurrent.Semaphore +import java.util.concurrent.CountDownLatch +import java.util.Properties + +class Lock() { + var finished = false + def jobWait() = { + synchronized { + while(!finished) { + this.wait() + } + } + } + + def jobFinished() = { + synchronized { + finished = true + this.notifyAll() + } + } +} + +object TaskThreadInfo { + val threadToLock = HashMap[Int, Lock]() + val threadToRunning = HashMap[Int, Boolean]() + val threadToStarted = HashMap[Int, CountDownLatch]() +} + +/* + * 1. each thread contains one job. + * 2. each job contains one stage. + * 3. each stage only contains one task. + * 4. each task(launched) must be lanched orderly(using threadToStarted) to make sure + * it will get cpu core resource, and will wait to finished after user manually + * release "Lock" and then cluster will contain another free cpu cores. + * 5. each task(pending) must use "sleep" to make sure it has been added to taskSetManager queue, + * thus it will be scheduled later when cluster has free cpu cores. + */ +class LocalSchedulerSuite extends FunSuite with LocalSparkContext { + + def createThread(threadIndex: Int, poolName: String, sc: SparkContext, sem: Semaphore) { + + TaskThreadInfo.threadToRunning(threadIndex) = false + val nums = sc.parallelize(threadIndex to threadIndex, 1) + TaskThreadInfo.threadToLock(threadIndex) = new Lock() + TaskThreadInfo.threadToStarted(threadIndex) = new CountDownLatch(1) + new Thread { + if (poolName != null) { + sc.addLocalProperties("spark.scheduler.cluster.fair.pool",poolName) + } + override def run() { + val ans = nums.map(number => { + TaskThreadInfo.threadToRunning(number) = true + TaskThreadInfo.threadToStarted(number).countDown() + TaskThreadInfo.threadToLock(number).jobWait() + TaskThreadInfo.threadToRunning(number) = false + number + }).collect() + assert(ans.toList === List(threadIndex)) + sem.release() + } + }.start() + } + + test("Local FIFO scheduler end-to-end test") { + System.setProperty("spark.cluster.schedulingmode", "FIFO") + sc = new SparkContext("local[4]", "test") + val sem = new Semaphore(0) + + createThread(1,null,sc,sem) + TaskThreadInfo.threadToStarted(1).await() + createThread(2,null,sc,sem) + TaskThreadInfo.threadToStarted(2).await() + createThread(3,null,sc,sem) + TaskThreadInfo.threadToStarted(3).await() + createThread(4,null,sc,sem) + TaskThreadInfo.threadToStarted(4).await() + // thread 5 and 6 (stage pending)must meet following two points + // 1. stages (taskSetManager) of jobs in thread 5 and 6 should be add to taskSetManager + // queue before executing TaskThreadInfo.threadToLock(1).jobFinished() + // 2. priority of stage in thread 5 should be prior to priority of stage in thread 6 + // So I just use "sleep" 1s here for each thread. + // TODO: any better solution? + createThread(5,null,sc,sem) + Thread.sleep(1000) + createThread(6,null,sc,sem) + Thread.sleep(1000) + + assert(TaskThreadInfo.threadToRunning(1) === true) + assert(TaskThreadInfo.threadToRunning(2) === true) + assert(TaskThreadInfo.threadToRunning(3) === true) + assert(TaskThreadInfo.threadToRunning(4) === true) + assert(TaskThreadInfo.threadToRunning(5) === false) + assert(TaskThreadInfo.threadToRunning(6) === false) + + TaskThreadInfo.threadToLock(1).jobFinished() + TaskThreadInfo.threadToStarted(5).await() + + assert(TaskThreadInfo.threadToRunning(1) === false) + assert(TaskThreadInfo.threadToRunning(2) === true) + assert(TaskThreadInfo.threadToRunning(3) === true) + assert(TaskThreadInfo.threadToRunning(4) === true) + assert(TaskThreadInfo.threadToRunning(5) === true) + assert(TaskThreadInfo.threadToRunning(6) === false) + + TaskThreadInfo.threadToLock(3).jobFinished() + TaskThreadInfo.threadToStarted(6).await() + + assert(TaskThreadInfo.threadToRunning(1) === false) + assert(TaskThreadInfo.threadToRunning(2) === true) + assert(TaskThreadInfo.threadToRunning(3) === false) + assert(TaskThreadInfo.threadToRunning(4) === true) + assert(TaskThreadInfo.threadToRunning(5) === true) + assert(TaskThreadInfo.threadToRunning(6) === true) + + TaskThreadInfo.threadToLock(2).jobFinished() + TaskThreadInfo.threadToLock(4).jobFinished() + TaskThreadInfo.threadToLock(5).jobFinished() + TaskThreadInfo.threadToLock(6).jobFinished() + sem.acquire(6) + } + + test("Local fair scheduler end-to-end test") { + sc = new SparkContext("local[8]", "LocalSchedulerSuite") + val sem = new Semaphore(0) + System.setProperty("spark.cluster.schedulingmode", "FAIR") + val xmlPath = getClass.getClassLoader.getResource("fairscheduler.xml").getFile() + System.setProperty("spark.fairscheduler.allocation.file", xmlPath) + + createThread(10,"1",sc,sem) + TaskThreadInfo.threadToStarted(10).await() + createThread(20,"2",sc,sem) + TaskThreadInfo.threadToStarted(20).await() + createThread(30,"3",sc,sem) + TaskThreadInfo.threadToStarted(30).await() + + assert(TaskThreadInfo.threadToRunning(10) === true) + assert(TaskThreadInfo.threadToRunning(20) === true) + assert(TaskThreadInfo.threadToRunning(30) === true) + + createThread(11,"1",sc,sem) + TaskThreadInfo.threadToStarted(11).await() + createThread(21,"2",sc,sem) + TaskThreadInfo.threadToStarted(21).await() + createThread(31,"3",sc,sem) + TaskThreadInfo.threadToStarted(31).await() + + assert(TaskThreadInfo.threadToRunning(11) === true) + assert(TaskThreadInfo.threadToRunning(21) === true) + assert(TaskThreadInfo.threadToRunning(31) === true) + + createThread(12,"1",sc,sem) + TaskThreadInfo.threadToStarted(12).await() + createThread(22,"2",sc,sem) + TaskThreadInfo.threadToStarted(22).await() + createThread(32,"3",sc,sem) + + assert(TaskThreadInfo.threadToRunning(12) === true) + assert(TaskThreadInfo.threadToRunning(22) === true) + assert(TaskThreadInfo.threadToRunning(32) === false) + + TaskThreadInfo.threadToLock(10).jobFinished() + TaskThreadInfo.threadToStarted(32).await() + + assert(TaskThreadInfo.threadToRunning(32) === true) + + //1. Similar with above scenario, sleep 1s for stage of 23 and 33 to be added to taskSetManager + // queue so that cluster will assign free cpu core to stage 23 after stage 11 finished. + //2. priority of 23 and 33 will be meaningless as using fair scheduler here. + createThread(23,"2",sc,sem) + createThread(33,"3",sc,sem) + Thread.sleep(1000) + + TaskThreadInfo.threadToLock(11).jobFinished() + TaskThreadInfo.threadToStarted(23).await() + + assert(TaskThreadInfo.threadToRunning(23) === true) + assert(TaskThreadInfo.threadToRunning(33) === false) + + TaskThreadInfo.threadToLock(12).jobFinished() + TaskThreadInfo.threadToStarted(33).await() + + assert(TaskThreadInfo.threadToRunning(33) === true) + + TaskThreadInfo.threadToLock(20).jobFinished() + TaskThreadInfo.threadToLock(21).jobFinished() + TaskThreadInfo.threadToLock(22).jobFinished() + TaskThreadInfo.threadToLock(23).jobFinished() + TaskThreadInfo.threadToLock(30).jobFinished() + TaskThreadInfo.threadToLock(31).jobFinished() + TaskThreadInfo.threadToLock(32).jobFinished() + TaskThreadInfo.threadToLock(33).jobFinished() + + sem.acquire(11) + } +} diff --git a/core/src/test/scala/spark/scheduler/SparkListenerSuite.scala b/core/src/test/scala/spark/scheduler/SparkListenerSuite.scala new file mode 100644 index 0000000000..48aa67c543 --- /dev/null +++ b/core/src/test/scala/spark/scheduler/SparkListenerSuite.scala @@ -0,0 +1,85 @@ +package spark.scheduler + +import org.scalatest.FunSuite +import spark.{SparkContext, LocalSparkContext} +import scala.collection.mutable +import org.scalatest.matchers.ShouldMatchers +import spark.SparkContext._ + +/** + * + */ + +class SparkListenerSuite extends FunSuite with LocalSparkContext with ShouldMatchers { + + test("local metrics") { + sc = new SparkContext("local[4]", "test") + val listener = new SaveStageInfo + sc.addSparkListener(listener) + sc.addSparkListener(new StatsReportListener) + //just to make sure some of the tasks take a noticeable amount of time + val w = {i:Int => + if (i == 0) + Thread.sleep(100) + i + } + + val d = sc.parallelize(1 to 1e4.toInt, 64).map{i => w(i)} + d.count + listener.stageInfos.size should be (1) + + val d2 = d.map{i => w(i) -> i * 2}.setName("shuffle input 1") + + val d3 = d.map{i => w(i) -> (0 to (i % 5))}.setName("shuffle input 2") + + val d4 = d2.cogroup(d3, 64).map{case(k,(v1,v2)) => w(k) -> (v1.size, v2.size)} + d4.setName("A Cogroup") + + d4.collectAsMap + + listener.stageInfos.size should be (4) + listener.stageInfos.foreach {stageInfo => + //small test, so some tasks might take less than 1 millisecond, but average should be greater than 1 ms + checkNonZeroAvg(stageInfo.taskInfos.map{_._1.duration}, stageInfo + " duration") + checkNonZeroAvg(stageInfo.taskInfos.map{_._2.executorRunTime.toLong}, stageInfo + " executorRunTime") + checkNonZeroAvg(stageInfo.taskInfos.map{_._2.executorDeserializeTime.toLong}, stageInfo + " executorDeserializeTime") + if (stageInfo.stage.rdd.name == d4.name) { + checkNonZeroAvg(stageInfo.taskInfos.map{_._2.shuffleReadMetrics.get.fetchWaitTime}, stageInfo + " fetchWaitTime") + } + + stageInfo.taskInfos.foreach{case (taskInfo, taskMetrics) => + taskMetrics.resultSize should be > (0l) + if (isStage(stageInfo, Set(d2.name, d3.name), Set(d4.name))) { + taskMetrics.shuffleWriteMetrics should be ('defined) + taskMetrics.shuffleWriteMetrics.get.shuffleBytesWritten should be > (0l) + } + if (stageInfo.stage.rdd.name == d4.name) { + taskMetrics.shuffleReadMetrics should be ('defined) + val sm = taskMetrics.shuffleReadMetrics.get + sm.totalBlocksFetched should be > (0) + sm.localBlocksFetched should be > (0) + sm.remoteBlocksFetched should be (0) + sm.remoteBytesRead should be (0l) + sm.remoteFetchTime should be (0l) + } + } + } + } + + def checkNonZeroAvg(m: Traversable[Long], msg: String) { + assert(m.sum / m.size.toDouble > 0.0, msg) + } + + def isStage(stageInfo: StageInfo, rddNames: Set[String], excludedNames: Set[String]) = { + val names = Set(stageInfo.stage.rdd.name) ++ stageInfo.stage.rdd.dependencies.map{_.rdd.name} + !names.intersect(rddNames).isEmpty && names.intersect(excludedNames).isEmpty + } + + class SaveStageInfo extends SparkListener { + val stageInfos = mutable.Buffer[StageInfo]() + override def onStageCompleted(stage: StageCompleted) { + stageInfos += stage.stageInfo + } + } + +} diff --git a/core/src/test/scala/spark/storage/BlockManagerSuite.scala b/core/src/test/scala/spark/storage/BlockManagerSuite.scala index 2d177bbf67..b9d5f9668e 100644 --- a/core/src/test/scala/spark/storage/BlockManagerSuite.scala +++ b/core/src/test/scala/spark/storage/BlockManagerSuite.scala @@ -12,10 +12,13 @@ import org.scalatest.concurrent.Timeouts._ import org.scalatest.matchers.ShouldMatchers._ import org.scalatest.time.SpanSugar._ +import spark.JavaSerializer import spark.KryoSerializer import spark.SizeEstimator +import spark.util.AkkaUtils import spark.util.ByteBufferInputStream + class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodTester { var store: BlockManager = null var store2: BlockManager = null @@ -30,8 +33,13 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT val serializer = new KryoSerializer before { - actorSystem = ActorSystem("test") - master = new BlockManagerMaster(actorSystem, true, true, "localhost", 7077) + val (actorSystem, boundPort) = AkkaUtils.createActorSystem("test", "localhost", 0) + this.actorSystem = actorSystem + System.setProperty("spark.driver.port", boundPort.toString) + System.setProperty("spark.hostPort", "localhost:" + boundPort) + + master = new BlockManagerMaster( + actorSystem.actorOf(Props(new spark.storage.BlockManagerMasterActor(true)))) // Set the arch to 64-bit and compressedOops to true to get a deterministic test-case oldArch = System.setProperty("os.arch", "amd64") @@ -39,9 +47,14 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT oldHeartBeat = System.setProperty("spark.storage.disableBlockManagerHeartBeat", "true") val initialize = PrivateMethod[Unit]('initialize) SizeEstimator invokePrivate initialize() + // Set some value ... + System.setProperty("spark.hostPort", spark.Utils.localHostName() + ":" + 1111) } after { + System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") + if (store != null) { store.stop() store = null @@ -86,9 +99,9 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT } test("BlockManagerId object caching") { - val id1 = BlockManagerId("e1", "XXX", 1) - val id2 = BlockManagerId("e1", "XXX", 1) // this should return the same object as id1 - val id3 = BlockManagerId("e1", "XXX", 2) // this should return a different object + val id1 = BlockManagerId("e1", "XXX", 1, 0) + val id2 = BlockManagerId("e1", "XXX", 1, 0) // this should return the same object as id1 + val id3 = BlockManagerId("e1", "XXX", 2, 0) // this should return a different object assert(id2 === id1, "id2 is not same as id1") assert(id2.eq(id1), "id2 is not the same object as id1") assert(id3 != id1, "id3 is same as id1") @@ -111,7 +124,7 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT // Putting a1, a2 and a3 in memory and telling master only about a1 and a2 store.putSingle("a1", a1, StorageLevel.MEMORY_ONLY) store.putSingle("a2", a2, StorageLevel.MEMORY_ONLY) - store.putSingle("a3", a3, StorageLevel.MEMORY_ONLY, false) + store.putSingle("a3", a3, StorageLevel.MEMORY_ONLY, tellMaster = false) // Checking whether blocks are in memory assert(store.getSingle("a1") != None, "a1 was not in store") @@ -157,7 +170,7 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT // Putting a1, a2 and a3 in memory and telling master only about a1 and a2 store.putSingle("a1-to-remove", a1, StorageLevel.MEMORY_ONLY) store.putSingle("a2-to-remove", a2, StorageLevel.MEMORY_ONLY) - store.putSingle("a3-to-remove", a3, StorageLevel.MEMORY_ONLY, false) + store.putSingle("a3-to-remove", a3, StorageLevel.MEMORY_ONLY, tellMaster = false) // Checking whether blocks are in memory and memory size val memStatus = master.getMemoryStatus.head._2 @@ -196,6 +209,39 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT } } + test("removing rdd") { + store = new BlockManager("<driver>", actorSystem, master, serializer, 2000) + val a1 = new Array[Byte](400) + val a2 = new Array[Byte](400) + val a3 = new Array[Byte](400) + // Putting a1, a2 and a3 in memory. + store.putSingle("rdd_0_0", a1, StorageLevel.MEMORY_ONLY) + store.putSingle("rdd_0_1", a2, StorageLevel.MEMORY_ONLY) + store.putSingle("nonrddblock", a3, StorageLevel.MEMORY_ONLY) + master.removeRdd(0, blocking = false) + + eventually(timeout(1000 milliseconds), interval(10 milliseconds)) { + store.getSingle("rdd_0_0") should be (None) + master.getLocations("rdd_0_0") should have size 0 + } + eventually(timeout(1000 milliseconds), interval(10 milliseconds)) { + store.getSingle("rdd_0_1") should be (None) + master.getLocations("rdd_0_1") should have size 0 + } + eventually(timeout(1000 milliseconds), interval(10 milliseconds)) { + store.getSingle("nonrddblock") should not be (None) + master.getLocations("nonrddblock") should have size (1) + } + + store.putSingle("rdd_0_0", a1, StorageLevel.MEMORY_ONLY) + store.putSingle("rdd_0_1", a2, StorageLevel.MEMORY_ONLY) + master.removeRdd(0, blocking = true) + store.getSingle("rdd_0_0") should be (None) + master.getLocations("rdd_0_0") should have size 0 + store.getSingle("rdd_0_1") should be (None) + master.getLocations("rdd_0_1") should have size 0 + } + test("reregistration on heart beat") { val heartBeat = PrivateMethod[Unit]('heartBeat) store = new BlockManager("<driver>", actorSystem, master, serializer, 2000) @@ -224,7 +270,7 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT master.removeExecutor(store.blockManagerId.executorId) assert(master.getLocations("a1").size == 0, "a1 was not removed from master") - store.putSingle("a2", a1, StorageLevel.MEMORY_ONLY) + store.putSingle("a2", a2, StorageLevel.MEMORY_ONLY) store.waitForAsyncReregister() assert(master.getLocations("a1").size > 0, "a1 was not reregistered with master") @@ -242,7 +288,7 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT master.removeExecutor(store.blockManagerId.executorId) val t1 = new Thread { override def run() { - store.put("a2", a2.iterator, StorageLevel.MEMORY_ONLY, true) + store.put("a2", a2.iterator, StorageLevel.MEMORY_ONLY, tellMaster = true) } } val t2 = new Thread { @@ -262,7 +308,7 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT t1.join() t2.join() t3.join() - + store.dropFromMemory("a1", null) store.dropFromMemory("a2", null) store.waitForAsyncReregister() @@ -452,9 +498,9 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT val list1 = List(new Array[Byte](200), new Array[Byte](200)) val list2 = List(new Array[Byte](200), new Array[Byte](200)) val list3 = List(new Array[Byte](200), new Array[Byte](200)) - store.put("list1", list1.iterator, StorageLevel.MEMORY_ONLY, true) - store.put("list2", list2.iterator, StorageLevel.MEMORY_ONLY, true) - store.put("list3", list3.iterator, StorageLevel.MEMORY_ONLY, true) + store.put("list1", list1.iterator, StorageLevel.MEMORY_ONLY, tellMaster = true) + store.put("list2", list2.iterator, StorageLevel.MEMORY_ONLY, tellMaster = true) + store.put("list3", list3.iterator, StorageLevel.MEMORY_ONLY, tellMaster = true) assert(store.get("list2") != None, "list2 was not in store") assert(store.get("list2").get.size == 2) assert(store.get("list3") != None, "list3 was not in store") @@ -463,7 +509,7 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT assert(store.get("list2") != None, "list2 was not in store") assert(store.get("list2").get.size == 2) // At this point list2 was gotten last, so LRU will getSingle rid of list3 - store.put("list1", list1.iterator, StorageLevel.MEMORY_ONLY, true) + store.put("list1", list1.iterator, StorageLevel.MEMORY_ONLY, tellMaster = true) assert(store.get("list1") != None, "list1 was not in store") assert(store.get("list1").get.size == 2) assert(store.get("list2") != None, "list2 was not in store") @@ -478,9 +524,9 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT val list3 = List(new Array[Byte](200), new Array[Byte](200)) val list4 = List(new Array[Byte](200), new Array[Byte](200)) // First store list1 and list2, both in memory, and list3, on disk only - store.put("list1", list1.iterator, StorageLevel.MEMORY_ONLY_SER, true) - store.put("list2", list2.iterator, StorageLevel.MEMORY_ONLY_SER, true) - store.put("list3", list3.iterator, StorageLevel.DISK_ONLY, true) + store.put("list1", list1.iterator, StorageLevel.MEMORY_ONLY_SER, tellMaster = true) + store.put("list2", list2.iterator, StorageLevel.MEMORY_ONLY_SER, tellMaster = true) + store.put("list3", list3.iterator, StorageLevel.DISK_ONLY, tellMaster = true) // At this point LRU should not kick in because list3 is only on disk assert(store.get("list1") != None, "list2 was not in store") assert(store.get("list1").get.size === 2) @@ -495,7 +541,7 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT assert(store.get("list3") != None, "list1 was not in store") assert(store.get("list3").get.size === 2) // Now let's add in list4, which uses both disk and memory; list1 should drop out - store.put("list4", list4.iterator, StorageLevel.MEMORY_AND_DISK_SER, true) + store.put("list4", list4.iterator, StorageLevel.MEMORY_AND_DISK_SER, tellMaster = true) assert(store.get("list1") === None, "list1 was in store") assert(store.get("list2") != None, "list3 was not in store") assert(store.get("list2").get.size === 2) @@ -582,4 +628,21 @@ class BlockManagerSuite extends FunSuite with BeforeAndAfter with PrivateMethodT System.clearProperty("spark.rdd.compress") } } + + test("block store put failure") { + // Use Java serializer so we can create an unserializable error. + store = new BlockManager("<driver>", actorSystem, master, new JavaSerializer, 1200) + + // The put should fail since a1 is not serializable. + class UnserializableClass + val a1 = new UnserializableClass + intercept[java.io.NotSerializableException] { + store.putSingle("a1", a1, StorageLevel.DISK_ONLY) + } + + // Make sure get a1 doesn't hang and returns None. + failAfter(1 second) { + assert(store.getSingle("a1") == None, "a1 should not be in store") + } + } } diff --git a/core/src/test/scala/spark/util/DistributionSuite.scala b/core/src/test/scala/spark/util/DistributionSuite.scala new file mode 100644 index 0000000000..cc6249b1dd --- /dev/null +++ b/core/src/test/scala/spark/util/DistributionSuite.scala @@ -0,0 +1,25 @@ +package spark.util + +import org.scalatest.FunSuite +import org.scalatest.matchers.ShouldMatchers + +/** + * + */ + +class DistributionSuite extends FunSuite with ShouldMatchers { + test("summary") { + val d = new Distribution((1 to 100).toArray.map{_.toDouble}) + val stats = d.statCounter + stats.count should be (100) + stats.mean should be (50.5) + stats.sum should be (50 * 101) + + val quantiles = d.getQuantiles() + quantiles(0) should be (1) + quantiles(1) should be (26) + quantiles(2) should be (51) + quantiles(3) should be (76) + quantiles(4) should be (100) + } +} diff --git a/core/src/test/scala/spark/util/NextIteratorSuite.scala b/core/src/test/scala/spark/util/NextIteratorSuite.scala new file mode 100644 index 0000000000..ed5b36da73 --- /dev/null +++ b/core/src/test/scala/spark/util/NextIteratorSuite.scala @@ -0,0 +1,68 @@ +package spark.util + +import org.scalatest.FunSuite +import org.scalatest.matchers.ShouldMatchers +import scala.collection.mutable.Buffer +import java.util.NoSuchElementException + +class NextIteratorSuite extends FunSuite with ShouldMatchers { + test("one iteration") { + val i = new StubIterator(Buffer(1)) + i.hasNext should be === true + i.next should be === 1 + i.hasNext should be === false + intercept[NoSuchElementException] { i.next() } + } + + test("two iterations") { + val i = new StubIterator(Buffer(1, 2)) + i.hasNext should be === true + i.next should be === 1 + i.hasNext should be === true + i.next should be === 2 + i.hasNext should be === false + intercept[NoSuchElementException] { i.next() } + } + + test("empty iteration") { + val i = new StubIterator(Buffer()) + i.hasNext should be === false + intercept[NoSuchElementException] { i.next() } + } + + test("close is called once for empty iterations") { + val i = new StubIterator(Buffer()) + i.hasNext should be === false + i.hasNext should be === false + i.closeCalled should be === 1 + } + + test("close is called once for non-empty iterations") { + val i = new StubIterator(Buffer(1, 2)) + i.next should be === 1 + i.next should be === 2 + // close isn't called until we check for the next element + i.closeCalled should be === 0 + i.hasNext should be === false + i.closeCalled should be === 1 + i.hasNext should be === false + i.closeCalled should be === 1 + } + + class StubIterator(ints: Buffer[Int]) extends NextIterator[Int] { + var closeCalled = 0 + + override def getNext() = { + if (ints.size == 0) { + finished = true + 0 + } else { + ints.remove(0) + } + } + + override def close() { + closeCalled += 1 + } + } +} diff --git a/docs/README.md b/docs/README.md index 887f407f18..c2b3497bb3 100644 --- a/docs/README.md +++ b/docs/README.md @@ -33,4 +33,4 @@ Similarly, you can build just the PySpark epydoc by running `epydoc --config epy When you run `jekyll` in the docs directory, it will also copy over the scaladoc for the various Spark subprojects into the docs directory (and then also into the _site directory). We use a jekyll plugin to run `sbt/sbt doc` before building the site so if you haven't run it (recently) it may take some time as it generates all of the scaladoc. The jekyll plugin also generates the PySpark docs using [epydoc](http://epydoc.sourceforge.net/). -NOTE: To skip the step of building and copying over the scaladoc when you build the docs, run `SKIP_SCALADOC=1 jekyll`. Similarly, `SKIP_EPYDOC=1 jekyll` will skip PySpark API doc generation. +NOTE: To skip the step of building and copying over the Scala and Python API docs, run `SKIP_API=1 jekyll`. diff --git a/docs/_config.yml b/docs/_config.yml index 09617e4a1e..5c135a0242 100644 --- a/docs/_config.yml +++ b/docs/_config.yml @@ -3,8 +3,8 @@ markdown: kramdown # These allow the documentation to be updated with nerw releases # of Spark, Scala, and Mesos. -SPARK_VERSION: 0.7.0-SNAPSHOT -SPARK_VERSION_SHORT: 0.7.0 -SCALA_VERSION: 2.9.2 +SPARK_VERSION: 0.8.0-SNAPSHOT +SPARK_VERSION_SHORT: 0.8.0 +SCALA_VERSION: 2.9.3 MESOS_VERSION: 0.9.0-incubating SPARK_ISSUE_TRACKER_URL: https://spark-project.atlassian.net diff --git a/docs/_layouts/global.html b/docs/_layouts/global.html index 94baa634aa..f06ab2d5b0 100755 --- a/docs/_layouts/global.html +++ b/docs/_layouts/global.html @@ -23,6 +23,22 @@ <script src="js/vendor/modernizr-2.6.1-respond-1.1.0.min.js"></script> <link rel="stylesheet" href="css/pygments-default.css"> + + <!-- Google analytics script --> + <script type="text/javascript"> + /* + var _gaq = _gaq || []; + _gaq.push(['_setAccount', 'UA-32518208-1']); + _gaq.push(['_trackPageview']); + + (function() { + var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; + ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; + var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); + })(); + */ + </script> + </head> <body> <!--[if lt IE 7]> @@ -53,11 +69,11 @@ </li> <li class="dropdown"> - <a href="#" class="dropdown-toggle" data-toggle="dropdown">API (Scaladoc)<b class="caret"></b></a> + <a href="#" class="dropdown-toggle" data-toggle="dropdown">API Docs<b class="caret"></b></a> <ul class="dropdown-menu"> - <li><a href="api/core/index.html">Spark Scala/Java (Scaladoc)</a></li> + <li><a href="api/core/index.html">Spark Java/Scala (Scaladoc)</a></li> <li><a href="api/pyspark/index.html">Spark Python (Epydoc)</a></li> - <li><a href="api/streaming/index.html">Spark Streaming Scala/Java (Scaladoc) </a></li> + <li><a href="api/streaming/index.html">Spark Streaming Java/Scala (Scaladoc) </a></li> </ul> </li> @@ -74,6 +90,7 @@ <li class="dropdown"> <a href="api.html" class="dropdown-toggle" data-toggle="dropdown">More<b class="caret"></b></a> <ul class="dropdown-menu"> + <li><a href="building-with-maven.html">Building Spark with Maven</a></li> <li><a href="configuration.html">Configuration</a></li> <li><a href="tuning.html">Tuning Guide</a></li> <li><a href="bagel-programming-guide.html">Bagel (Pregel on Spark)</a></li> diff --git a/docs/_plugins/copy_api_dirs.rb b/docs/_plugins/copy_api_dirs.rb index e400dec619..c10ae595de 100644 --- a/docs/_plugins/copy_api_dirs.rb +++ b/docs/_plugins/copy_api_dirs.rb @@ -1,7 +1,8 @@ require 'fileutils' include FileUtils -if ENV['SKIP_SCALADOC'] != '1' +if ENV['SKIP_API'] != '1' + # Build Scaladoc for Java/Scala projects = ["core", "examples", "repl", "bagel", "streaming"] puts "Moving to project root and building scaladoc." @@ -17,7 +18,7 @@ if ENV['SKIP_SCALADOC'] != '1' # Copy over the scaladoc from each project into the docs directory. # This directory will be copied over to _site when `jekyll` command is run. projects.each do |project_name| - source = "../" + project_name + "/target/scala-2.9.2/api" + source = "../" + project_name + "/target/scala-2.9.3/api" dest = "api/" + project_name puts "echo making directory " + dest @@ -27,9 +28,8 @@ if ENV['SKIP_SCALADOC'] != '1' puts "cp -r " + source + "/. " + dest cp_r(source + "/.", dest) end -end -if ENV['SKIP_EPYDOC'] != '1' + # Build Epydoc for Python puts "Moving to python directory and building epydoc." cd("../python") puts `epydoc --config epydoc.conf` diff --git a/docs/building-with-maven.md b/docs/building-with-maven.md new file mode 100644 index 0000000000..04cd79d039 --- /dev/null +++ b/docs/building-with-maven.md @@ -0,0 +1,66 @@ +--- +layout: global +title: Building Spark with Maven +--- + +* This will become a table of contents (this text will be scraped). +{:toc} + +Building Spark using Maven Requires Maven 3 (the build process is tested with Maven 3.0.4) and Java 1.6 or newer. + +Building with Maven requires that a Hadoop profile be specified explicitly at the command line, there is no default. There are two profiles to choose from, one for building for Hadoop 1 or Hadoop 2. + +for Hadoop 1 (using 0.20.205.0) use: + + $ mvn -Phadoop1 clean install + + +for Hadoop 2 (using 2.0.0-mr1-cdh4.1.1) use: + + $ mvn -Phadoop2 clean install + +It uses the scala-maven-plugin which supports incremental and continuous compilation. E.g. + + $ mvn -Phadoop2 scala:cc + +…should run continuous compilation (i.e. wait for changes). However, this has not been tested extensively. + +## Spark Tests in Maven ## + +Tests are run by default via the scalatest-maven-plugin. With this you can do things like: + +Skip test execution (but not compilation): + + $ mvn -DskipTests -Phadoop2 clean install + +To run a specific test suite: + + $ mvn -Phadoop2 -Dsuites=spark.repl.ReplSuite test + + +## Setting up JVM Memory Usage Via Maven ## + +You might run into the following errors if you're using a vanilla installation of Maven: + + [INFO] Compiling 203 Scala sources and 9 Java sources to /Users/me/Development/spark/core/target/scala-{{site.SCALA_VERSION}}/classes... + [ERROR] PermGen space -> [Help 1] + + [INFO] Compiling 203 Scala sources and 9 Java sources to /Users/me/Development/spark/core/target/scala-{{site.SCALA_VERSION}}/classes... + [ERROR] Java heap space -> [Help 1] + +To fix these, you can do the following: + + export MAVEN_OPTS="-Xmx1024m -XX:MaxPermSize=128M" + + +## Using With IntelliJ IDEA ## + +This setup works fine in IntelliJ IDEA 11.1.4. After opening the project via the pom.xml file in the project root folder, you only need to activate either the hadoop1 or hadoop2 profile in the "Maven Properties" popout. We have not tried Eclipse/Scala IDE with this. + +## Building Spark Debian Packages ## + +It includes support for building a Debian package containing a 'fat-jar' which includes the repl, the examples and bagel. This can be created by specifying the deb profile: + + $ mvn -Phadoop2,deb clean install + +The debian package can then be found under repl/target. We added the short commit hash to the file name so that we can distinguish individual packages build for SNAPSHOT versions. diff --git a/docs/configuration.md b/docs/configuration.md index f1ca77aa78..2de512f896 100644 --- a/docs/configuration.md +++ b/docs/configuration.md @@ -134,6 +134,13 @@ Apart from these, the following properties are also available, and may be useful </td> </tr> <tr> + <td>spark.ui.port</td> + <td>(random)</td> + <td> + Port for your application's dashboard, which shows memory usage of each RDD. + </td> +</tr> +<tr> <td>spark.shuffle.compress</td> <td>true</td> <td> @@ -183,7 +190,7 @@ Apart from these, the following properties are also available, and may be useful </tr> <tr> <td>spark.broadcast.factory</td> - <td>spark.broadcast. HttpBroadcastFactory</td> + <td>spark.broadcast.HttpBroadcastFactory</td> <td> Which broadcast implementation to use. </td> @@ -244,15 +251,22 @@ Apart from these, the following properties are also available, and may be useful </td> </tr> <tr> - <td>spark.cleaner.delay</td> + <td>spark.cleaner.ttl</td> <td>(disable)</td> <td> - Duration (minutes) of how long Spark will remember any metadata (stages generated, tasks generated, etc.). + Duration (seconds) of how long Spark will remember any metadata (stages generated, tasks generated, etc.). Periodic cleanups will ensure that metadata older than this duration will be forgetten. This is useful for running Spark for many hours / days (for example, running 24/7 in case of Spark Streaming applications). Note that any RDD that persists in memory for more than this duration will be cleared as well. </td> </tr> +<tr> + <td>spark.streaming.blockInterval</td> + <td>200</td> + <td> + Duration (milliseconds) of how long to batch new objects coming from network receivers. + </td> +</tr> </table> diff --git a/docs/css/bootstrap.css b/docs/css/bootstrap.css index 0664207aee..b51ef7b42b 100644 --- a/docs/css/bootstrap.css +++ b/docs/css/bootstrap.css @@ -2719,16 +2719,16 @@ table .span24 { .dropdown-menu .active > a:hover { color: #ffffff; text-decoration: none; - background-color: #0088cc; - background-color: #0081c2; - background-image: linear-gradient(to bottom, #0088cc, #0077b3); - background-image: -moz-linear-gradient(top, #0088cc, #0077b3); - background-image: -webkit-gradient(linear, 0 0, 0 100%, from(#0088cc), to(#0077b3)); - background-image: -webkit-linear-gradient(top, #0088cc, #0077b3); - background-image: -o-linear-gradient(top, #0088cc, #0077b3); + background-color: #0098cc; + background-color: #0098cc; + background-image: linear-gradient(to bottom, #0098cc, #0087b3); + background-image: -moz-linear-gradient(top, #0098cc, #0087b3); + background-image: -webkit-gradient(linear, 0 0, 0 100%, from(#0098cc), to(#0087b3)); + background-image: -webkit-linear-gradient(top, #0098cc, #0087b3); + background-image: -o-linear-gradient(top, #0098cc, #0087b3); background-repeat: repeat-x; outline: 0; - filter: progid:dximagetransform.microsoft.gradient(startColorstr='#ff0088cc', endColorstr='#ff0077b3', GradientType=0); + filter: progid:dximagetransform.microsoft.gradient(startColorstr='#ff0098cc', endColorstr='#ff0087b3', GradientType=0); } .dropdown-menu .disabled > a, @@ -4025,18 +4025,18 @@ input[type="submit"].btn.btn-mini { min-height: 40px; padding-right: 20px; padding-left: 20px; - background-color: #fafafa; - background-image: -moz-linear-gradient(top, #ffffff, #f2f2f2); - background-image: -webkit-gradient(linear, 0 0, 0 100%, from(#ffffff), to(#f2f2f2)); - background-image: -webkit-linear-gradient(top, #ffffff, #f2f2f2); - background-image: -o-linear-gradient(top, #ffffff, #f2f2f2); - background-image: linear-gradient(to bottom, #ffffff, #f2f2f2); + background-color: #fadafa; + background-image: -moz-linear-gradient(top, #ffddff, #f2d2f2); + background-image: -webkit-gradient(linear, 0 0, 0 100%, from(#ffddff), to(#f2d2f2)); + background-image: -webkit-linear-gradient(top, #ffddff, #f2d2f2); + background-image: -o-linear-gradient(top, #ffddff, #f2d2f2); + background-image: linear-gradient(to bottom, #ffddff, #f2d2f2); background-repeat: repeat-x; border: 1px solid #d4d4d4; -webkit-border-radius: 4px; -moz-border-radius: 4px; border-radius: 4px; - filter: progid:dximagetransform.microsoft.gradient(startColorstr='#ffffffff', endColorstr='#fff2f2f2', GradientType=0); + filter: progid:dximagetransform.microsoft.gradient(startColorstr='#ffffddff', endColorstr='#fff2d2f2', GradientType=0); -webkit-box-shadow: 0 1px 4px rgba(0, 0, 0, 0.065); -moz-box-shadow: 0 1px 4px rgba(0, 0, 0, 0.065); box-shadow: 0 1px 4px rgba(0, 0, 0, 0.065); diff --git a/docs/css/bootstrap.min.css b/docs/css/bootstrap.min.css index 3119038a6b..eb48138e08 100644 --- a/docs/css/bootstrap.min.css +++ b/docs/css/bootstrap.min.css @@ -6,4 +6,4 @@ * http://www.apache.org/licenses/LICENSE-2.0 * * Designed and built with all the love in the world @twitter by @mdo and @fat. - */article,aside,details,figcaption,figure,footer,header,hgroup,nav,section{display:block}audio,canvas,video{display:inline-block;*display:inline;*zoom:1}audio:not([controls]){display:none}html{font-size:100%;-webkit-text-size-adjust:100%;-ms-text-size-adjust:100%}a:focus{outline:thin dotted #333;outline:5px auto -webkit-focus-ring-color;outline-offset:-2px}a:hover,a:active{outline:0}sub,sup{position:relative;font-size:75%;line-height:0;vertical-align:baseline}sup{top:-0.5em}sub{bottom:-0.25em}img{height:auto;max-width:100%;vertical-align:middle;border:0;-ms-interpolation-mode:bicubic}#map_canvas img{max-width:none}button,input,select,textarea{margin:0;font-size:100%;vertical-align:middle}button,input{*overflow:visible;line-height:normal}button::-moz-focus-inner,input::-moz-focus-inner{padding:0;border:0}button,input[type="button"],input[type="reset"],input[type="submit"]{cursor:pointer;-webkit-appearance:button}input[type="search"]{-webkit-box-sizing:content-box;-moz-box-sizing:content-box;box-sizing:content-box;-webkit-appearance:textfield}input[type="search"]::-webkit-search-decoration,input[type="search"]::-webkit-search-cancel-button{-webkit-appearance:none}textarea{overflow:auto;vertical-align:top}.clearfix{*zoom:1}.clearfix:before,.clearfix:after{display:table;line-height:0;content:""}.clearfix:after{clear:both}.hide-text{font:0/0 a;color:transparent;text-shadow:none;background-color:transparent;border:0}.input-block-level{display:block;width:100%;min-height:30px;-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box}body{margin:0;font-family:"Helvetica Neue",Helvetica,Arial,sans-serif;font-size:14px;line-height:20px;color:#333;background-color:#fff}a{color:#08c;text-decoration:none}a:hover{color:#005580;text-decoration:underline}.img-rounded{-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.img-polaroid{padding:4px;background-color:#fff;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.2);-webkit-box-shadow:0 1px 3px rgba(0,0,0,0.1);-moz-box-shadow:0 1px 3px rgba(0,0,0,0.1);box-shadow:0 1px 3px rgba(0,0,0,0.1)}.img-circle{-webkit-border-radius:500px;-moz-border-radius:500px;border-radius:500px}.row{margin-left:-20px;*zoom:1}.row:before,.row:after{display:table;line-height:0;content:""}.row:after{clear:both}[class*="span"]{float:left;margin-left:20px}.container,.navbar-static-top .container,.navbar-fixed-top .container,.navbar-fixed-bottom .container{width:940px}.span12{width:940px}.span11{width:860px}.span10{width:780px}.span9{width:700px}.span8{width:620px}.span7{width:540px}.span6{width:460px}.span5{width:380px}.span4{width:300px}.span3{width:220px}.span2{width:140px}.span1{width:60px}.offset12{margin-left:980px}.offset11{margin-left:900px}.offset10{margin-left:820px}.offset9{margin-left:740px}.offset8{margin-left:660px}.offset7{margin-left:580px}.offset6{margin-left:500px}.offset5{margin-left:420px}.offset4{margin-left:340px}.offset3{margin-left:260px}.offset2{margin-left:180px}.offset1{margin-left:100px}.row-fluid{width:100%;*zoom:1}.row-fluid:before,.row-fluid:after{display:table;line-height:0;content:""}.row-fluid:after{clear:both}.row-fluid [class*="span"]{display:block;float:left;width:100%;min-height:30px;margin-left:2.127659574468085%;*margin-left:2.074468085106383%;-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box}.row-fluid [class*="span"]:first-child{margin-left:0}.row-fluid .span12{width:100%;*width:99.94680851063829%}.row-fluid .span11{width:91.48936170212765%;*width:91.43617021276594%}.row-fluid .span10{width:82.97872340425532%;*width:82.92553191489361%}.row-fluid .span9{width:74.46808510638297%;*width:74.41489361702126%}.row-fluid .span8{width:65.95744680851064%;*width:65.90425531914893%}.row-fluid .span7{width:57.44680851063829%;*width:57.39361702127659%}.row-fluid .span6{width:48.93617021276595%;*width:48.88297872340425%}.row-fluid .span5{width:40.42553191489362%;*width:40.37234042553192%}.row-fluid .span4{width:31.914893617021278%;*width:31.861702127659576%}.row-fluid .span3{width:23.404255319148934%;*width:23.351063829787233%}.row-fluid .span2{width:14.893617021276595%;*width:14.840425531914894%}.row-fluid .span1{width:6.382978723404255%;*width:6.329787234042553%}.row-fluid .offset12{margin-left:104.25531914893617%;*margin-left:104.14893617021275%}.row-fluid .offset12:first-child{margin-left:102.12765957446808%;*margin-left:102.02127659574467%}.row-fluid .offset11{margin-left:95.74468085106382%;*margin-left:95.6382978723404%}.row-fluid .offset11:first-child{margin-left:93.61702127659574%;*margin-left:93.51063829787232%}.row-fluid .offset10{margin-left:87.23404255319149%;*margin-left:87.12765957446807%}.row-fluid .offset10:first-child{margin-left:85.1063829787234%;*margin-left:84.99999999999999%}.row-fluid .offset9{margin-left:78.72340425531914%;*margin-left:78.61702127659572%}.row-fluid .offset9:first-child{margin-left:76.59574468085106%;*margin-left:76.48936170212764%}.row-fluid .offset8{margin-left:70.2127659574468%;*margin-left:70.10638297872339%}.row-fluid .offset8:first-child{margin-left:68.08510638297872%;*margin-left:67.9787234042553%}.row-fluid .offset7{margin-left:61.70212765957446%;*margin-left:61.59574468085106%}.row-fluid .offset7:first-child{margin-left:59.574468085106375%;*margin-left:59.46808510638297%}.row-fluid .offset6{margin-left:53.191489361702125%;*margin-left:53.085106382978715%}.row-fluid .offset6:first-child{margin-left:51.063829787234035%;*margin-left:50.95744680851063%}.row-fluid .offset5{margin-left:44.68085106382979%;*margin-left:44.57446808510638%}.row-fluid .offset5:first-child{margin-left:42.5531914893617%;*margin-left:42.4468085106383%}.row-fluid .offset4{margin-left:36.170212765957444%;*margin-left:36.06382978723405%}.row-fluid .offset4:first-child{margin-left:34.04255319148936%;*margin-left:33.93617021276596%}.row-fluid .offset3{margin-left:27.659574468085104%;*margin-left:27.5531914893617%}.row-fluid .offset3:first-child{margin-left:25.53191489361702%;*margin-left:25.425531914893618%}.row-fluid .offset2{margin-left:19.148936170212764%;*margin-left:19.04255319148936%}.row-fluid .offset2:first-child{margin-left:17.02127659574468%;*margin-left:16.914893617021278%}.row-fluid .offset1{margin-left:10.638297872340425%;*margin-left:10.53191489361702%}.row-fluid .offset1:first-child{margin-left:8.51063829787234%;*margin-left:8.404255319148938%}[class*="span"].hide,.row-fluid [class*="span"].hide{display:none}[class*="span"].pull-right,.row-fluid [class*="span"].pull-right{float:right}.container{margin-right:auto;margin-left:auto;*zoom:1}.container:before,.container:after{display:table;line-height:0;content:""}.container:after{clear:both}.container-fluid{padding-right:20px;padding-left:20px;*zoom:1}.container-fluid:before,.container-fluid:after{display:table;line-height:0;content:""}.container-fluid:after{clear:both}p{margin:0 0 10px}.lead{margin-bottom:20px;font-size:20px;font-weight:200;line-height:30px}small{font-size:85%}strong{font-weight:bold}em{font-style:italic}cite{font-style:normal}.muted{color:#999}h1,h2,h3,h4,h5,h6{margin:10px 0;font-family:inherit;font-weight:bold;line-height:1;color:inherit;text-rendering:optimizelegibility}h1 small,h2 small,h3 small,h4 small,h5 small,h6 small{font-weight:normal;line-height:1;color:#999}h1{font-size:36px;line-height:40px}h2{font-size:30px;line-height:40px}h3{font-size:24px;line-height:40px}h4{font-size:18px;line-height:20px}h5{font-size:14px;line-height:20px}h6{font-size:12px;line-height:20px}h1 small{font-size:24px}h2 small{font-size:18px}h3 small{font-size:14px}h4 small{font-size:14px}.page-header{padding-bottom:9px;margin:20px 0 30px;border-bottom:1px solid #eee}ul,ol{padding:0;margin:0 0 10px 25px}ul ul,ul ol,ol ol,ol ul{margin-bottom:0}li{line-height:20px}ul.unstyled,ol.unstyled{margin-left:0;list-style:none}dl{margin-bottom:20px}dt,dd{line-height:20px}dt{font-weight:bold}dd{margin-left:10px}.dl-horizontal dt{float:left;width:120px;overflow:hidden;clear:left;text-align:right;text-overflow:ellipsis;white-space:nowrap}.dl-horizontal dd{margin-left:130px}hr{margin:20px 0;border:0;border-top:1px solid #eee;border-bottom:1px solid #fff}abbr[title]{cursor:help;border-bottom:1px dotted #999}abbr.initialism{font-size:90%;text-transform:uppercase}blockquote{padding:0 0 0 15px;margin:0 0 20px;border-left:5px solid #eee}blockquote p{margin-bottom:0;font-size:16px;font-weight:300;line-height:25px}blockquote small{display:block;line-height:20px;color:#999}blockquote small:before{content:'\2014 \00A0'}blockquote.pull-right{float:right;padding-right:15px;padding-left:0;border-right:5px solid #eee;border-left:0}blockquote.pull-right p,blockquote.pull-right small{text-align:right}blockquote.pull-right small:before{content:''}blockquote.pull-right small:after{content:'\00A0 \2014'}q:before,q:after,blockquote:before,blockquote:after{content:""}address{display:block;margin-bottom:20px;font-style:normal;line-height:20px}code,pre{padding:0 3px 2px;font-family:Monaco,Menlo,Consolas,"Courier New",monospace;font-size:12px;color:#333;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}code{padding:2px 4px;color:#d14;background-color:#f7f7f9;border:1px solid #e1e1e8}pre{display:block;padding:9.5px;margin:0 0 10px;font-size:13px;line-height:20px;word-break:break-all;word-wrap:break-word;white-space:pre;white-space:pre-wrap;background-color:#f5f5f5;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.15);-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}pre.prettyprint{margin-bottom:20px}pre code{padding:0;color:inherit;background-color:transparent;border:0}.pre-scrollable{max-height:340px;overflow-y:scroll}form{margin:0 0 20px}fieldset{padding:0;margin:0;border:0}legend{display:block;width:100%;padding:0;margin-bottom:20px;font-size:21px;line-height:40px;color:#333;border:0;border-bottom:1px solid #e5e5e5}legend small{font-size:15px;color:#999}label,input,button,select,textarea{font-size:14px;font-weight:normal;line-height:20px}input,button,select,textarea{font-family:"Helvetica Neue",Helvetica,Arial,sans-serif}label{display:block;margin-bottom:5px}select,textarea,input[type="text"],input[type="password"],input[type="datetime"],input[type="datetime-local"],input[type="date"],input[type="month"],input[type="time"],input[type="week"],input[type="number"],input[type="email"],input[type="url"],input[type="search"],input[type="tel"],input[type="color"],.uneditable-input{display:inline-block;height:20px;padding:4px 6px;margin-bottom:9px;font-size:14px;line-height:20px;color:#555;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}input,textarea{width:210px}textarea{height:auto}textarea,input[type="text"],input[type="password"],input[type="datetime"],input[type="datetime-local"],input[type="date"],input[type="month"],input[type="time"],input[type="week"],input[type="number"],input[type="email"],input[type="url"],input[type="search"],input[type="tel"],input[type="color"],.uneditable-input{background-color:#fff;border:1px solid #ccc;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-webkit-transition:border linear .2s,box-shadow linear .2s;-moz-transition:border linear .2s,box-shadow linear .2s;-o-transition:border linear .2s,box-shadow linear .2s;transition:border linear .2s,box-shadow linear .2s}textarea:focus,input[type="text"]:focus,input[type="password"]:focus,input[type="datetime"]:focus,input[type="datetime-local"]:focus,input[type="date"]:focus,input[type="month"]:focus,input[type="time"]:focus,input[type="week"]:focus,input[type="number"]:focus,input[type="email"]:focus,input[type="url"]:focus,input[type="search"]:focus,input[type="tel"]:focus,input[type="color"]:focus,.uneditable-input:focus{border-color:rgba(82,168,236,0.8);outline:0;outline:thin dotted \9;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 8px rgba(82,168,236,0.6);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 8px rgba(82,168,236,0.6);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 8px rgba(82,168,236,0.6)}input[type="radio"],input[type="checkbox"]{margin:4px 0 0;margin-top:1px \9;*margin-top:0;line-height:normal;cursor:pointer}input[type="file"],input[type="image"],input[type="submit"],input[type="reset"],input[type="button"],input[type="radio"],input[type="checkbox"]{width:auto}select,input[type="file"]{height:30px;*margin-top:4px;line-height:30px}select{width:220px;background-color:#fff;border:1px solid #bbb}select[multiple],select[size]{height:auto}select:focus,input[type="file"]:focus,input[type="radio"]:focus,input[type="checkbox"]:focus{outline:thin dotted #333;outline:5px auto -webkit-focus-ring-color;outline-offset:-2px}.uneditable-input,.uneditable-textarea{color:#999;cursor:not-allowed;background-color:#fcfcfc;border-color:#ccc;-webkit-box-shadow:inset 0 1px 2px rgba(0,0,0,0.025);-moz-box-shadow:inset 0 1px 2px rgba(0,0,0,0.025);box-shadow:inset 0 1px 2px rgba(0,0,0,0.025)}.uneditable-input{overflow:hidden;white-space:nowrap}.uneditable-textarea{width:auto;height:auto}input:-moz-placeholder,textarea:-moz-placeholder{color:#999}input:-ms-input-placeholder,textarea:-ms-input-placeholder{color:#999}input::-webkit-input-placeholder,textarea::-webkit-input-placeholder{color:#999}.radio,.checkbox{min-height:18px;padding-left:18px}.radio input[type="radio"],.checkbox input[type="checkbox"]{float:left;margin-left:-18px}.controls>.radio:first-child,.controls>.checkbox:first-child{padding-top:5px}.radio.inline,.checkbox.inline{display:inline-block;padding-top:5px;margin-bottom:0;vertical-align:middle}.radio.inline+.radio.inline,.checkbox.inline+.checkbox.inline{margin-left:10px}.input-mini{width:60px}.input-small{width:90px}.input-medium{width:150px}.input-large{width:210px}.input-xlarge{width:270px}.input-xxlarge{width:530px}input[class*="span"],select[class*="span"],textarea[class*="span"],.uneditable-input[class*="span"],.row-fluid input[class*="span"],.row-fluid select[class*="span"],.row-fluid textarea[class*="span"],.row-fluid .uneditable-input[class*="span"]{float:none;margin-left:0}.input-append input[class*="span"],.input-append .uneditable-input[class*="span"],.input-prepend input[class*="span"],.input-prepend .uneditable-input[class*="span"],.row-fluid input[class*="span"],.row-fluid select[class*="span"],.row-fluid textarea[class*="span"],.row-fluid .uneditable-input[class*="span"],.row-fluid .input-prepend [class*="span"],.row-fluid .input-append [class*="span"]{display:inline-block}input,textarea,.uneditable-input{margin-left:0}.controls-row [class*="span"]+[class*="span"]{margin-left:20px}input.span12,textarea.span12,.uneditable-input.span12{width:926px}input.span11,textarea.span11,.uneditable-input.span11{width:846px}input.span10,textarea.span10,.uneditable-input.span10{width:766px}input.span9,textarea.span9,.uneditable-input.span9{width:686px}input.span8,textarea.span8,.uneditable-input.span8{width:606px}input.span7,textarea.span7,.uneditable-input.span7{width:526px}input.span6,textarea.span6,.uneditable-input.span6{width:446px}input.span5,textarea.span5,.uneditable-input.span5{width:366px}input.span4,textarea.span4,.uneditable-input.span4{width:286px}input.span3,textarea.span3,.uneditable-input.span3{width:206px}input.span2,textarea.span2,.uneditable-input.span2{width:126px}input.span1,textarea.span1,.uneditable-input.span1{width:46px}.controls-row{*zoom:1}.controls-row:before,.controls-row:after{display:table;line-height:0;content:""}.controls-row:after{clear:both}.controls-row [class*="span"]{float:left}input[disabled],select[disabled],textarea[disabled],input[readonly],select[readonly],textarea[readonly]{cursor:not-allowed;background-color:#eee}input[type="radio"][disabled],input[type="checkbox"][disabled],input[type="radio"][readonly],input[type="checkbox"][readonly]{background-color:transparent}.control-group.warning>label,.control-group.warning .help-block,.control-group.warning .help-inline{color:#c09853}.control-group.warning .checkbox,.control-group.warning .radio,.control-group.warning input,.control-group.warning select,.control-group.warning textarea{color:#c09853;border-color:#c09853;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075)}.control-group.warning .checkbox:focus,.control-group.warning .radio:focus,.control-group.warning input:focus,.control-group.warning select:focus,.control-group.warning textarea:focus{border-color:#a47e3c;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #dbc59e;-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #dbc59e;box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #dbc59e}.control-group.warning .input-prepend .add-on,.control-group.warning .input-append .add-on{color:#c09853;background-color:#fcf8e3;border-color:#c09853}.control-group.error>label,.control-group.error .help-block,.control-group.error .help-inline{color:#b94a48}.control-group.error .checkbox,.control-group.error .radio,.control-group.error input,.control-group.error select,.control-group.error textarea{color:#b94a48;border-color:#b94a48;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075)}.control-group.error .checkbox:focus,.control-group.error .radio:focus,.control-group.error input:focus,.control-group.error select:focus,.control-group.error textarea:focus{border-color:#953b39;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #d59392;-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #d59392;box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #d59392}.control-group.error .input-prepend .add-on,.control-group.error .input-append .add-on{color:#b94a48;background-color:#f2dede;border-color:#b94a48}.control-group.success>label,.control-group.success .help-block,.control-group.success .help-inline{color:#468847}.control-group.success .checkbox,.control-group.success .radio,.control-group.success input,.control-group.success select,.control-group.success textarea{color:#468847;border-color:#468847;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075)}.control-group.success .checkbox:focus,.control-group.success .radio:focus,.control-group.success input:focus,.control-group.success select:focus,.control-group.success textarea:focus{border-color:#356635;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #7aba7b;-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #7aba7b;box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #7aba7b}.control-group.success .input-prepend .add-on,.control-group.success .input-append .add-on{color:#468847;background-color:#dff0d8;border-color:#468847}input:focus:required:invalid,textarea:focus:required:invalid,select:focus:required:invalid{color:#b94a48;border-color:#ee5f5b}input:focus:required:invalid:focus,textarea:focus:required:invalid:focus,select:focus:required:invalid:focus{border-color:#e9322d;-webkit-box-shadow:0 0 6px #f8b9b7;-moz-box-shadow:0 0 6px #f8b9b7;box-shadow:0 0 6px #f8b9b7}.form-actions{padding:19px 20px 20px;margin-top:20px;margin-bottom:20px;background-color:#f5f5f5;border-top:1px solid #e5e5e5;*zoom:1}.form-actions:before,.form-actions:after{display:table;line-height:0;content:""}.form-actions:after{clear:both}.help-block,.help-inline{color:#595959}.help-block{display:block;margin-bottom:10px}.help-inline{display:inline-block;*display:inline;padding-left:5px;vertical-align:middle;*zoom:1}.input-append,.input-prepend{margin-bottom:5px;font-size:0;white-space:nowrap}.input-append input,.input-prepend input,.input-append select,.input-prepend select,.input-append .uneditable-input,.input-prepend .uneditable-input{position:relative;margin-bottom:0;*margin-left:0;font-size:14px;vertical-align:top;-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}.input-append input:focus,.input-prepend input:focus,.input-append select:focus,.input-prepend select:focus,.input-append .uneditable-input:focus,.input-prepend .uneditable-input:focus{z-index:2}.input-append .add-on,.input-prepend .add-on{display:inline-block;width:auto;height:20px;min-width:16px;padding:4px 5px;font-size:14px;font-weight:normal;line-height:20px;text-align:center;text-shadow:0 1px 0 #fff;background-color:#eee;border:1px solid #ccc}.input-append .add-on,.input-prepend .add-on,.input-append .btn,.input-prepend .btn{margin-left:-1px;vertical-align:top;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.input-append .active,.input-prepend .active{background-color:#a9dba9;border-color:#46a546}.input-prepend .add-on,.input-prepend .btn{margin-right:-1px}.input-prepend .add-on:first-child,.input-prepend .btn:first-child{-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.input-append input,.input-append select,.input-append .uneditable-input{-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.input-append .add-on:last-child,.input-append .btn:last-child{-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}.input-prepend.input-append input,.input-prepend.input-append select,.input-prepend.input-append .uneditable-input{-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.input-prepend.input-append .add-on:first-child,.input-prepend.input-append .btn:first-child{margin-right:-1px;-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.input-prepend.input-append .add-on:last-child,.input-prepend.input-append .btn:last-child{margin-left:-1px;-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}input.search-query{padding-right:14px;padding-right:4px \9;padding-left:14px;padding-left:4px \9;margin-bottom:0;-webkit-border-radius:15px;-moz-border-radius:15px;border-radius:15px}.form-search .input-append .search-query,.form-search .input-prepend .search-query{-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.form-search .input-append .search-query{-webkit-border-radius:14px 0 0 14px;-moz-border-radius:14px 0 0 14px;border-radius:14px 0 0 14px}.form-search .input-append .btn{-webkit-border-radius:0 14px 14px 0;-moz-border-radius:0 14px 14px 0;border-radius:0 14px 14px 0}.form-search .input-prepend .search-query{-webkit-border-radius:0 14px 14px 0;-moz-border-radius:0 14px 14px 0;border-radius:0 14px 14px 0}.form-search .input-prepend .btn{-webkit-border-radius:14px 0 0 14px;-moz-border-radius:14px 0 0 14px;border-radius:14px 0 0 14px}.form-search input,.form-inline input,.form-horizontal input,.form-search textarea,.form-inline textarea,.form-horizontal textarea,.form-search select,.form-inline select,.form-horizontal select,.form-search .help-inline,.form-inline .help-inline,.form-horizontal .help-inline,.form-search .uneditable-input,.form-inline .uneditable-input,.form-horizontal .uneditable-input,.form-search .input-prepend,.form-inline .input-prepend,.form-horizontal .input-prepend,.form-search .input-append,.form-inline .input-append,.form-horizontal .input-append{display:inline-block;*display:inline;margin-bottom:0;vertical-align:middle;*zoom:1}.form-search .hide,.form-inline .hide,.form-horizontal .hide{display:none}.form-search label,.form-inline label,.form-search .btn-group,.form-inline .btn-group{display:inline-block}.form-search .input-append,.form-inline .input-append,.form-search .input-prepend,.form-inline .input-prepend{margin-bottom:0}.form-search .radio,.form-search .checkbox,.form-inline .radio,.form-inline .checkbox{padding-left:0;margin-bottom:0;vertical-align:middle}.form-search .radio input[type="radio"],.form-search .checkbox input[type="checkbox"],.form-inline .radio input[type="radio"],.form-inline .checkbox input[type="checkbox"]{float:left;margin-right:3px;margin-left:0}.control-group{margin-bottom:10px}legend+.control-group{margin-top:20px;-webkit-margin-top-collapse:separate}.form-horizontal .control-group{margin-bottom:20px;*zoom:1}.form-horizontal .control-group:before,.form-horizontal .control-group:after{display:table;line-height:0;content:""}.form-horizontal .control-group:after{clear:both}.form-horizontal .control-label{float:left;width:140px;padding-top:5px;text-align:right}.form-horizontal .controls{*display:inline-block;*padding-left:20px;margin-left:160px;*margin-left:0}.form-horizontal .controls:first-child{*padding-left:160px}.form-horizontal .help-block{margin-top:10px;margin-bottom:0}.form-horizontal .form-actions{padding-left:160px}table{max-width:100%;background-color:transparent;border-collapse:collapse;border-spacing:0}.table{width:100%;margin-bottom:20px}.table th,.table td{padding:8px;line-height:20px;text-align:left;vertical-align:top;border-top:1px solid #ddd}.table th{font-weight:bold}.table thead th{vertical-align:bottom}.table caption+thead tr:first-child th,.table caption+thead tr:first-child td,.table colgroup+thead tr:first-child th,.table colgroup+thead tr:first-child td,.table thead:first-child tr:first-child th,.table thead:first-child tr:first-child td{border-top:0}.table tbody+tbody{border-top:2px solid #ddd}.table-condensed th,.table-condensed td{padding:4px 5px}.table-bordered{border:1px solid #ddd;border-collapse:separate;*border-collapse:collapse;border-left:0;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.table-bordered th,.table-bordered td{border-left:1px solid #ddd}.table-bordered caption+thead tr:first-child th,.table-bordered caption+tbody tr:first-child th,.table-bordered caption+tbody tr:first-child td,.table-bordered colgroup+thead tr:first-child th,.table-bordered colgroup+tbody tr:first-child th,.table-bordered colgroup+tbody tr:first-child td,.table-bordered thead:first-child tr:first-child th,.table-bordered tbody:first-child tr:first-child th,.table-bordered tbody:first-child tr:first-child td{border-top:0}.table-bordered thead:first-child tr:first-child th:first-child,.table-bordered tbody:first-child tr:first-child td:first-child{-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-topleft:4px}.table-bordered thead:first-child tr:first-child th:last-child,.table-bordered tbody:first-child tr:first-child td:last-child{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-moz-border-radius-topright:4px}.table-bordered thead:last-child tr:last-child th:first-child,.table-bordered tbody:last-child tr:last-child td:first-child,.table-bordered tfoot:last-child tr:last-child td:first-child{-webkit-border-radius:0 0 0 4px;-moz-border-radius:0 0 0 4px;border-radius:0 0 0 4px;-webkit-border-bottom-left-radius:4px;border-bottom-left-radius:4px;-moz-border-radius-bottomleft:4px}.table-bordered thead:last-child tr:last-child th:last-child,.table-bordered tbody:last-child tr:last-child td:last-child,.table-bordered tfoot:last-child tr:last-child td:last-child{-webkit-border-bottom-right-radius:4px;border-bottom-right-radius:4px;-moz-border-radius-bottomright:4px}.table-bordered caption+thead tr:first-child th:first-child,.table-bordered caption+tbody tr:first-child td:first-child,.table-bordered colgroup+thead tr:first-child th:first-child,.table-bordered colgroup+tbody tr:first-child td:first-child{-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-topleft:4px}.table-bordered caption+thead tr:first-child th:last-child,.table-bordered caption+tbody tr:first-child td:last-child,.table-bordered colgroup+thead tr:first-child th:last-child,.table-bordered colgroup+tbody tr:first-child td:last-child{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-moz-border-right-topleft:4px}.table-striped tbody tr:nth-child(odd) td,.table-striped tbody tr:nth-child(odd) th{background-color:#f9f9f9}.table-hover tbody tr:hover td,.table-hover tbody tr:hover th{background-color:#f5f5f5}table [class*=span],.row-fluid table [class*=span]{display:table-cell;float:none;margin-left:0}table .span1{float:none;width:44px;margin-left:0}table .span2{float:none;width:124px;margin-left:0}table .span3{float:none;width:204px;margin-left:0}table .span4{float:none;width:284px;margin-left:0}table .span5{float:none;width:364px;margin-left:0}table .span6{float:none;width:444px;margin-left:0}table .span7{float:none;width:524px;margin-left:0}table .span8{float:none;width:604px;margin-left:0}table .span9{float:none;width:684px;margin-left:0}table .span10{float:none;width:764px;margin-left:0}table .span11{float:none;width:844px;margin-left:0}table .span12{float:none;width:924px;margin-left:0}table .span13{float:none;width:1004px;margin-left:0}table .span14{float:none;width:1084px;margin-left:0}table .span15{float:none;width:1164px;margin-left:0}table .span16{float:none;width:1244px;margin-left:0}table .span17{float:none;width:1324px;margin-left:0}table .span18{float:none;width:1404px;margin-left:0}table .span19{float:none;width:1484px;margin-left:0}table .span20{float:none;width:1564px;margin-left:0}table .span21{float:none;width:1644px;margin-left:0}table .span22{float:none;width:1724px;margin-left:0}table .span23{float:none;width:1804px;margin-left:0}table .span24{float:none;width:1884px;margin-left:0}.table tbody tr.success td{background-color:#dff0d8}.table tbody tr.error td{background-color:#f2dede}.table tbody tr.info td{background-color:#d9edf7}[class^="icon-"],[class*=" icon-"]{display:inline-block;width:14px;height:14px;margin-top:1px;*margin-right:.3em;line-height:14px;vertical-align:text-top;background-image:url("../img/glyphicons-halflings.png");background-position:14px 14px;background-repeat:no-repeat}.icon-white,.nav>.active>a>[class^="icon-"],.nav>.active>a>[class*=" icon-"],.dropdown-menu>li>a:hover>[class^="icon-"],.dropdown-menu>li>a:hover>[class*=" icon-"],.dropdown-menu>.active>a>[class^="icon-"],.dropdown-menu>.active>a>[class*=" icon-"]{background-image:url("../img/glyphicons-halflings-white.png")}.icon-glass{background-position:0 0}.icon-music{background-position:-24px 0}.icon-search{background-position:-48px 0}.icon-envelope{background-position:-72px 0}.icon-heart{background-position:-96px 0}.icon-star{background-position:-120px 0}.icon-star-empty{background-position:-144px 0}.icon-user{background-position:-168px 0}.icon-film{background-position:-192px 0}.icon-th-large{background-position:-216px 0}.icon-th{background-position:-240px 0}.icon-th-list{background-position:-264px 0}.icon-ok{background-position:-288px 0}.icon-remove{background-position:-312px 0}.icon-zoom-in{background-position:-336px 0}.icon-zoom-out{background-position:-360px 0}.icon-off{background-position:-384px 0}.icon-signal{background-position:-408px 0}.icon-cog{background-position:-432px 0}.icon-trash{background-position:-456px 0}.icon-home{background-position:0 -24px}.icon-file{background-position:-24px -24px}.icon-time{background-position:-48px -24px}.icon-road{background-position:-72px -24px}.icon-download-alt{background-position:-96px -24px}.icon-download{background-position:-120px -24px}.icon-upload{background-position:-144px -24px}.icon-inbox{background-position:-168px -24px}.icon-play-circle{background-position:-192px -24px}.icon-repeat{background-position:-216px -24px}.icon-refresh{background-position:-240px -24px}.icon-list-alt{background-position:-264px -24px}.icon-lock{background-position:-287px -24px}.icon-flag{background-position:-312px -24px}.icon-headphones{background-position:-336px -24px}.icon-volume-off{background-position:-360px -24px}.icon-volume-down{background-position:-384px -24px}.icon-volume-up{background-position:-408px -24px}.icon-qrcode{background-position:-432px -24px}.icon-barcode{background-position:-456px -24px}.icon-tag{background-position:0 -48px}.icon-tags{background-position:-25px -48px}.icon-book{background-position:-48px -48px}.icon-bookmark{background-position:-72px -48px}.icon-print{background-position:-96px -48px}.icon-camera{background-position:-120px -48px}.icon-font{background-position:-144px -48px}.icon-bold{background-position:-167px -48px}.icon-italic{background-position:-192px -48px}.icon-text-height{background-position:-216px -48px}.icon-text-width{background-position:-240px -48px}.icon-align-left{background-position:-264px -48px}.icon-align-center{background-position:-288px -48px}.icon-align-right{background-position:-312px -48px}.icon-align-justify{background-position:-336px -48px}.icon-list{background-position:-360px -48px}.icon-indent-left{background-position:-384px -48px}.icon-indent-right{background-position:-408px -48px}.icon-facetime-video{background-position:-432px -48px}.icon-picture{background-position:-456px -48px}.icon-pencil{background-position:0 -72px}.icon-map-marker{background-position:-24px -72px}.icon-adjust{background-position:-48px -72px}.icon-tint{background-position:-72px -72px}.icon-edit{background-position:-96px -72px}.icon-share{background-position:-120px -72px}.icon-check{background-position:-144px -72px}.icon-move{background-position:-168px -72px}.icon-step-backward{background-position:-192px -72px}.icon-fast-backward{background-position:-216px -72px}.icon-backward{background-position:-240px -72px}.icon-play{background-position:-264px -72px}.icon-pause{background-position:-288px -72px}.icon-stop{background-position:-312px -72px}.icon-forward{background-position:-336px -72px}.icon-fast-forward{background-position:-360px -72px}.icon-step-forward{background-position:-384px -72px}.icon-eject{background-position:-408px -72px}.icon-chevron-left{background-position:-432px -72px}.icon-chevron-right{background-position:-456px -72px}.icon-plus-sign{background-position:0 -96px}.icon-minus-sign{background-position:-24px -96px}.icon-remove-sign{background-position:-48px -96px}.icon-ok-sign{background-position:-72px -96px}.icon-question-sign{background-position:-96px -96px}.icon-info-sign{background-position:-120px -96px}.icon-screenshot{background-position:-144px -96px}.icon-remove-circle{background-position:-168px -96px}.icon-ok-circle{background-position:-192px -96px}.icon-ban-circle{background-position:-216px -96px}.icon-arrow-left{background-position:-240px -96px}.icon-arrow-right{background-position:-264px -96px}.icon-arrow-up{background-position:-289px -96px}.icon-arrow-down{background-position:-312px -96px}.icon-share-alt{background-position:-336px -96px}.icon-resize-full{background-position:-360px -96px}.icon-resize-small{background-position:-384px -96px}.icon-plus{background-position:-408px -96px}.icon-minus{background-position:-433px -96px}.icon-asterisk{background-position:-456px -96px}.icon-exclamation-sign{background-position:0 -120px}.icon-gift{background-position:-24px -120px}.icon-leaf{background-position:-48px -120px}.icon-fire{background-position:-72px -120px}.icon-eye-open{background-position:-96px -120px}.icon-eye-close{background-position:-120px -120px}.icon-warning-sign{background-position:-144px -120px}.icon-plane{background-position:-168px -120px}.icon-calendar{background-position:-192px -120px}.icon-random{width:16px;background-position:-216px -120px}.icon-comment{background-position:-240px -120px}.icon-magnet{background-position:-264px -120px}.icon-chevron-up{background-position:-288px -120px}.icon-chevron-down{background-position:-313px -119px}.icon-retweet{background-position:-336px -120px}.icon-shopping-cart{background-position:-360px -120px}.icon-folder-close{background-position:-384px -120px}.icon-folder-open{width:16px;background-position:-408px -120px}.icon-resize-vertical{background-position:-432px -119px}.icon-resize-horizontal{background-position:-456px -118px}.icon-hdd{background-position:0 -144px}.icon-bullhorn{background-position:-24px -144px}.icon-bell{background-position:-48px -144px}.icon-certificate{background-position:-72px -144px}.icon-thumbs-up{background-position:-96px -144px}.icon-thumbs-down{background-position:-120px -144px}.icon-hand-right{background-position:-144px -144px}.icon-hand-left{background-position:-168px -144px}.icon-hand-up{background-position:-192px -144px}.icon-hand-down{background-position:-216px -144px}.icon-circle-arrow-right{background-position:-240px -144px}.icon-circle-arrow-left{background-position:-264px -144px}.icon-circle-arrow-up{background-position:-288px -144px}.icon-circle-arrow-down{background-position:-312px -144px}.icon-globe{background-position:-336px -144px}.icon-wrench{background-position:-360px -144px}.icon-tasks{background-position:-384px -144px}.icon-filter{background-position:-408px -144px}.icon-briefcase{background-position:-432px -144px}.icon-fullscreen{background-position:-456px -144px}.dropup,.dropdown{position:relative}.dropdown-toggle{*margin-bottom:-3px}.dropdown-toggle:active,.open .dropdown-toggle{outline:0}.caret{display:inline-block;width:0;height:0;vertical-align:top;border-top:4px solid #000;border-right:4px solid transparent;border-left:4px solid transparent;content:""}.dropdown .caret{margin-top:8px;margin-left:2px}.dropdown-menu{position:absolute;top:100%;left:0;z-index:1000;display:none;float:left;min-width:160px;padding:5px 0;margin:2px 0 0;list-style:none;background-color:#fff;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.2);*border-right-width:2px;*border-bottom-width:2px;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px;-webkit-box-shadow:0 5px 10px rgba(0,0,0,0.2);-moz-box-shadow:0 5px 10px rgba(0,0,0,0.2);box-shadow:0 5px 10px rgba(0,0,0,0.2);-webkit-background-clip:padding-box;-moz-background-clip:padding;background-clip:padding-box}.dropdown-menu.pull-right{right:0;left:auto}.dropdown-menu .divider{*width:100%;height:1px;margin:9px 1px;*margin:-5px 0 5px;overflow:hidden;background-color:#e5e5e5;border-bottom:1px solid #fff}.dropdown-menu a{display:block;padding:3px 20px;clear:both;font-weight:normal;line-height:20px;color:#333;white-space:nowrap}.dropdown-menu li>a:hover,.dropdown-menu li>a:focus,.dropdown-submenu:hover>a{color:#fff;text-decoration:none;background-color:#08c;background-color:#0081c2;background-image:-moz-linear-gradient(top,#08c,#0077b3);background-image:-webkit-gradient(linear,0 0,0 100%,from(#08c),to(#0077b3));background-image:-webkit-linear-gradient(top,#08c,#0077b3);background-image:-o-linear-gradient(top,#08c,#0077b3);background-image:linear-gradient(to bottom,#08c,#0077b3);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff0088cc',endColorstr='#ff0077b3',GradientType=0)}.dropdown-menu .active>a,.dropdown-menu .active>a:hover{color:#fff;text-decoration:none;background-color:#08c;background-color:#0081c2;background-image:linear-gradient(to bottom,#08c,#0077b3);background-image:-moz-linear-gradient(top,#08c,#0077b3);background-image:-webkit-gradient(linear,0 0,0 100%,from(#08c),to(#0077b3));background-image:-webkit-linear-gradient(top,#08c,#0077b3);background-image:-o-linear-gradient(top,#08c,#0077b3);background-repeat:repeat-x;outline:0;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff0088cc',endColorstr='#ff0077b3',GradientType=0)}.dropdown-menu .disabled>a,.dropdown-menu .disabled>a:hover{color:#999}.dropdown-menu .disabled>a:hover{text-decoration:none;cursor:default;background-color:transparent}.open{*z-index:1000}.open>.dropdown-menu{display:block}.pull-right>.dropdown-menu{right:0;left:auto}.dropup .caret,.navbar-fixed-bottom .dropdown .caret{border-top:0;border-bottom:4px solid #000;content:"\2191"}.dropup .dropdown-menu,.navbar-fixed-bottom .dropdown .dropdown-menu{top:auto;bottom:100%;margin-bottom:1px}.dropdown-submenu{position:relative}.dropdown-submenu>.dropdown-menu{top:0;left:100%;margin-top:-6px;margin-left:-1px;-webkit-border-radius:0 6px 6px 6px;-moz-border-radius:0 6px 6px 6px;border-radius:0 6px 6px 6px}.dropdown-submenu:hover .dropdown-menu{display:block}.dropdown-submenu>a:after{display:block;float:right;width:0;height:0;margin-top:5px;margin-right:-10px;border-color:transparent;border-left-color:#ccc;border-style:solid;border-width:5px 0 5px 5px;content:" "}.dropdown-submenu:hover>a:after{border-left-color:#fff}.dropdown .dropdown-menu .nav-header{padding-right:20px;padding-left:20px}.typeahead{margin-top:2px;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.well{min-height:20px;padding:19px;margin-bottom:20px;background-color:#f5f5f5;border:1px solid #e3e3e3;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.05);box-shadow:inset 0 1px 1px rgba(0,0,0,0.05)}.well blockquote{border-color:#ddd;border-color:rgba(0,0,0,0.15)}.well-large{padding:24px;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.well-small{padding:9px;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}.fade{opacity:0;-webkit-transition:opacity .15s linear;-moz-transition:opacity .15s linear;-o-transition:opacity .15s linear;transition:opacity .15s linear}.fade.in{opacity:1}.collapse{position:relative;height:0;overflow:hidden;overflow:visible \9;-webkit-transition:height .35s ease;-moz-transition:height .35s ease;-o-transition:height .35s ease;transition:height .35s ease}.collapse.in{height:auto}.close{float:right;font-size:20px;font-weight:bold;line-height:20px;color:#000;text-shadow:0 1px 0 #fff;opacity:.2;filter:alpha(opacity=20)}.close:hover{color:#000;text-decoration:none;cursor:pointer;opacity:.4;filter:alpha(opacity=40)}button.close{padding:0;cursor:pointer;background:transparent;border:0;-webkit-appearance:none}.btn{display:inline-block;*display:inline;padding:4px 14px;margin-bottom:0;*margin-left:.3em;font-size:14px;line-height:20px;*line-height:20px;color:#333;text-align:center;text-shadow:0 1px 1px rgba(255,255,255,0.75);vertical-align:middle;cursor:pointer;background-color:#f5f5f5;*background-color:#e6e6e6;background-image:-webkit-gradient(linear,0 0,0 100%,from(#fff),to(#e6e6e6));background-image:-webkit-linear-gradient(top,#fff,#e6e6e6);background-image:-o-linear-gradient(top,#fff,#e6e6e6);background-image:linear-gradient(to bottom,#fff,#e6e6e6);background-image:-moz-linear-gradient(top,#fff,#e6e6e6);background-repeat:repeat-x;border:1px solid #bbb;*border:0;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);border-color:#e6e6e6 #e6e6e6 #bfbfbf;border-bottom-color:#a2a2a2;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffffffff',endColorstr='#ffe6e6e6',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false);*zoom:1;-webkit-box-shadow:inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05)}.btn:hover,.btn:active,.btn.active,.btn.disabled,.btn[disabled]{color:#333;background-color:#e6e6e6;*background-color:#d9d9d9}.btn:active,.btn.active{background-color:#ccc \9}.btn:first-child{*margin-left:0}.btn:hover{color:#333;text-decoration:none;background-color:#e6e6e6;*background-color:#d9d9d9;background-position:0 -15px;-webkit-transition:background-position .1s linear;-moz-transition:background-position .1s linear;-o-transition:background-position .1s linear;transition:background-position .1s linear}.btn:focus{outline:thin dotted #333;outline:5px auto -webkit-focus-ring-color;outline-offset:-2px}.btn.active,.btn:active{background-color:#e6e6e6;background-color:#d9d9d9 \9;background-image:none;outline:0;-webkit-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05)}.btn.disabled,.btn[disabled]{cursor:default;background-color:#e6e6e6;background-image:none;opacity:.65;filter:alpha(opacity=65);-webkit-box-shadow:none;-moz-box-shadow:none;box-shadow:none}.btn-large{padding:9px 14px;font-size:16px;line-height:normal;-webkit-border-radius:5px;-moz-border-radius:5px;border-radius:5px}.btn-large [class^="icon-"]{margin-top:2px}.btn-small{padding:3px 9px;font-size:12px;line-height:18px}.btn-small [class^="icon-"]{margin-top:0}.btn-mini{padding:2px 6px;font-size:11px;line-height:16px}.btn-block{display:block;width:100%;padding-right:0;padding-left:0;-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box}.btn-block+.btn-block{margin-top:5px}.btn-primary.active,.btn-warning.active,.btn-danger.active,.btn-success.active,.btn-info.active,.btn-inverse.active{color:rgba(255,255,255,0.75)}.btn{border-color:#c5c5c5;border-color:rgba(0,0,0,0.15) rgba(0,0,0,0.15) rgba(0,0,0,0.25)}.btn-primary{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#006dcc;*background-color:#04c;background-image:-webkit-gradient(linear,0 0,0 100%,from(#08c),to(#04c));background-image:-webkit-linear-gradient(top,#08c,#04c);background-image:-o-linear-gradient(top,#08c,#04c);background-image:linear-gradient(to bottom,#08c,#04c);background-image:-moz-linear-gradient(top,#08c,#04c);background-repeat:repeat-x;border-color:#04c #04c #002a80;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff0088cc',endColorstr='#ff0044cc',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-primary:hover,.btn-primary:active,.btn-primary.active,.btn-primary.disabled,.btn-primary[disabled]{color:#fff;background-color:#04c;*background-color:#003bb3}.btn-primary:active,.btn-primary.active{background-color:#039 \9}.btn-warning{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#faa732;*background-color:#f89406;background-image:-webkit-gradient(linear,0 0,0 100%,from(#fbb450),to(#f89406));background-image:-webkit-linear-gradient(top,#fbb450,#f89406);background-image:-o-linear-gradient(top,#fbb450,#f89406);background-image:linear-gradient(to bottom,#fbb450,#f89406);background-image:-moz-linear-gradient(top,#fbb450,#f89406);background-repeat:repeat-x;border-color:#f89406 #f89406 #ad6704;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fffbb450',endColorstr='#fff89406',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-warning:hover,.btn-warning:active,.btn-warning.active,.btn-warning.disabled,.btn-warning[disabled]{color:#fff;background-color:#f89406;*background-color:#df8505}.btn-warning:active,.btn-warning.active{background-color:#c67605 \9}.btn-danger{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#da4f49;*background-color:#bd362f;background-image:-webkit-gradient(linear,0 0,0 100%,from(#ee5f5b),to(#bd362f));background-image:-webkit-linear-gradient(top,#ee5f5b,#bd362f);background-image:-o-linear-gradient(top,#ee5f5b,#bd362f);background-image:linear-gradient(to bottom,#ee5f5b,#bd362f);background-image:-moz-linear-gradient(top,#ee5f5b,#bd362f);background-repeat:repeat-x;border-color:#bd362f #bd362f #802420;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffee5f5b',endColorstr='#ffbd362f',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-danger:hover,.btn-danger:active,.btn-danger.active,.btn-danger.disabled,.btn-danger[disabled]{color:#fff;background-color:#bd362f;*background-color:#a9302a}.btn-danger:active,.btn-danger.active{background-color:#942a25 \9}.btn-success{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#5bb75b;*background-color:#51a351;background-image:-webkit-gradient(linear,0 0,0 100%,from(#62c462),to(#51a351));background-image:-webkit-linear-gradient(top,#62c462,#51a351);background-image:-o-linear-gradient(top,#62c462,#51a351);background-image:linear-gradient(to bottom,#62c462,#51a351);background-image:-moz-linear-gradient(top,#62c462,#51a351);background-repeat:repeat-x;border-color:#51a351 #51a351 #387038;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff62c462',endColorstr='#ff51a351',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-success:hover,.btn-success:active,.btn-success.active,.btn-success.disabled,.btn-success[disabled]{color:#fff;background-color:#51a351;*background-color:#499249}.btn-success:active,.btn-success.active{background-color:#408140 \9}.btn-info{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#49afcd;*background-color:#2f96b4;background-image:-webkit-gradient(linear,0 0,0 100%,from(#5bc0de),to(#2f96b4));background-image:-webkit-linear-gradient(top,#5bc0de,#2f96b4);background-image:-o-linear-gradient(top,#5bc0de,#2f96b4);background-image:linear-gradient(to bottom,#5bc0de,#2f96b4);background-image:-moz-linear-gradient(top,#5bc0de,#2f96b4);background-repeat:repeat-x;border-color:#2f96b4 #2f96b4 #1f6377;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff5bc0de',endColorstr='#ff2f96b4',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-info:hover,.btn-info:active,.btn-info.active,.btn-info.disabled,.btn-info[disabled]{color:#fff;background-color:#2f96b4;*background-color:#2a85a0}.btn-info:active,.btn-info.active{background-color:#24748c \9}.btn-inverse{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#363636;*background-color:#222;background-image:-webkit-gradient(linear,0 0,0 100%,from(#444),to(#222));background-image:-webkit-linear-gradient(top,#444,#222);background-image:-o-linear-gradient(top,#444,#222);background-image:linear-gradient(to bottom,#444,#222);background-image:-moz-linear-gradient(top,#444,#222);background-repeat:repeat-x;border-color:#222 #222 #000;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff444444',endColorstr='#ff222222',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-inverse:hover,.btn-inverse:active,.btn-inverse.active,.btn-inverse.disabled,.btn-inverse[disabled]{color:#fff;background-color:#222;*background-color:#151515}.btn-inverse:active,.btn-inverse.active{background-color:#080808 \9}button.btn,input[type="submit"].btn{*padding-top:3px;*padding-bottom:3px}button.btn::-moz-focus-inner,input[type="submit"].btn::-moz-focus-inner{padding:0;border:0}button.btn.btn-large,input[type="submit"].btn.btn-large{*padding-top:7px;*padding-bottom:7px}button.btn.btn-small,input[type="submit"].btn.btn-small{*padding-top:3px;*padding-bottom:3px}button.btn.btn-mini,input[type="submit"].btn.btn-mini{*padding-top:1px;*padding-bottom:1px}.btn-link,.btn-link:active{background-color:transparent;background-image:none;-webkit-box-shadow:none;-moz-box-shadow:none;box-shadow:none}.btn-link{color:#08c;cursor:pointer;border-color:transparent;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.btn-link:hover{color:#005580;text-decoration:underline;background-color:transparent}.btn-group{position:relative;*margin-left:.3em;font-size:0;white-space:nowrap}.btn-group:first-child{*margin-left:0}.btn-group+.btn-group{margin-left:5px}.btn-toolbar{margin-top:10px;margin-bottom:10px;font-size:0}.btn-toolbar .btn-group{display:inline-block;*display:inline;*zoom:1}.btn-toolbar .btn+.btn,.btn-toolbar .btn-group+.btn,.btn-toolbar .btn+.btn-group{margin-left:5px}.btn-group>.btn{position:relative;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.btn-group>.btn+.btn{margin-left:-1px}.btn-group>.btn,.btn-group>.dropdown-menu{font-size:14px}.btn-group>.btn-mini{font-size:11px}.btn-group>.btn-small{font-size:12px}.btn-group>.btn-large{font-size:16px}.btn-group>.btn:first-child{margin-left:0;-webkit-border-bottom-left-radius:4px;border-bottom-left-radius:4px;-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-bottomleft:4px;-moz-border-radius-topleft:4px}.btn-group>.btn:last-child,.btn-group>.dropdown-toggle{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-webkit-border-bottom-right-radius:4px;border-bottom-right-radius:4px;-moz-border-radius-topright:4px;-moz-border-radius-bottomright:4px}.btn-group>.btn.large:first-child{margin-left:0;-webkit-border-bottom-left-radius:6px;border-bottom-left-radius:6px;-webkit-border-top-left-radius:6px;border-top-left-radius:6px;-moz-border-radius-bottomleft:6px;-moz-border-radius-topleft:6px}.btn-group>.btn.large:last-child,.btn-group>.large.dropdown-toggle{-webkit-border-top-right-radius:6px;border-top-right-radius:6px;-webkit-border-bottom-right-radius:6px;border-bottom-right-radius:6px;-moz-border-radius-topright:6px;-moz-border-radius-bottomright:6px}.btn-group>.btn:hover,.btn-group>.btn:focus,.btn-group>.btn:active,.btn-group>.btn.active{z-index:2}.btn-group .dropdown-toggle:active,.btn-group.open .dropdown-toggle{outline:0}.btn-group>.btn+.dropdown-toggle{*padding-top:5px;padding-right:8px;*padding-bottom:5px;padding-left:8px;-webkit-box-shadow:inset 1px 0 0 rgba(255,255,255,0.125),inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 1px 0 0 rgba(255,255,255,0.125),inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 1px 0 0 rgba(255,255,255,0.125),inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05)}.btn-group>.btn-mini+.dropdown-toggle{*padding-top:2px;padding-right:5px;*padding-bottom:2px;padding-left:5px}.btn-group>.btn-small+.dropdown-toggle{*padding-top:5px;*padding-bottom:4px}.btn-group>.btn-large+.dropdown-toggle{*padding-top:7px;padding-right:12px;*padding-bottom:7px;padding-left:12px}.btn-group.open .dropdown-toggle{background-image:none;-webkit-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05)}.btn-group.open .btn.dropdown-toggle{background-color:#e6e6e6}.btn-group.open .btn-primary.dropdown-toggle{background-color:#04c}.btn-group.open .btn-warning.dropdown-toggle{background-color:#f89406}.btn-group.open .btn-danger.dropdown-toggle{background-color:#bd362f}.btn-group.open .btn-success.dropdown-toggle{background-color:#51a351}.btn-group.open .btn-info.dropdown-toggle{background-color:#2f96b4}.btn-group.open .btn-inverse.dropdown-toggle{background-color:#222}.btn .caret{margin-top:8px;margin-left:0}.btn-mini .caret,.btn-small .caret,.btn-large .caret{margin-top:6px}.btn-large .caret{border-top-width:5px;border-right-width:5px;border-left-width:5px}.dropup .btn-large .caret{border-top:0;border-bottom:5px solid #000}.btn-primary .caret,.btn-warning .caret,.btn-danger .caret,.btn-info .caret,.btn-success .caret,.btn-inverse .caret{border-top-color:#fff;border-bottom-color:#fff}.btn-group-vertical{display:inline-block;*display:inline;*zoom:1}.btn-group-vertical .btn{display:block;float:none;width:100%;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.btn-group-vertical .btn+.btn{margin-top:-1px;margin-left:0}.btn-group-vertical .btn:first-child{-webkit-border-radius:4px 4px 0 0;-moz-border-radius:4px 4px 0 0;border-radius:4px 4px 0 0}.btn-group-vertical .btn:last-child{-webkit-border-radius:0 0 4px 4px;-moz-border-radius:0 0 4px 4px;border-radius:0 0 4px 4px}.btn-group-vertical .btn-large:first-child{-webkit-border-radius:6px 6px 0 0;-moz-border-radius:6px 6px 0 0;border-radius:6px 6px 0 0}.btn-group-vertical .btn-large:last-child{-webkit-border-radius:0 0 6px 6px;-moz-border-radius:0 0 6px 6px;border-radius:0 0 6px 6px}.alert{padding:8px 35px 8px 14px;margin-bottom:20px;color:#c09853;text-shadow:0 1px 0 rgba(255,255,255,0.5);background-color:#fcf8e3;border:1px solid #fbeed5;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.alert h4{margin:0}.alert .close{position:relative;top:-2px;right:-21px;line-height:20px}.alert-success{color:#468847;background-color:#dff0d8;border-color:#d6e9c6}.alert-danger,.alert-error{color:#b94a48;background-color:#f2dede;border-color:#eed3d7}.alert-info{color:#3a87ad;background-color:#d9edf7;border-color:#bce8f1}.alert-block{padding-top:14px;padding-bottom:14px}.alert-block>p,.alert-block>ul{margin-bottom:0}.alert-block p+p{margin-top:5px}.nav{margin-bottom:20px;margin-left:0;list-style:none}.nav>li>a{display:block}.nav>li>a:hover{text-decoration:none;background-color:#eee}.nav>.pull-right{float:right}.nav-header{display:block;padding:3px 15px;font-size:11px;font-weight:bold;line-height:20px;color:#999;text-shadow:0 1px 0 rgba(255,255,255,0.5);text-transform:uppercase}.nav li+.nav-header{margin-top:9px}.nav-list{padding-right:15px;padding-left:15px;margin-bottom:0}.nav-list>li>a,.nav-list .nav-header{margin-right:-15px;margin-left:-15px;text-shadow:0 1px 0 rgba(255,255,255,0.5)}.nav-list>li>a{padding:3px 15px}.nav-list>.active>a,.nav-list>.active>a:hover{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.2);background-color:#08c}.nav-list [class^="icon-"]{margin-right:2px}.nav-list .divider{*width:100%;height:1px;margin:9px 1px;*margin:-5px 0 5px;overflow:hidden;background-color:#e5e5e5;border-bottom:1px solid #fff}.nav-tabs,.nav-pills{*zoom:1}.nav-tabs:before,.nav-pills:before,.nav-tabs:after,.nav-pills:after{display:table;line-height:0;content:""}.nav-tabs:after,.nav-pills:after{clear:both}.nav-tabs>li,.nav-pills>li{float:left}.nav-tabs>li>a,.nav-pills>li>a{padding-right:12px;padding-left:12px;margin-right:2px;line-height:14px}.nav-tabs{border-bottom:1px solid #ddd}.nav-tabs>li{margin-bottom:-1px}.nav-tabs>li>a{padding-top:8px;padding-bottom:8px;line-height:20px;border:1px solid transparent;-webkit-border-radius:4px 4px 0 0;-moz-border-radius:4px 4px 0 0;border-radius:4px 4px 0 0}.nav-tabs>li>a:hover{border-color:#eee #eee #ddd}.nav-tabs>.active>a,.nav-tabs>.active>a:hover{color:#555;cursor:default;background-color:#fff;border:1px solid #ddd;border-bottom-color:transparent}.nav-pills>li>a{padding-top:8px;padding-bottom:8px;margin-top:2px;margin-bottom:2px;-webkit-border-radius:5px;-moz-border-radius:5px;border-radius:5px}.nav-pills>.active>a,.nav-pills>.active>a:hover{color:#fff;background-color:#08c}.nav-stacked>li{float:none}.nav-stacked>li>a{margin-right:0}.nav-tabs.nav-stacked{border-bottom:0}.nav-tabs.nav-stacked>li>a{border:1px solid #ddd;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.nav-tabs.nav-stacked>li:first-child>a{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-topright:4px;-moz-border-radius-topleft:4px}.nav-tabs.nav-stacked>li:last-child>a{-webkit-border-bottom-right-radius:4px;border-bottom-right-radius:4px;-webkit-border-bottom-left-radius:4px;border-bottom-left-radius:4px;-moz-border-radius-bottomright:4px;-moz-border-radius-bottomleft:4px}.nav-tabs.nav-stacked>li>a:hover{z-index:2;border-color:#ddd}.nav-pills.nav-stacked>li>a{margin-bottom:3px}.nav-pills.nav-stacked>li:last-child>a{margin-bottom:1px}.nav-tabs .dropdown-menu{-webkit-border-radius:0 0 6px 6px;-moz-border-radius:0 0 6px 6px;border-radius:0 0 6px 6px}.nav-pills .dropdown-menu{-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.nav .dropdown-toggle .caret{margin-top:6px;border-top-color:#08c;border-bottom-color:#08c}.nav .dropdown-toggle:hover .caret{border-top-color:#005580;border-bottom-color:#005580}.nav-tabs .dropdown-toggle .caret{margin-top:8px}.nav .active .dropdown-toggle .caret{border-top-color:#fff;border-bottom-color:#fff}.nav-tabs .active .dropdown-toggle .caret{border-top-color:#555;border-bottom-color:#555}.nav>.dropdown.active>a:hover{cursor:pointer}.nav-tabs .open .dropdown-toggle,.nav-pills .open .dropdown-toggle,.nav>li.dropdown.open.active>a:hover{color:#fff;background-color:#999;border-color:#999}.nav li.dropdown.open .caret,.nav li.dropdown.open.active .caret,.nav li.dropdown.open a:hover .caret{border-top-color:#fff;border-bottom-color:#fff;opacity:1;filter:alpha(opacity=100)}.tabs-stacked .open>a:hover{border-color:#999}.tabbable{*zoom:1}.tabbable:before,.tabbable:after{display:table;line-height:0;content:""}.tabbable:after{clear:both}.tab-content{overflow:auto}.tabs-below>.nav-tabs,.tabs-right>.nav-tabs,.tabs-left>.nav-tabs{border-bottom:0}.tab-content>.tab-pane,.pill-content>.pill-pane{display:none}.tab-content>.active,.pill-content>.active{display:block}.tabs-below>.nav-tabs{border-top:1px solid #ddd}.tabs-below>.nav-tabs>li{margin-top:-1px;margin-bottom:0}.tabs-below>.nav-tabs>li>a{-webkit-border-radius:0 0 4px 4px;-moz-border-radius:0 0 4px 4px;border-radius:0 0 4px 4px}.tabs-below>.nav-tabs>li>a:hover{border-top-color:#ddd;border-bottom-color:transparent}.tabs-below>.nav-tabs>.active>a,.tabs-below>.nav-tabs>.active>a:hover{border-color:transparent #ddd #ddd #ddd}.tabs-left>.nav-tabs>li,.tabs-right>.nav-tabs>li{float:none}.tabs-left>.nav-tabs>li>a,.tabs-right>.nav-tabs>li>a{min-width:74px;margin-right:0;margin-bottom:3px}.tabs-left>.nav-tabs{float:left;margin-right:19px;border-right:1px solid #ddd}.tabs-left>.nav-tabs>li>a{margin-right:-1px;-webkit-border-radius:4px 0 0 4px;-moz-border-radius:4px 0 0 4px;border-radius:4px 0 0 4px}.tabs-left>.nav-tabs>li>a:hover{border-color:#eee #ddd #eee #eee}.tabs-left>.nav-tabs .active>a,.tabs-left>.nav-tabs .active>a:hover{border-color:#ddd transparent #ddd #ddd;*border-right-color:#fff}.tabs-right>.nav-tabs{float:right;margin-left:19px;border-left:1px solid #ddd}.tabs-right>.nav-tabs>li>a{margin-left:-1px;-webkit-border-radius:0 4px 4px 0;-moz-border-radius:0 4px 4px 0;border-radius:0 4px 4px 0}.tabs-right>.nav-tabs>li>a:hover{border-color:#eee #eee #eee #ddd}.tabs-right>.nav-tabs .active>a,.tabs-right>.nav-tabs .active>a:hover{border-color:#ddd #ddd #ddd transparent;*border-left-color:#fff}.nav>.disabled>a{color:#999}.nav>.disabled>a:hover{text-decoration:none;cursor:default;background-color:transparent}.navbar{*position:relative;*z-index:2;margin-bottom:20px;overflow:visible;color:#555}.navbar-inner{min-height:40px;padding-right:20px;padding-left:20px;background-color:#fafafa;background-image:-moz-linear-gradient(top,#fff,#f2f2f2);background-image:-webkit-gradient(linear,0 0,0 100%,from(#fff),to(#f2f2f2));background-image:-webkit-linear-gradient(top,#fff,#f2f2f2);background-image:-o-linear-gradient(top,#fff,#f2f2f2);background-image:linear-gradient(to bottom,#fff,#f2f2f2);background-repeat:repeat-x;border:1px solid #d4d4d4;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffffffff',endColorstr='#fff2f2f2',GradientType=0);-webkit-box-shadow:0 1px 4px rgba(0,0,0,0.065);-moz-box-shadow:0 1px 4px rgba(0,0,0,0.065);box-shadow:0 1px 4px rgba(0,0,0,0.065)}.navbar .container{width:auto}.nav-collapse.collapse{height:auto}.navbar .brand{display:block;float:left;padding:10px 20px 10px;margin-left:-20px;font-size:20px;font-weight:200;color:#555;text-shadow:0 1px 0 #fff}.navbar .brand:hover{text-decoration:none}.navbar-text{margin-bottom:0;line-height:40px}.navbar-link{color:#555}.navbar-link:hover{color:#333}.navbar .divider-vertical{height:40px;margin:0 9px;border-right:1px solid #fff;border-left:1px solid #f2f2f2}.navbar .btn,.navbar .btn-group{margin-top:6px}.navbar .btn-group .btn{margin:0}.navbar-form{margin-bottom:0;*zoom:1}.navbar-form:before,.navbar-form:after{display:table;line-height:0;content:""}.navbar-form:after{clear:both}.navbar-form input,.navbar-form select,.navbar-form .radio,.navbar-form .checkbox{margin-top:5px}.navbar-form input,.navbar-form select,.navbar-form .btn{display:inline-block;margin-bottom:0}.navbar-form input[type="image"],.navbar-form input[type="checkbox"],.navbar-form input[type="radio"]{margin-top:3px}.navbar-form .input-append,.navbar-form .input-prepend{margin-top:6px;white-space:nowrap}.navbar-form .input-append input,.navbar-form .input-prepend input{margin-top:0}.navbar-search{position:relative;float:left;margin-top:5px;margin-bottom:0}.navbar-search .search-query{padding:4px 14px;margin-bottom:0;font-family:"Helvetica Neue",Helvetica,Arial,sans-serif;font-size:13px;font-weight:normal;line-height:1;-webkit-border-radius:15px;-moz-border-radius:15px;border-radius:15px}.navbar-static-top{position:static;width:100%;margin-bottom:0}.navbar-static-top .navbar-inner{-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.navbar-fixed-top,.navbar-fixed-bottom{position:fixed;right:0;left:0;z-index:1030;margin-bottom:0}.navbar-fixed-top .navbar-inner,.navbar-fixed-bottom .navbar-inner,.navbar-static-top .navbar-inner{border:0}.navbar-fixed-top .navbar-inner,.navbar-fixed-bottom .navbar-inner{padding-right:0;padding-left:0;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.navbar-static-top .container,.navbar-fixed-top .container,.navbar-fixed-bottom .container{width:940px}.navbar-fixed-top{top:0}.navbar-fixed-top .navbar-inner,.navbar-static-top .navbar-inner{-webkit-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.1),0 1px 10px rgba(0,0,0,0.1);-moz-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.1),0 1px 10px rgba(0,0,0,0.1);box-shadow:inset 0 -1px 0 rgba(0,0,0,0.1),0 1px 10px rgba(0,0,0,0.1)}.navbar-fixed-bottom{bottom:0}.navbar-fixed-bottom .navbar-inner{-webkit-box-shadow:inset 0 1px 0 rgba(0,0,0,0.1),0 -1px 10px rgba(0,0,0,0.1);-moz-box-shadow:inset 0 1px 0 rgba(0,0,0,0.1),0 -1px 10px rgba(0,0,0,0.1);box-shadow:inset 0 1px 0 rgba(0,0,0,0.1),0 -1px 10px rgba(0,0,0,0.1)}.navbar .nav{position:relative;left:0;display:block;float:left;margin:0 10px 0 0}.navbar .nav.pull-right{float:right}.navbar .nav>li{float:left}.navbar .nav>li>a{float:none;padding:10px 15px 10px;color:#555;text-decoration:none;text-shadow:0 1px 0 #fff}.navbar .nav .dropdown-toggle .caret{margin-top:8px}.navbar .nav>li>a:focus,.navbar .nav>li>a:hover{color:#333;text-decoration:none;background-color:transparent}.navbar .nav>.active>a,.navbar .nav>.active>a:hover,.navbar .nav>.active>a:focus{color:#555;text-decoration:none;background-color:#e5e5e5;-webkit-box-shadow:inset 0 3px 8px rgba(0,0,0,0.125);-moz-box-shadow:inset 0 3px 8px rgba(0,0,0,0.125);box-shadow:inset 0 3px 8px rgba(0,0,0,0.125)}.navbar .btn-navbar{display:none;float:right;padding:7px 10px;margin-right:5px;margin-left:5px;color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#ededed;*background-color:#e5e5e5;background-image:-webkit-gradient(linear,0 0,0 100%,from(#f2f2f2),to(#e5e5e5));background-image:-webkit-linear-gradient(top,#f2f2f2,#e5e5e5);background-image:-o-linear-gradient(top,#f2f2f2,#e5e5e5);background-image:linear-gradient(to bottom,#f2f2f2,#e5e5e5);background-image:-moz-linear-gradient(top,#f2f2f2,#e5e5e5);background-repeat:repeat-x;border-color:#e5e5e5 #e5e5e5 #bfbfbf;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fff2f2f2',endColorstr='#ffe5e5e5',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false);-webkit-box-shadow:inset 0 1px 0 rgba(255,255,255,0.1),0 1px 0 rgba(255,255,255,0.075);-moz-box-shadow:inset 0 1px 0 rgba(255,255,255,0.1),0 1px 0 rgba(255,255,255,0.075);box-shadow:inset 0 1px 0 rgba(255,255,255,0.1),0 1px 0 rgba(255,255,255,0.075)}.navbar .btn-navbar:hover,.navbar .btn-navbar:active,.navbar .btn-navbar.active,.navbar .btn-navbar.disabled,.navbar .btn-navbar[disabled]{color:#fff;background-color:#e5e5e5;*background-color:#d9d9d9}.navbar .btn-navbar:active,.navbar .btn-navbar.active{background-color:#ccc \9}.navbar .btn-navbar .icon-bar{display:block;width:18px;height:2px;background-color:#f5f5f5;-webkit-border-radius:1px;-moz-border-radius:1px;border-radius:1px;-webkit-box-shadow:0 1px 0 rgba(0,0,0,0.25);-moz-box-shadow:0 1px 0 rgba(0,0,0,0.25);box-shadow:0 1px 0 rgba(0,0,0,0.25)}.btn-navbar .icon-bar+.icon-bar{margin-top:3px}.navbar .nav>li>.dropdown-menu:before{position:absolute;top:-7px;left:9px;display:inline-block;border-right:7px solid transparent;border-bottom:7px solid #ccc;border-left:7px solid transparent;border-bottom-color:rgba(0,0,0,0.2);content:''}.navbar .nav>li>.dropdown-menu:after{position:absolute;top:-6px;left:10px;display:inline-block;border-right:6px solid transparent;border-bottom:6px solid #fff;border-left:6px solid transparent;content:''}.navbar-fixed-bottom .nav>li>.dropdown-menu:before{top:auto;bottom:-7px;border-top:7px solid #ccc;border-bottom:0;border-top-color:rgba(0,0,0,0.2)}.navbar-fixed-bottom .nav>li>.dropdown-menu:after{top:auto;bottom:-6px;border-top:6px solid #fff;border-bottom:0}.navbar .nav li.dropdown.open>.dropdown-toggle,.navbar .nav li.dropdown.active>.dropdown-toggle,.navbar .nav li.dropdown.open.active>.dropdown-toggle{color:#555;background-color:#e5e5e5}.navbar .nav li.dropdown>.dropdown-toggle .caret{border-top-color:#555;border-bottom-color:#555}.navbar .nav li.dropdown.open>.dropdown-toggle .caret,.navbar .nav li.dropdown.active>.dropdown-toggle .caret,.navbar .nav li.dropdown.open.active>.dropdown-toggle .caret{border-top-color:#555;border-bottom-color:#555}.navbar .pull-right>li>.dropdown-menu,.navbar .nav>li>.dropdown-menu.pull-right{right:0;left:auto}.navbar .pull-right>li>.dropdown-menu:before,.navbar .nav>li>.dropdown-menu.pull-right:before{right:12px;left:auto}.navbar .pull-right>li>.dropdown-menu:after,.navbar .nav>li>.dropdown-menu.pull-right:after{right:13px;left:auto}.navbar .pull-right>li>.dropdown-menu .dropdown-menu,.navbar .nav>li>.dropdown-menu.pull-right .dropdown-menu{right:100%;left:auto;margin-right:-1px;margin-left:0;-webkit-border-radius:6px 0 6px 6px;-moz-border-radius:6px 0 6px 6px;border-radius:6px 0 6px 6px}.navbar-inverse{color:#999}.navbar-inverse .navbar-inner{background-color:#1b1b1b;background-image:-moz-linear-gradient(top,#222,#111);background-image:-webkit-gradient(linear,0 0,0 100%,from(#222),to(#111));background-image:-webkit-linear-gradient(top,#222,#111);background-image:-o-linear-gradient(top,#222,#111);background-image:linear-gradient(to bottom,#222,#111);background-repeat:repeat-x;border-color:#252525;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff222222',endColorstr='#ff111111',GradientType=0)}.navbar-inverse .brand,.navbar-inverse .nav>li>a{color:#999;text-shadow:0 -1px 0 rgba(0,0,0,0.25)}.navbar-inverse .brand:hover,.navbar-inverse .nav>li>a:hover{color:#fff}.navbar-inverse .nav>li>a:focus,.navbar-inverse .nav>li>a:hover{color:#fff;background-color:transparent}.navbar-inverse .nav .active>a,.navbar-inverse .nav .active>a:hover,.navbar-inverse .nav .active>a:focus{color:#fff;background-color:#111}.navbar-inverse .navbar-link{color:#999}.navbar-inverse .navbar-link:hover{color:#fff}.navbar-inverse .divider-vertical{border-right-color:#222;border-left-color:#111}.navbar-inverse .nav li.dropdown.open>.dropdown-toggle,.navbar-inverse .nav li.dropdown.active>.dropdown-toggle,.navbar-inverse .nav li.dropdown.open.active>.dropdown-toggle{color:#fff;background-color:#111}.navbar-inverse .nav li.dropdown>.dropdown-toggle .caret{border-top-color:#999;border-bottom-color:#999}.navbar-inverse .nav li.dropdown.open>.dropdown-toggle .caret,.navbar-inverse .nav li.dropdown.active>.dropdown-toggle .caret,.navbar-inverse .nav li.dropdown.open.active>.dropdown-toggle .caret{border-top-color:#fff;border-bottom-color:#fff}.navbar-inverse .navbar-search .search-query{color:#fff;background-color:#515151;border-color:#111;-webkit-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1),0 1px 0 rgba(255,255,255,0.15);-moz-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1),0 1px 0 rgba(255,255,255,0.15);box-shadow:inset 0 1px 2px rgba(0,0,0,0.1),0 1px 0 rgba(255,255,255,0.15);-webkit-transition:none;-moz-transition:none;-o-transition:none;transition:none}.navbar-inverse .navbar-search .search-query:-moz-placeholder{color:#ccc}.navbar-inverse .navbar-search .search-query:-ms-input-placeholder{color:#ccc}.navbar-inverse .navbar-search .search-query::-webkit-input-placeholder{color:#ccc}.navbar-inverse .navbar-search .search-query:focus,.navbar-inverse .navbar-search .search-query.focused{padding:5px 15px;color:#333;text-shadow:0 1px 0 #fff;background-color:#fff;border:0;outline:0;-webkit-box-shadow:0 0 3px rgba(0,0,0,0.15);-moz-box-shadow:0 0 3px rgba(0,0,0,0.15);box-shadow:0 0 3px rgba(0,0,0,0.15)}.navbar-inverse .btn-navbar{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#0e0e0e;*background-color:#040404;background-image:-webkit-gradient(linear,0 0,0 100%,from(#151515),to(#040404));background-image:-webkit-linear-gradient(top,#151515,#040404);background-image:-o-linear-gradient(top,#151515,#040404);background-image:linear-gradient(to bottom,#151515,#040404);background-image:-moz-linear-gradient(top,#151515,#040404);background-repeat:repeat-x;border-color:#040404 #040404 #000;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff151515',endColorstr='#ff040404',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.navbar-inverse .btn-navbar:hover,.navbar-inverse .btn-navbar:active,.navbar-inverse .btn-navbar.active,.navbar-inverse .btn-navbar.disabled,.navbar-inverse .btn-navbar[disabled]{color:#fff;background-color:#040404;*background-color:#000}.navbar-inverse .btn-navbar:active,.navbar-inverse .btn-navbar.active{background-color:#000 \9}.breadcrumb{padding:8px 15px;margin:0 0 20px;list-style:none;background-color:#f5f5f5;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.breadcrumb li{display:inline-block;*display:inline;text-shadow:0 1px 0 #fff;*zoom:1}.breadcrumb .divider{padding:0 5px;color:#ccc}.breadcrumb .active{color:#999}.pagination{height:40px;margin:20px 0}.pagination ul{display:inline-block;*display:inline;margin-bottom:0;margin-left:0;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px;*zoom:1;-webkit-box-shadow:0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:0 1px 2px rgba(0,0,0,0.05);box-shadow:0 1px 2px rgba(0,0,0,0.05)}.pagination li{display:inline}.pagination a,.pagination span{float:left;padding:0 14px;line-height:38px;text-decoration:none;background-color:#fff;border:1px solid #ddd;border-left-width:0}.pagination a:hover,.pagination .active a,.pagination .active span{background-color:#f5f5f5}.pagination .active a,.pagination .active span{color:#999;cursor:default}.pagination .disabled span,.pagination .disabled a,.pagination .disabled a:hover{color:#999;cursor:default;background-color:transparent}.pagination li:first-child a,.pagination li:first-child span{border-left-width:1px;-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.pagination li:last-child a,.pagination li:last-child span{-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}.pagination-centered{text-align:center}.pagination-right{text-align:right}.pager{margin:20px 0;text-align:center;list-style:none;*zoom:1}.pager:before,.pager:after{display:table;line-height:0;content:""}.pager:after{clear:both}.pager li{display:inline}.pager a{display:inline-block;padding:5px 14px;background-color:#fff;border:1px solid #ddd;-webkit-border-radius:15px;-moz-border-radius:15px;border-radius:15px}.pager a:hover{text-decoration:none;background-color:#f5f5f5}.pager .next a{float:right}.pager .previous a{float:left}.pager .disabled a,.pager .disabled a:hover{color:#999;cursor:default;background-color:#fff}.modal-open .dropdown-menu{z-index:2050}.modal-open .dropdown.open{*z-index:2050}.modal-open .popover{z-index:2060}.modal-open .tooltip{z-index:2080}.modal-backdrop{position:fixed;top:0;right:0;bottom:0;left:0;z-index:1040;background-color:#000}.modal-backdrop.fade{opacity:0}.modal-backdrop,.modal-backdrop.fade.in{opacity:.8;filter:alpha(opacity=80)}.modal{position:fixed;top:50%;left:50%;z-index:1050;width:560px;margin:-250px 0 0 -280px;overflow:auto;background-color:#fff;border:1px solid #999;border:1px solid rgba(0,0,0,0.3);*border:1px solid #999;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px;-webkit-box-shadow:0 3px 7px rgba(0,0,0,0.3);-moz-box-shadow:0 3px 7px rgba(0,0,0,0.3);box-shadow:0 3px 7px rgba(0,0,0,0.3);-webkit-background-clip:padding-box;-moz-background-clip:padding-box;background-clip:padding-box}.modal.fade{top:-25%;-webkit-transition:opacity .3s linear,top .3s ease-out;-moz-transition:opacity .3s linear,top .3s ease-out;-o-transition:opacity .3s linear,top .3s ease-out;transition:opacity .3s linear,top .3s ease-out}.modal.fade.in{top:50%}.modal-header{padding:9px 15px;border-bottom:1px solid #eee}.modal-header .close{margin-top:2px}.modal-header h3{margin:0;line-height:30px}.modal-body{max-height:400px;padding:15px;overflow-y:auto}.modal-form{margin-bottom:0}.modal-footer{padding:14px 15px 15px;margin-bottom:0;text-align:right;background-color:#f5f5f5;border-top:1px solid #ddd;-webkit-border-radius:0 0 6px 6px;-moz-border-radius:0 0 6px 6px;border-radius:0 0 6px 6px;*zoom:1;-webkit-box-shadow:inset 0 1px 0 #fff;-moz-box-shadow:inset 0 1px 0 #fff;box-shadow:inset 0 1px 0 #fff}.modal-footer:before,.modal-footer:after{display:table;line-height:0;content:""}.modal-footer:after{clear:both}.modal-footer .btn+.btn{margin-bottom:0;margin-left:5px}.modal-footer .btn-group .btn+.btn{margin-left:-1px}.tooltip{position:absolute;z-index:1030;display:block;padding:5px;font-size:11px;opacity:0;filter:alpha(opacity=0);visibility:visible}.tooltip.in{opacity:.8;filter:alpha(opacity=80)}.tooltip.top{margin-top:-3px}.tooltip.right{margin-left:3px}.tooltip.bottom{margin-top:3px}.tooltip.left{margin-left:-3px}.tooltip-inner{max-width:200px;padding:3px 8px;color:#fff;text-align:center;text-decoration:none;background-color:#000;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.tooltip-arrow{position:absolute;width:0;height:0;border-color:transparent;border-style:solid}.tooltip.top .tooltip-arrow{bottom:0;left:50%;margin-left:-5px;border-top-color:#000;border-width:5px 5px 0}.tooltip.right .tooltip-arrow{top:50%;left:0;margin-top:-5px;border-right-color:#000;border-width:5px 5px 5px 0}.tooltip.left .tooltip-arrow{top:50%;right:0;margin-top:-5px;border-left-color:#000;border-width:5px 0 5px 5px}.tooltip.bottom .tooltip-arrow{top:0;left:50%;margin-left:-5px;border-bottom-color:#000;border-width:0 5px 5px}.popover{position:absolute;top:0;left:0;z-index:1010;display:none;width:236px;padding:1px;background-color:#fff;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.2);-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px;-webkit-box-shadow:0 5px 10px rgba(0,0,0,0.2);-moz-box-shadow:0 5px 10px rgba(0,0,0,0.2);box-shadow:0 5px 10px rgba(0,0,0,0.2);-webkit-background-clip:padding-box;-moz-background-clip:padding;background-clip:padding-box}.popover.top{margin-bottom:10px}.popover.right{margin-left:10px}.popover.bottom{margin-top:10px}.popover.left{margin-right:10px}.popover-title{padding:8px 14px;margin:0;font-size:14px;font-weight:normal;line-height:18px;background-color:#f7f7f7;border-bottom:1px solid #ebebeb;-webkit-border-radius:5px 5px 0 0;-moz-border-radius:5px 5px 0 0;border-radius:5px 5px 0 0}.popover-content{padding:9px 14px}.popover-content p,.popover-content ul,.popover-content ol{margin-bottom:0}.popover .arrow,.popover .arrow:after{position:absolute;display:inline-block;width:0;height:0;border-color:transparent;border-style:solid}.popover .arrow:after{z-index:-1;content:""}.popover.top .arrow{bottom:-10px;left:50%;margin-left:-10px;border-top-color:#fff;border-width:10px 10px 0}.popover.top .arrow:after{bottom:-1px;left:-11px;border-top-color:rgba(0,0,0,0.25);border-width:11px 11px 0}.popover.right .arrow{top:50%;left:-10px;margin-top:-10px;border-right-color:#fff;border-width:10px 10px 10px 0}.popover.right .arrow:after{bottom:-11px;left:-1px;border-right-color:rgba(0,0,0,0.25);border-width:11px 11px 11px 0}.popover.bottom .arrow{top:-10px;left:50%;margin-left:-10px;border-bottom-color:#fff;border-width:0 10px 10px}.popover.bottom .arrow:after{top:-1px;left:-11px;border-bottom-color:rgba(0,0,0,0.25);border-width:0 11px 11px}.popover.left .arrow{top:50%;right:-10px;margin-top:-10px;border-left-color:#fff;border-width:10px 0 10px 10px}.popover.left .arrow:after{right:-1px;bottom:-11px;border-left-color:rgba(0,0,0,0.25);border-width:11px 0 11px 11px}.thumbnails{margin-left:-20px;list-style:none;*zoom:1}.thumbnails:before,.thumbnails:after{display:table;line-height:0;content:""}.thumbnails:after{clear:both}.row-fluid .thumbnails{margin-left:0}.thumbnails>li{float:left;margin-bottom:20px;margin-left:20px}.thumbnail{display:block;padding:4px;line-height:20px;border:1px solid #ddd;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;-webkit-box-shadow:0 1px 3px rgba(0,0,0,0.055);-moz-box-shadow:0 1px 3px rgba(0,0,0,0.055);box-shadow:0 1px 3px rgba(0,0,0,0.055);-webkit-transition:all .2s ease-in-out;-moz-transition:all .2s ease-in-out;-o-transition:all .2s ease-in-out;transition:all .2s ease-in-out}a.thumbnail:hover{border-color:#08c;-webkit-box-shadow:0 1px 4px rgba(0,105,214,0.25);-moz-box-shadow:0 1px 4px rgba(0,105,214,0.25);box-shadow:0 1px 4px rgba(0,105,214,0.25)}.thumbnail>img{display:block;max-width:100%;margin-right:auto;margin-left:auto}.thumbnail .caption{padding:9px;color:#555}.label,.badge{font-size:11.844px;font-weight:bold;line-height:14px;color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);white-space:nowrap;vertical-align:baseline;background-color:#999}.label{padding:1px 4px 2px;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}.badge{padding:1px 9px 2px;-webkit-border-radius:9px;-moz-border-radius:9px;border-radius:9px}a.label:hover,a.badge:hover{color:#fff;text-decoration:none;cursor:pointer}.label-important,.badge-important{background-color:#b94a48}.label-important[href],.badge-important[href]{background-color:#953b39}.label-warning,.badge-warning{background-color:#f89406}.label-warning[href],.badge-warning[href]{background-color:#c67605}.label-success,.badge-success{background-color:#468847}.label-success[href],.badge-success[href]{background-color:#356635}.label-info,.badge-info{background-color:#3a87ad}.label-info[href],.badge-info[href]{background-color:#2d6987}.label-inverse,.badge-inverse{background-color:#333}.label-inverse[href],.badge-inverse[href]{background-color:#1a1a1a}.btn .label,.btn .badge{position:relative;top:-1px}.btn-mini .label,.btn-mini .badge{top:0}@-webkit-keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}@-moz-keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}@-ms-keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}@-o-keyframes progress-bar-stripes{from{background-position:0 0}to{background-position:40px 0}}@keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}.progress{height:20px;margin-bottom:20px;overflow:hidden;background-color:#f7f7f7;background-image:-moz-linear-gradient(top,#f5f5f5,#f9f9f9);background-image:-webkit-gradient(linear,0 0,0 100%,from(#f5f5f5),to(#f9f9f9));background-image:-webkit-linear-gradient(top,#f5f5f5,#f9f9f9);background-image:-o-linear-gradient(top,#f5f5f5,#f9f9f9);background-image:linear-gradient(to bottom,#f5f5f5,#f9f9f9);background-repeat:repeat-x;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fff5f5f5',endColorstr='#fff9f9f9',GradientType=0);-webkit-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1);-moz-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1);box-shadow:inset 0 1px 2px rgba(0,0,0,0.1)}.progress .bar{float:left;width:0;height:100%;font-size:12px;color:#fff;text-align:center;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#0e90d2;background-image:-moz-linear-gradient(top,#149bdf,#0480be);background-image:-webkit-gradient(linear,0 0,0 100%,from(#149bdf),to(#0480be));background-image:-webkit-linear-gradient(top,#149bdf,#0480be);background-image:-o-linear-gradient(top,#149bdf,#0480be);background-image:linear-gradient(to bottom,#149bdf,#0480be);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff149bdf',endColorstr='#ff0480be',GradientType=0);-webkit-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.15);-moz-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.15);box-shadow:inset 0 -1px 0 rgba(0,0,0,0.15);-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box;-webkit-transition:width .6s ease;-moz-transition:width .6s ease;-o-transition:width .6s ease;transition:width .6s ease}.progress .bar+.bar{-webkit-box-shadow:inset 1px 0 0 rgba(0,0,0,0.15),inset 0 -1px 0 rgba(0,0,0,0.15);-moz-box-shadow:inset 1px 0 0 rgba(0,0,0,0.15),inset 0 -1px 0 rgba(0,0,0,0.15);box-shadow:inset 1px 0 0 rgba(0,0,0,0.15),inset 0 -1px 0 rgba(0,0,0,0.15)}.progress-striped .bar{background-color:#149bdf;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);-webkit-background-size:40px 40px;-moz-background-size:40px 40px;-o-background-size:40px 40px;background-size:40px 40px}.progress.active .bar{-webkit-animation:progress-bar-stripes 2s linear infinite;-moz-animation:progress-bar-stripes 2s linear infinite;-ms-animation:progress-bar-stripes 2s linear infinite;-o-animation:progress-bar-stripes 2s linear infinite;animation:progress-bar-stripes 2s linear infinite}.progress-danger .bar,.progress .bar-danger{background-color:#dd514c;background-image:-moz-linear-gradient(top,#ee5f5b,#c43c35);background-image:-webkit-gradient(linear,0 0,0 100%,from(#ee5f5b),to(#c43c35));background-image:-webkit-linear-gradient(top,#ee5f5b,#c43c35);background-image:-o-linear-gradient(top,#ee5f5b,#c43c35);background-image:linear-gradient(to bottom,#ee5f5b,#c43c35);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffee5f5b',endColorstr='#ffc43c35',GradientType=0)}.progress-danger.progress-striped .bar,.progress-striped .bar-danger{background-color:#ee5f5b;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.progress-success .bar,.progress .bar-success{background-color:#5eb95e;background-image:-moz-linear-gradient(top,#62c462,#57a957);background-image:-webkit-gradient(linear,0 0,0 100%,from(#62c462),to(#57a957));background-image:-webkit-linear-gradient(top,#62c462,#57a957);background-image:-o-linear-gradient(top,#62c462,#57a957);background-image:linear-gradient(to bottom,#62c462,#57a957);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff62c462',endColorstr='#ff57a957',GradientType=0)}.progress-success.progress-striped .bar,.progress-striped .bar-success{background-color:#62c462;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.progress-info .bar,.progress .bar-info{background-color:#4bb1cf;background-image:-moz-linear-gradient(top,#5bc0de,#339bb9);background-image:-webkit-gradient(linear,0 0,0 100%,from(#5bc0de),to(#339bb9));background-image:-webkit-linear-gradient(top,#5bc0de,#339bb9);background-image:-o-linear-gradient(top,#5bc0de,#339bb9);background-image:linear-gradient(to bottom,#5bc0de,#339bb9);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff5bc0de',endColorstr='#ff339bb9',GradientType=0)}.progress-info.progress-striped .bar,.progress-striped .bar-info{background-color:#5bc0de;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.progress-warning .bar,.progress .bar-warning{background-color:#faa732;background-image:-moz-linear-gradient(top,#fbb450,#f89406);background-image:-webkit-gradient(linear,0 0,0 100%,from(#fbb450),to(#f89406));background-image:-webkit-linear-gradient(top,#fbb450,#f89406);background-image:-o-linear-gradient(top,#fbb450,#f89406);background-image:linear-gradient(to bottom,#fbb450,#f89406);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fffbb450',endColorstr='#fff89406',GradientType=0)}.progress-warning.progress-striped .bar,.progress-striped .bar-warning{background-color:#fbb450;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.accordion{margin-bottom:20px}.accordion-group{margin-bottom:2px;border:1px solid #e5e5e5;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.accordion-heading{border-bottom:0}.accordion-heading .accordion-toggle{display:block;padding:8px 15px}.accordion-toggle{cursor:pointer}.accordion-inner{padding:9px 15px;border-top:1px solid #e5e5e5}.carousel{position:relative;margin-bottom:20px;line-height:1}.carousel-inner{position:relative;width:100%;overflow:hidden}.carousel .item{position:relative;display:none;-webkit-transition:.6s ease-in-out left;-moz-transition:.6s ease-in-out left;-o-transition:.6s ease-in-out left;transition:.6s ease-in-out left}.carousel .item>img{display:block;line-height:1}.carousel .active,.carousel .next,.carousel .prev{display:block}.carousel .active{left:0}.carousel .next,.carousel .prev{position:absolute;top:0;width:100%}.carousel .next{left:100%}.carousel .prev{left:-100%}.carousel .next.left,.carousel .prev.right{left:0}.carousel .active.left{left:-100%}.carousel .active.right{left:100%}.carousel-control{position:absolute;top:40%;left:15px;width:40px;height:40px;margin-top:-20px;font-size:60px;font-weight:100;line-height:30px;color:#fff;text-align:center;background:#222;border:3px solid #fff;-webkit-border-radius:23px;-moz-border-radius:23px;border-radius:23px;opacity:.5;filter:alpha(opacity=50)}.carousel-control.right{right:15px;left:auto}.carousel-control:hover{color:#fff;text-decoration:none;opacity:.9;filter:alpha(opacity=90)}.carousel-caption{position:absolute;right:0;bottom:0;left:0;padding:15px;background:#333;background:rgba(0,0,0,0.75)}.carousel-caption h4,.carousel-caption p{line-height:20px;color:#fff}.carousel-caption h4{margin:0 0 5px}.carousel-caption p{margin-bottom:0}.hero-unit{padding:60px;margin-bottom:30px;background-color:#eee;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.hero-unit h1{margin-bottom:0;font-size:60px;line-height:1;letter-spacing:-1px;color:inherit}.hero-unit p{font-size:18px;font-weight:200;line-height:30px;color:inherit}.pull-right{float:right}.pull-left{float:left}.hide{display:none}.show{display:block}.invisible{visibility:hidden}.affix{position:fixed} + */article,aside,details,figcaption,figure,footer,header,hgroup,nav,section{display:block}audio,canvas,video{display:inline-block;*display:inline;*zoom:1}audio:not([controls]){display:none}html{font-size:100%;-webkit-text-size-adjust:100%;-ms-text-size-adjust:100%}a:focus{outline:thin dotted #333;outline:5px auto -webkit-focus-ring-color;outline-offset:-2px}a:hover,a:active{outline:0}sub,sup{position:relative;font-size:75%;line-height:0;vertical-align:baseline}sup{top:-0.5em}sub{bottom:-0.25em}img{height:auto;max-width:100%;vertical-align:middle;border:0;-ms-interpolation-mode:bicubic}#map_canvas img{max-width:none}button,input,select,textarea{margin:0;font-size:100%;vertical-align:middle}button,input{*overflow:visible;line-height:normal}button::-moz-focus-inner,input::-moz-focus-inner{padding:0;border:0}button,input[type="button"],input[type="reset"],input[type="submit"]{cursor:pointer;-webkit-appearance:button}input[type="search"]{-webkit-box-sizing:content-box;-moz-box-sizing:content-box;box-sizing:content-box;-webkit-appearance:textfield}input[type="search"]::-webkit-search-decoration,input[type="search"]::-webkit-search-cancel-button{-webkit-appearance:none}textarea{overflow:auto;vertical-align:top}.clearfix{*zoom:1}.clearfix:before,.clearfix:after{display:table;line-height:0;content:""}.clearfix:after{clear:both}.hide-text{font:0/0 a;color:transparent;text-shadow:none;background-color:transparent;border:0}.input-block-level{display:block;width:100%;min-height:30px;-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box}body{margin:0;font-family:"Helvetica Neue",Helvetica,Arial,sans-serif;font-size:14px;line-height:20px;color:#333;background-color:#fff}a{color:#08c;text-decoration:none}a:hover{color:#005580;text-decoration:underline}.img-rounded{-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.img-polaroid{padding:4px;background-color:#fff;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.2);-webkit-box-shadow:0 1px 3px rgba(0,0,0,0.1);-moz-box-shadow:0 1px 3px rgba(0,0,0,0.1);box-shadow:0 1px 3px rgba(0,0,0,0.1)}.img-circle{-webkit-border-radius:500px;-moz-border-radius:500px;border-radius:500px}.row{margin-left:-20px;*zoom:1}.row:before,.row:after{display:table;line-height:0;content:""}.row:after{clear:both}[class*="span"]{float:left;margin-left:20px}.container,.navbar-static-top .container,.navbar-fixed-top .container,.navbar-fixed-bottom .container{width:940px}.span12{width:940px}.span11{width:860px}.span10{width:780px}.span9{width:700px}.span8{width:620px}.span7{width:540px}.span6{width:460px}.span5{width:380px}.span4{width:300px}.span3{width:220px}.span2{width:140px}.span1{width:60px}.offset12{margin-left:980px}.offset11{margin-left:900px}.offset10{margin-left:820px}.offset9{margin-left:740px}.offset8{margin-left:660px}.offset7{margin-left:580px}.offset6{margin-left:500px}.offset5{margin-left:420px}.offset4{margin-left:340px}.offset3{margin-left:260px}.offset2{margin-left:180px}.offset1{margin-left:100px}.row-fluid{width:100%;*zoom:1}.row-fluid:before,.row-fluid:after{display:table;line-height:0;content:""}.row-fluid:after{clear:both}.row-fluid [class*="span"]{display:block;float:left;width:100%;min-height:30px;margin-left:2.127659574468085%;*margin-left:2.074468085106383%;-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box}.row-fluid [class*="span"]:first-child{margin-left:0}.row-fluid .span12{width:100%;*width:99.94680851063829%}.row-fluid .span11{width:91.48936170212765%;*width:91.43617021276594%}.row-fluid .span10{width:82.97872340425532%;*width:82.92553191489361%}.row-fluid .span9{width:74.46808510638297%;*width:74.41489361702126%}.row-fluid .span8{width:65.95744680851064%;*width:65.90425531914893%}.row-fluid .span7{width:57.44680851063829%;*width:57.39361702127659%}.row-fluid .span6{width:48.93617021276595%;*width:48.88297872340425%}.row-fluid .span5{width:40.42553191489362%;*width:40.37234042553192%}.row-fluid .span4{width:31.914893617021278%;*width:31.861702127659576%}.row-fluid .span3{width:23.404255319148934%;*width:23.351063829787233%}.row-fluid .span2{width:14.893617021276595%;*width:14.840425531914894%}.row-fluid .span1{width:6.382978723404255%;*width:6.329787234042553%}.row-fluid .offset12{margin-left:104.25531914893617%;*margin-left:104.14893617021275%}.row-fluid .offset12:first-child{margin-left:102.12765957446808%;*margin-left:102.02127659574467%}.row-fluid .offset11{margin-left:95.74468085106382%;*margin-left:95.6382978723404%}.row-fluid .offset11:first-child{margin-left:93.61702127659574%;*margin-left:93.51063829787232%}.row-fluid .offset10{margin-left:87.23404255319149%;*margin-left:87.12765957446807%}.row-fluid .offset10:first-child{margin-left:85.1063829787234%;*margin-left:84.99999999999999%}.row-fluid .offset9{margin-left:78.72340425531914%;*margin-left:78.61702127659572%}.row-fluid .offset9:first-child{margin-left:76.59574468085106%;*margin-left:76.48936170212764%}.row-fluid .offset8{margin-left:70.2127659574468%;*margin-left:70.10638297872339%}.row-fluid .offset8:first-child{margin-left:68.08510638297872%;*margin-left:67.9787234042553%}.row-fluid .offset7{margin-left:61.70212765957446%;*margin-left:61.59574468085106%}.row-fluid .offset7:first-child{margin-left:59.574468085106375%;*margin-left:59.46808510638297%}.row-fluid .offset6{margin-left:53.191489361702125%;*margin-left:53.085106382978715%}.row-fluid .offset6:first-child{margin-left:51.063829787234035%;*margin-left:50.95744680851063%}.row-fluid .offset5{margin-left:44.68085106382979%;*margin-left:44.57446808510638%}.row-fluid .offset5:first-child{margin-left:42.5531914893617%;*margin-left:42.4468085106383%}.row-fluid .offset4{margin-left:36.170212765957444%;*margin-left:36.06382978723405%}.row-fluid .offset4:first-child{margin-left:34.04255319148936%;*margin-left:33.93617021276596%}.row-fluid .offset3{margin-left:27.659574468085104%;*margin-left:27.5531914893617%}.row-fluid .offset3:first-child{margin-left:25.53191489361702%;*margin-left:25.425531914893618%}.row-fluid .offset2{margin-left:19.148936170212764%;*margin-left:19.04255319148936%}.row-fluid .offset2:first-child{margin-left:17.02127659574468%;*margin-left:16.914893617021278%}.row-fluid .offset1{margin-left:10.638297872340425%;*margin-left:10.53191489361702%}.row-fluid .offset1:first-child{margin-left:8.51063829787234%;*margin-left:8.404255319148938%}[class*="span"].hide,.row-fluid [class*="span"].hide{display:none}[class*="span"].pull-right,.row-fluid [class*="span"].pull-right{float:right}.container{margin-right:auto;margin-left:auto;*zoom:1}.container:before,.container:after{display:table;line-height:0;content:""}.container:after{clear:both}.container-fluid{padding-right:20px;padding-left:20px;*zoom:1}.container-fluid:before,.container-fluid:after{display:table;line-height:0;content:""}.container-fluid:after{clear:both}p{margin:0 0 10px}.lead{margin-bottom:20px;font-size:20px;font-weight:200;line-height:30px}small{font-size:85%}strong{font-weight:bold}em{font-style:italic}cite{font-style:normal}.muted{color:#999}h1,h2,h3,h4,h5,h6{margin:10px 0;font-family:inherit;font-weight:bold;line-height:1;color:inherit;text-rendering:optimizelegibility}h1 small,h2 small,h3 small,h4 small,h5 small,h6 small{font-weight:normal;line-height:1;color:#999}h1{font-size:36px;line-height:40px}h2{font-size:30px;line-height:40px}h3{font-size:24px;line-height:40px}h4{font-size:18px;line-height:20px}h5{font-size:14px;line-height:20px}h6{font-size:12px;line-height:20px}h1 small{font-size:24px}h2 small{font-size:18px}h3 small{font-size:14px}h4 small{font-size:14px}.page-header{padding-bottom:9px;margin:20px 0 30px;border-bottom:1px solid #eee}ul,ol{padding:0;margin:0 0 10px 25px}ul ul,ul ol,ol ol,ol ul{margin-bottom:0}li{line-height:20px}ul.unstyled,ol.unstyled{margin-left:0;list-style:none}dl{margin-bottom:20px}dt,dd{line-height:20px}dt{font-weight:bold}dd{margin-left:10px}.dl-horizontal dt{float:left;width:120px;overflow:hidden;clear:left;text-align:right;text-overflow:ellipsis;white-space:nowrap}.dl-horizontal dd{margin-left:130px}hr{margin:20px 0;border:0;border-top:1px solid #eee;border-bottom:1px solid #fff}abbr[title]{cursor:help;border-bottom:1px dotted #999}abbr.initialism{font-size:90%;text-transform:uppercase}blockquote{padding:0 0 0 15px;margin:0 0 20px;border-left:5px solid #eee}blockquote p{margin-bottom:0;font-size:16px;font-weight:300;line-height:25px}blockquote small{display:block;line-height:20px;color:#999}blockquote small:before{content:'\2014 \00A0'}blockquote.pull-right{float:right;padding-right:15px;padding-left:0;border-right:5px solid #eee;border-left:0}blockquote.pull-right p,blockquote.pull-right small{text-align:right}blockquote.pull-right small:before{content:''}blockquote.pull-right small:after{content:'\00A0 \2014'}q:before,q:after,blockquote:before,blockquote:after{content:""}address{display:block;margin-bottom:20px;font-style:normal;line-height:20px}code,pre{padding:0 3px 2px;font-family:Monaco,Menlo,Consolas,"Courier New",monospace;font-size:12px;color:#333;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}code{padding:2px 4px;color:#d14;background-color:#f7f7f9;border:1px solid #e1e1e8}pre{display:block;padding:9.5px;margin:0 0 10px;font-size:13px;line-height:20px;word-break:break-all;word-wrap:break-word;white-space:pre;white-space:pre-wrap;background-color:#f5f5f5;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.15);-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}pre.prettyprint{margin-bottom:20px}pre code{padding:0;color:inherit;background-color:transparent;border:0}.pre-scrollable{max-height:340px;overflow-y:scroll}form{margin:0 0 20px}fieldset{padding:0;margin:0;border:0}legend{display:block;width:100%;padding:0;margin-bottom:20px;font-size:21px;line-height:40px;color:#333;border:0;border-bottom:1px solid #e5e5e5}legend small{font-size:15px;color:#999}label,input,button,select,textarea{font-size:14px;font-weight:normal;line-height:20px}input,button,select,textarea{font-family:"Helvetica Neue",Helvetica,Arial,sans-serif}label{display:block;margin-bottom:5px}select,textarea,input[type="text"],input[type="password"],input[type="datetime"],input[type="datetime-local"],input[type="date"],input[type="month"],input[type="time"],input[type="week"],input[type="number"],input[type="email"],input[type="url"],input[type="search"],input[type="tel"],input[type="color"],.uneditable-input{display:inline-block;height:20px;padding:4px 6px;margin-bottom:9px;font-size:14px;line-height:20px;color:#555;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}input,textarea{width:210px}textarea{height:auto}textarea,input[type="text"],input[type="password"],input[type="datetime"],input[type="datetime-local"],input[type="date"],input[type="month"],input[type="time"],input[type="week"],input[type="number"],input[type="email"],input[type="url"],input[type="search"],input[type="tel"],input[type="color"],.uneditable-input{background-color:#fff;border:1px solid #ccc;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-webkit-transition:border linear .2s,box-shadow linear .2s;-moz-transition:border linear .2s,box-shadow linear .2s;-o-transition:border linear .2s,box-shadow linear .2s;transition:border linear .2s,box-shadow linear .2s}textarea:focus,input[type="text"]:focus,input[type="password"]:focus,input[type="datetime"]:focus,input[type="datetime-local"]:focus,input[type="date"]:focus,input[type="month"]:focus,input[type="time"]:focus,input[type="week"]:focus,input[type="number"]:focus,input[type="email"]:focus,input[type="url"]:focus,input[type="search"]:focus,input[type="tel"]:focus,input[type="color"]:focus,.uneditable-input:focus{border-color:rgba(82,168,236,0.8);outline:0;outline:thin dotted \9;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 8px rgba(82,168,236,0.6);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 8px rgba(82,168,236,0.6);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 8px rgba(82,168,236,0.6)}input[type="radio"],input[type="checkbox"]{margin:4px 0 0;margin-top:1px \9;*margin-top:0;line-height:normal;cursor:pointer}input[type="file"],input[type="image"],input[type="submit"],input[type="reset"],input[type="button"],input[type="radio"],input[type="checkbox"]{width:auto}select,input[type="file"]{height:30px;*margin-top:4px;line-height:30px}select{width:220px;background-color:#fff;border:1px solid #bbb}select[multiple],select[size]{height:auto}select:focus,input[type="file"]:focus,input[type="radio"]:focus,input[type="checkbox"]:focus{outline:thin dotted #333;outline:5px auto -webkit-focus-ring-color;outline-offset:-2px}.uneditable-input,.uneditable-textarea{color:#999;cursor:not-allowed;background-color:#fcfcfc;border-color:#ccc;-webkit-box-shadow:inset 0 1px 2px rgba(0,0,0,0.025);-moz-box-shadow:inset 0 1px 2px rgba(0,0,0,0.025);box-shadow:inset 0 1px 2px rgba(0,0,0,0.025)}.uneditable-input{overflow:hidden;white-space:nowrap}.uneditable-textarea{width:auto;height:auto}input:-moz-placeholder,textarea:-moz-placeholder{color:#999}input:-ms-input-placeholder,textarea:-ms-input-placeholder{color:#999}input::-webkit-input-placeholder,textarea::-webkit-input-placeholder{color:#999}.radio,.checkbox{min-height:18px;padding-left:18px}.radio input[type="radio"],.checkbox input[type="checkbox"]{float:left;margin-left:-18px}.controls>.radio:first-child,.controls>.checkbox:first-child{padding-top:5px}.radio.inline,.checkbox.inline{display:inline-block;padding-top:5px;margin-bottom:0;vertical-align:middle}.radio.inline+.radio.inline,.checkbox.inline+.checkbox.inline{margin-left:10px}.input-mini{width:60px}.input-small{width:90px}.input-medium{width:150px}.input-large{width:210px}.input-xlarge{width:270px}.input-xxlarge{width:530px}input[class*="span"],select[class*="span"],textarea[class*="span"],.uneditable-input[class*="span"],.row-fluid input[class*="span"],.row-fluid select[class*="span"],.row-fluid textarea[class*="span"],.row-fluid .uneditable-input[class*="span"]{float:none;margin-left:0}.input-append input[class*="span"],.input-append .uneditable-input[class*="span"],.input-prepend input[class*="span"],.input-prepend .uneditable-input[class*="span"],.row-fluid input[class*="span"],.row-fluid select[class*="span"],.row-fluid textarea[class*="span"],.row-fluid .uneditable-input[class*="span"],.row-fluid .input-prepend [class*="span"],.row-fluid .input-append [class*="span"]{display:inline-block}input,textarea,.uneditable-input{margin-left:0}.controls-row [class*="span"]+[class*="span"]{margin-left:20px}input.span12,textarea.span12,.uneditable-input.span12{width:926px}input.span11,textarea.span11,.uneditable-input.span11{width:846px}input.span10,textarea.span10,.uneditable-input.span10{width:766px}input.span9,textarea.span9,.uneditable-input.span9{width:686px}input.span8,textarea.span8,.uneditable-input.span8{width:606px}input.span7,textarea.span7,.uneditable-input.span7{width:526px}input.span6,textarea.span6,.uneditable-input.span6{width:446px}input.span5,textarea.span5,.uneditable-input.span5{width:366px}input.span4,textarea.span4,.uneditable-input.span4{width:286px}input.span3,textarea.span3,.uneditable-input.span3{width:206px}input.span2,textarea.span2,.uneditable-input.span2{width:126px}input.span1,textarea.span1,.uneditable-input.span1{width:46px}.controls-row{*zoom:1}.controls-row:before,.controls-row:after{display:table;line-height:0;content:""}.controls-row:after{clear:both}.controls-row [class*="span"]{float:left}input[disabled],select[disabled],textarea[disabled],input[readonly],select[readonly],textarea[readonly]{cursor:not-allowed;background-color:#eee}input[type="radio"][disabled],input[type="checkbox"][disabled],input[type="radio"][readonly],input[type="checkbox"][readonly]{background-color:transparent}.control-group.warning>label,.control-group.warning .help-block,.control-group.warning .help-inline{color:#c09853}.control-group.warning .checkbox,.control-group.warning .radio,.control-group.warning input,.control-group.warning select,.control-group.warning textarea{color:#c09853;border-color:#c09853;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075)}.control-group.warning .checkbox:focus,.control-group.warning .radio:focus,.control-group.warning input:focus,.control-group.warning select:focus,.control-group.warning textarea:focus{border-color:#a47e3c;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #dbc59e;-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #dbc59e;box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #dbc59e}.control-group.warning .input-prepend .add-on,.control-group.warning .input-append .add-on{color:#c09853;background-color:#fcf8e3;border-color:#c09853}.control-group.error>label,.control-group.error .help-block,.control-group.error .help-inline{color:#b94a48}.control-group.error .checkbox,.control-group.error .radio,.control-group.error input,.control-group.error select,.control-group.error textarea{color:#b94a48;border-color:#b94a48;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075)}.control-group.error .checkbox:focus,.control-group.error .radio:focus,.control-group.error input:focus,.control-group.error select:focus,.control-group.error textarea:focus{border-color:#953b39;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #d59392;-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #d59392;box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #d59392}.control-group.error .input-prepend .add-on,.control-group.error .input-append .add-on{color:#b94a48;background-color:#f2dede;border-color:#b94a48}.control-group.success>label,.control-group.success .help-block,.control-group.success .help-inline{color:#468847}.control-group.success .checkbox,.control-group.success .radio,.control-group.success input,.control-group.success select,.control-group.success textarea{color:#468847;border-color:#468847;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075);box-shadow:inset 0 1px 1px rgba(0,0,0,0.075)}.control-group.success .checkbox:focus,.control-group.success .radio:focus,.control-group.success input:focus,.control-group.success select:focus,.control-group.success textarea:focus{border-color:#356635;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #7aba7b;-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #7aba7b;box-shadow:inset 0 1px 1px rgba(0,0,0,0.075),0 0 6px #7aba7b}.control-group.success .input-prepend .add-on,.control-group.success .input-append .add-on{color:#468847;background-color:#dff0d8;border-color:#468847}input:focus:required:invalid,textarea:focus:required:invalid,select:focus:required:invalid{color:#b94a48;border-color:#ee5f5b}input:focus:required:invalid:focus,textarea:focus:required:invalid:focus,select:focus:required:invalid:focus{border-color:#e9322d;-webkit-box-shadow:0 0 6px #f8b9b7;-moz-box-shadow:0 0 6px #f8b9b7;box-shadow:0 0 6px #f8b9b7}.form-actions{padding:19px 20px 20px;margin-top:20px;margin-bottom:20px;background-color:#f5f5f5;border-top:1px solid #e5e5e5;*zoom:1}.form-actions:before,.form-actions:after{display:table;line-height:0;content:""}.form-actions:after{clear:both}.help-block,.help-inline{color:#595959}.help-block{display:block;margin-bottom:10px}.help-inline{display:inline-block;*display:inline;padding-left:5px;vertical-align:middle;*zoom:1}.input-append,.input-prepend{margin-bottom:5px;font-size:0;white-space:nowrap}.input-append input,.input-prepend input,.input-append select,.input-prepend select,.input-append .uneditable-input,.input-prepend .uneditable-input{position:relative;margin-bottom:0;*margin-left:0;font-size:14px;vertical-align:top;-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}.input-append input:focus,.input-prepend input:focus,.input-append select:focus,.input-prepend select:focus,.input-append .uneditable-input:focus,.input-prepend .uneditable-input:focus{z-index:2}.input-append .add-on,.input-prepend .add-on{display:inline-block;width:auto;height:20px;min-width:16px;padding:4px 5px;font-size:14px;font-weight:normal;line-height:20px;text-align:center;text-shadow:0 1px 0 #fff;background-color:#eee;border:1px solid #ccc}.input-append .add-on,.input-prepend .add-on,.input-append .btn,.input-prepend .btn{margin-left:-1px;vertical-align:top;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.input-append .active,.input-prepend .active{background-color:#a9dba9;border-color:#46a546}.input-prepend .add-on,.input-prepend .btn{margin-right:-1px}.input-prepend .add-on:first-child,.input-prepend .btn:first-child{-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.input-append input,.input-append select,.input-append .uneditable-input{-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.input-append .add-on:last-child,.input-append .btn:last-child{-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}.input-prepend.input-append input,.input-prepend.input-append select,.input-prepend.input-append .uneditable-input{-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.input-prepend.input-append .add-on:first-child,.input-prepend.input-append .btn:first-child{margin-right:-1px;-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.input-prepend.input-append .add-on:last-child,.input-prepend.input-append .btn:last-child{margin-left:-1px;-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}input.search-query{padding-right:14px;padding-right:4px \9;padding-left:14px;padding-left:4px \9;margin-bottom:0;-webkit-border-radius:15px;-moz-border-radius:15px;border-radius:15px}.form-search .input-append .search-query,.form-search .input-prepend .search-query{-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.form-search .input-append .search-query{-webkit-border-radius:14px 0 0 14px;-moz-border-radius:14px 0 0 14px;border-radius:14px 0 0 14px}.form-search .input-append .btn{-webkit-border-radius:0 14px 14px 0;-moz-border-radius:0 14px 14px 0;border-radius:0 14px 14px 0}.form-search .input-prepend .search-query{-webkit-border-radius:0 14px 14px 0;-moz-border-radius:0 14px 14px 0;border-radius:0 14px 14px 0}.form-search .input-prepend .btn{-webkit-border-radius:14px 0 0 14px;-moz-border-radius:14px 0 0 14px;border-radius:14px 0 0 14px}.form-search input,.form-inline input,.form-horizontal input,.form-search textarea,.form-inline textarea,.form-horizontal textarea,.form-search select,.form-inline select,.form-horizontal select,.form-search .help-inline,.form-inline .help-inline,.form-horizontal .help-inline,.form-search .uneditable-input,.form-inline .uneditable-input,.form-horizontal .uneditable-input,.form-search .input-prepend,.form-inline .input-prepend,.form-horizontal .input-prepend,.form-search .input-append,.form-inline .input-append,.form-horizontal .input-append{display:inline-block;*display:inline;margin-bottom:0;vertical-align:middle;*zoom:1}.form-search .hide,.form-inline .hide,.form-horizontal .hide{display:none}.form-search label,.form-inline label,.form-search .btn-group,.form-inline .btn-group{display:inline-block}.form-search .input-append,.form-inline .input-append,.form-search .input-prepend,.form-inline .input-prepend{margin-bottom:0}.form-search .radio,.form-search .checkbox,.form-inline .radio,.form-inline .checkbox{padding-left:0;margin-bottom:0;vertical-align:middle}.form-search .radio input[type="radio"],.form-search .checkbox input[type="checkbox"],.form-inline .radio input[type="radio"],.form-inline .checkbox input[type="checkbox"]{float:left;margin-right:3px;margin-left:0}.control-group{margin-bottom:10px}legend+.control-group{margin-top:20px;-webkit-margin-top-collapse:separate}.form-horizontal .control-group{margin-bottom:20px;*zoom:1}.form-horizontal .control-group:before,.form-horizontal .control-group:after{display:table;line-height:0;content:""}.form-horizontal .control-group:after{clear:both}.form-horizontal .control-label{float:left;width:140px;padding-top:5px;text-align:right}.form-horizontal .controls{*display:inline-block;*padding-left:20px;margin-left:160px;*margin-left:0}.form-horizontal .controls:first-child{*padding-left:160px}.form-horizontal .help-block{margin-top:10px;margin-bottom:0}.form-horizontal .form-actions{padding-left:160px}table{max-width:100%;background-color:transparent;border-collapse:collapse;border-spacing:0}.table{width:100%;margin-bottom:20px}.table th,.table td{padding:8px;line-height:20px;text-align:left;vertical-align:top;border-top:1px solid #ddd}.table th{font-weight:bold}.table thead th{vertical-align:bottom}.table caption+thead tr:first-child th,.table caption+thead tr:first-child td,.table colgroup+thead tr:first-child th,.table colgroup+thead tr:first-child td,.table thead:first-child tr:first-child th,.table thead:first-child tr:first-child td{border-top:0}.table tbody+tbody{border-top:2px solid #ddd}.table-condensed th,.table-condensed td{padding:4px 5px}.table-bordered{border:1px solid #ddd;border-collapse:separate;*border-collapse:collapse;border-left:0;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.table-bordered th,.table-bordered td{border-left:1px solid #ddd}.table-bordered caption+thead tr:first-child th,.table-bordered caption+tbody tr:first-child th,.table-bordered caption+tbody tr:first-child td,.table-bordered colgroup+thead tr:first-child th,.table-bordered colgroup+tbody tr:first-child th,.table-bordered colgroup+tbody tr:first-child td,.table-bordered thead:first-child tr:first-child th,.table-bordered tbody:first-child tr:first-child th,.table-bordered tbody:first-child tr:first-child td{border-top:0}.table-bordered thead:first-child tr:first-child th:first-child,.table-bordered tbody:first-child tr:first-child td:first-child{-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-topleft:4px}.table-bordered thead:first-child tr:first-child th:last-child,.table-bordered tbody:first-child tr:first-child td:last-child{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-moz-border-radius-topright:4px}.table-bordered thead:last-child tr:last-child th:first-child,.table-bordered tbody:last-child tr:last-child td:first-child,.table-bordered tfoot:last-child tr:last-child td:first-child{-webkit-border-radius:0 0 0 4px;-moz-border-radius:0 0 0 4px;border-radius:0 0 0 4px;-webkit-border-bottom-left-radius:4px;border-bottom-left-radius:4px;-moz-border-radius-bottomleft:4px}.table-bordered thead:last-child tr:last-child th:last-child,.table-bordered tbody:last-child tr:last-child td:last-child,.table-bordered tfoot:last-child tr:last-child td:last-child{-webkit-border-bottom-right-radius:4px;border-bottom-right-radius:4px;-moz-border-radius-bottomright:4px}.table-bordered caption+thead tr:first-child th:first-child,.table-bordered caption+tbody tr:first-child td:first-child,.table-bordered colgroup+thead tr:first-child th:first-child,.table-bordered colgroup+tbody tr:first-child td:first-child{-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-topleft:4px}.table-bordered caption+thead tr:first-child th:last-child,.table-bordered caption+tbody tr:first-child td:last-child,.table-bordered colgroup+thead tr:first-child th:last-child,.table-bordered colgroup+tbody tr:first-child td:last-child{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-moz-border-right-topleft:4px}.table-striped tbody tr:nth-child(odd) td,.table-striped tbody tr:nth-child(odd) th{background-color:#f9f9f9}.table-hover tbody tr:hover td,.table-hover tbody tr:hover th{background-color:#f5f5f5}table [class*=span],.row-fluid table [class*=span]{display:table-cell;float:none;margin-left:0}table .span1{float:none;width:44px;margin-left:0}table .span2{float:none;width:124px;margin-left:0}table .span3{float:none;width:204px;margin-left:0}table .span4{float:none;width:284px;margin-left:0}table .span5{float:none;width:364px;margin-left:0}table .span6{float:none;width:444px;margin-left:0}table .span7{float:none;width:524px;margin-left:0}table .span8{float:none;width:604px;margin-left:0}table .span9{float:none;width:684px;margin-left:0}table .span10{float:none;width:764px;margin-left:0}table .span11{float:none;width:844px;margin-left:0}table .span12{float:none;width:924px;margin-left:0}table .span13{float:none;width:1004px;margin-left:0}table .span14{float:none;width:1084px;margin-left:0}table .span15{float:none;width:1164px;margin-left:0}table .span16{float:none;width:1244px;margin-left:0}table .span17{float:none;width:1324px;margin-left:0}table .span18{float:none;width:1404px;margin-left:0}table .span19{float:none;width:1484px;margin-left:0}table .span20{float:none;width:1564px;margin-left:0}table .span21{float:none;width:1644px;margin-left:0}table .span22{float:none;width:1724px;margin-left:0}table .span23{float:none;width:1804px;margin-left:0}table .span24{float:none;width:1884px;margin-left:0}.table tbody tr.success td{background-color:#dff0d8}.table tbody tr.error td{background-color:#f2dede}.table tbody tr.info td{background-color:#d9edf7}[class^="icon-"],[class*=" icon-"]{display:inline-block;width:14px;height:14px;margin-top:1px;*margin-right:.3em;line-height:14px;vertical-align:text-top;background-image:url("../img/glyphicons-halflings.png");background-position:14px 14px;background-repeat:no-repeat}.icon-white,.nav>.active>a>[class^="icon-"],.nav>.active>a>[class*=" icon-"],.dropdown-menu>li>a:hover>[class^="icon-"],.dropdown-menu>li>a:hover>[class*=" icon-"],.dropdown-menu>.active>a>[class^="icon-"],.dropdown-menu>.active>a>[class*=" icon-"]{background-image:url("../img/glyphicons-halflings-white.png")}.icon-glass{background-position:0 0}.icon-music{background-position:-24px 0}.icon-search{background-position:-48px 0}.icon-envelope{background-position:-72px 0}.icon-heart{background-position:-96px 0}.icon-star{background-position:-120px 0}.icon-star-empty{background-position:-144px 0}.icon-user{background-position:-168px 0}.icon-film{background-position:-192px 0}.icon-th-large{background-position:-216px 0}.icon-th{background-position:-240px 0}.icon-th-list{background-position:-264px 0}.icon-ok{background-position:-288px 0}.icon-remove{background-position:-312px 0}.icon-zoom-in{background-position:-336px 0}.icon-zoom-out{background-position:-360px 0}.icon-off{background-position:-384px 0}.icon-signal{background-position:-408px 0}.icon-cog{background-position:-432px 0}.icon-trash{background-position:-456px 0}.icon-home{background-position:0 -24px}.icon-file{background-position:-24px -24px}.icon-time{background-position:-48px -24px}.icon-road{background-position:-72px -24px}.icon-download-alt{background-position:-96px -24px}.icon-download{background-position:-120px -24px}.icon-upload{background-position:-144px -24px}.icon-inbox{background-position:-168px -24px}.icon-play-circle{background-position:-192px -24px}.icon-repeat{background-position:-216px -24px}.icon-refresh{background-position:-240px -24px}.icon-list-alt{background-position:-264px -24px}.icon-lock{background-position:-287px -24px}.icon-flag{background-position:-312px -24px}.icon-headphones{background-position:-336px -24px}.icon-volume-off{background-position:-360px -24px}.icon-volume-down{background-position:-384px -24px}.icon-volume-up{background-position:-408px -24px}.icon-qrcode{background-position:-432px -24px}.icon-barcode{background-position:-456px -24px}.icon-tag{background-position:0 -48px}.icon-tags{background-position:-25px -48px}.icon-book{background-position:-48px -48px}.icon-bookmark{background-position:-72px -48px}.icon-print{background-position:-96px -48px}.icon-camera{background-position:-120px -48px}.icon-font{background-position:-144px -48px}.icon-bold{background-position:-167px -48px}.icon-italic{background-position:-192px -48px}.icon-text-height{background-position:-216px -48px}.icon-text-width{background-position:-240px -48px}.icon-align-left{background-position:-264px -48px}.icon-align-center{background-position:-288px -48px}.icon-align-right{background-position:-312px -48px}.icon-align-justify{background-position:-336px -48px}.icon-list{background-position:-360px -48px}.icon-indent-left{background-position:-384px -48px}.icon-indent-right{background-position:-408px -48px}.icon-facetime-video{background-position:-432px -48px}.icon-picture{background-position:-456px -48px}.icon-pencil{background-position:0 -72px}.icon-map-marker{background-position:-24px -72px}.icon-adjust{background-position:-48px -72px}.icon-tint{background-position:-72px -72px}.icon-edit{background-position:-96px -72px}.icon-share{background-position:-120px -72px}.icon-check{background-position:-144px -72px}.icon-move{background-position:-168px -72px}.icon-step-backward{background-position:-192px -72px}.icon-fast-backward{background-position:-216px -72px}.icon-backward{background-position:-240px -72px}.icon-play{background-position:-264px -72px}.icon-pause{background-position:-288px -72px}.icon-stop{background-position:-312px -72px}.icon-forward{background-position:-336px -72px}.icon-fast-forward{background-position:-360px -72px}.icon-step-forward{background-position:-384px -72px}.icon-eject{background-position:-408px -72px}.icon-chevron-left{background-position:-432px -72px}.icon-chevron-right{background-position:-456px -72px}.icon-plus-sign{background-position:0 -96px}.icon-minus-sign{background-position:-24px -96px}.icon-remove-sign{background-position:-48px -96px}.icon-ok-sign{background-position:-72px -96px}.icon-question-sign{background-position:-96px -96px}.icon-info-sign{background-position:-120px -96px}.icon-screenshot{background-position:-144px -96px}.icon-remove-circle{background-position:-168px -96px}.icon-ok-circle{background-position:-192px -96px}.icon-ban-circle{background-position:-216px -96px}.icon-arrow-left{background-position:-240px -96px}.icon-arrow-right{background-position:-264px -96px}.icon-arrow-up{background-position:-289px -96px}.icon-arrow-down{background-position:-312px -96px}.icon-share-alt{background-position:-336px -96px}.icon-resize-full{background-position:-360px -96px}.icon-resize-small{background-position:-384px -96px}.icon-plus{background-position:-408px -96px}.icon-minus{background-position:-433px -96px}.icon-asterisk{background-position:-456px -96px}.icon-exclamation-sign{background-position:0 -120px}.icon-gift{background-position:-24px -120px}.icon-leaf{background-position:-48px -120px}.icon-fire{background-position:-72px -120px}.icon-eye-open{background-position:-96px -120px}.icon-eye-close{background-position:-120px -120px}.icon-warning-sign{background-position:-144px -120px}.icon-plane{background-position:-168px -120px}.icon-calendar{background-position:-192px -120px}.icon-random{width:16px;background-position:-216px -120px}.icon-comment{background-position:-240px -120px}.icon-magnet{background-position:-264px -120px}.icon-chevron-up{background-position:-288px -120px}.icon-chevron-down{background-position:-313px -119px}.icon-retweet{background-position:-336px -120px}.icon-shopping-cart{background-position:-360px -120px}.icon-folder-close{background-position:-384px -120px}.icon-folder-open{width:16px;background-position:-408px -120px}.icon-resize-vertical{background-position:-432px -119px}.icon-resize-horizontal{background-position:-456px -118px}.icon-hdd{background-position:0 -144px}.icon-bullhorn{background-position:-24px -144px}.icon-bell{background-position:-48px -144px}.icon-certificate{background-position:-72px -144px}.icon-thumbs-up{background-position:-96px -144px}.icon-thumbs-down{background-position:-120px -144px}.icon-hand-right{background-position:-144px -144px}.icon-hand-left{background-position:-168px -144px}.icon-hand-up{background-position:-192px -144px}.icon-hand-down{background-position:-216px -144px}.icon-circle-arrow-right{background-position:-240px -144px}.icon-circle-arrow-left{background-position:-264px -144px}.icon-circle-arrow-up{background-position:-288px -144px}.icon-circle-arrow-down{background-position:-312px -144px}.icon-globe{background-position:-336px -144px}.icon-wrench{background-position:-360px -144px}.icon-tasks{background-position:-384px -144px}.icon-filter{background-position:-408px -144px}.icon-briefcase{background-position:-432px -144px}.icon-fullscreen{background-position:-456px -144px}.dropup,.dropdown{position:relative}.dropdown-toggle{*margin-bottom:-3px}.dropdown-toggle:active,.open .dropdown-toggle{outline:0}.caret{display:inline-block;width:0;height:0;vertical-align:top;border-top:4px solid #000;border-right:4px solid transparent;border-left:4px solid transparent;content:""}.dropdown .caret{margin-top:8px;margin-left:2px}.dropdown-menu{position:absolute;top:100%;left:0;z-index:1000;display:none;float:left;min-width:160px;padding:5px 0;margin:2px 0 0;list-style:none;background-color:#fff;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.2);*border-right-width:2px;*border-bottom-width:2px;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px;-webkit-box-shadow:0 5px 10px rgba(0,0,0,0.2);-moz-box-shadow:0 5px 10px rgba(0,0,0,0.2);box-shadow:0 5px 10px rgba(0,0,0,0.2);-webkit-background-clip:padding-box;-moz-background-clip:padding;background-clip:padding-box}.dropdown-menu.pull-right{right:0;left:auto}.dropdown-menu .divider{*width:100%;height:1px;margin:9px 1px;*margin:-5px 0 5px;overflow:hidden;background-color:#e5e5e5;border-bottom:1px solid #fff}.dropdown-menu a{display:block;padding:3px 20px;clear:both;font-weight:normal;line-height:20px;color:#333;white-space:nowrap}.dropdown-menu li>a:hover,.dropdown-menu li>a:focus,.dropdown-submenu:hover>a{color:#fff;text-decoration:none;background-color:#0098cc;background-color:#0098cc;background-image:-moz-linear-gradient(top,#0098cc,#0087b3);background-image:-webkit-gradient(linear,0 0,0 100%,from(#0098cc),to(#0087b3));background-image:-webkit-linear-gradient(top,#0098cc,#0087b3);background-image:-o-linear-gradient(top,#0098cc,#0087b3);background-image:linear-gradient(to bottom,#0098cc,#0087b3);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff0098cc',endColorstr='#ff0087b3',GradientType=0)}.dropdown-menu .active>a,.dropdown-menu .active>a:hover{color:#fff;text-decoration:none;background-color:#0098cc;background-color:#0081c2;background-image:linear-gradient(to bottom,#0098cc,#0087b3);background-image:-moz-linear-gradient(top,#0098cc,#0087b3);background-image:-webkit-gradient(linear,0 0,0 100%,from(#0098cc),to(#0087b3));background-image:-webkit-linear-gradient(top,#0098cc,#0087b3);background-image:-o-linear-gradient(top,#0098cc,#0087b3);background-repeat:repeat-x;outline:0;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff0098cc',endColorstr='#ff0087b3',GradientType=0)}.dropdown-menu .disabled>a,.dropdown-menu .disabled>a:hover{color:#999}.dropdown-menu .disabled>a:hover{text-decoration:none;cursor:default;background-color:transparent}.open{*z-index:1000}.open>.dropdown-menu{display:block}.pull-right>.dropdown-menu{right:0;left:auto}.dropup .caret,.navbar-fixed-bottom .dropdown .caret{border-top:0;border-bottom:4px solid #000;content:"\2191"}.dropup .dropdown-menu,.navbar-fixed-bottom .dropdown .dropdown-menu{top:auto;bottom:100%;margin-bottom:1px}.dropdown-submenu{position:relative}.dropdown-submenu>.dropdown-menu{top:0;left:100%;margin-top:-6px;margin-left:-1px;-webkit-border-radius:0 6px 6px 6px;-moz-border-radius:0 6px 6px 6px;border-radius:0 6px 6px 6px}.dropdown-submenu:hover .dropdown-menu{display:block}.dropdown-submenu>a:after{display:block;float:right;width:0;height:0;margin-top:5px;margin-right:-10px;border-color:transparent;border-left-color:#ccc;border-style:solid;border-width:5px 0 5px 5px;content:" "}.dropdown-submenu:hover>a:after{border-left-color:#fff}.dropdown .dropdown-menu .nav-header{padding-right:20px;padding-left:20px}.typeahead{margin-top:2px;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.well{min-height:20px;padding:19px;margin-bottom:20px;background-color:#f5f5f5;border:1px solid #e3e3e3;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;-webkit-box-shadow:inset 0 1px 1px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 1px 1px rgba(0,0,0,0.05);box-shadow:inset 0 1px 1px rgba(0,0,0,0.05)}.well blockquote{border-color:#ddd;border-color:rgba(0,0,0,0.15)}.well-large{padding:24px;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.well-small{padding:9px;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}.fade{opacity:0;-webkit-transition:opacity .15s linear;-moz-transition:opacity .15s linear;-o-transition:opacity .15s linear;transition:opacity .15s linear}.fade.in{opacity:1}.collapse{position:relative;height:0;overflow:hidden;overflow:visible \9;-webkit-transition:height .35s ease;-moz-transition:height .35s ease;-o-transition:height .35s ease;transition:height .35s ease}.collapse.in{height:auto}.close{float:right;font-size:20px;font-weight:bold;line-height:20px;color:#000;text-shadow:0 1px 0 #fff;opacity:.2;filter:alpha(opacity=20)}.close:hover{color:#000;text-decoration:none;cursor:pointer;opacity:.4;filter:alpha(opacity=40)}button.close{padding:0;cursor:pointer;background:transparent;border:0;-webkit-appearance:none}.btn{display:inline-block;*display:inline;padding:4px 14px;margin-bottom:0;*margin-left:.3em;font-size:14px;line-height:20px;*line-height:20px;color:#333;text-align:center;text-shadow:0 1px 1px rgba(255,255,255,0.75);vertical-align:middle;cursor:pointer;background-color:#f5f5f5;*background-color:#e6e6e6;background-image:-webkit-gradient(linear,0 0,0 100%,from(#fff),to(#e6e6e6));background-image:-webkit-linear-gradient(top,#fff,#e6e6e6);background-image:-o-linear-gradient(top,#fff,#e6e6e6);background-image:linear-gradient(to bottom,#fff,#e6e6e6);background-image:-moz-linear-gradient(top,#fff,#e6e6e6);background-repeat:repeat-x;border:1px solid #bbb;*border:0;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);border-color:#e6e6e6 #e6e6e6 #bfbfbf;border-bottom-color:#a2a2a2;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffffffff',endColorstr='#ffe6e6e6',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false);*zoom:1;-webkit-box-shadow:inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05)}.btn:hover,.btn:active,.btn.active,.btn.disabled,.btn[disabled]{color:#333;background-color:#e6e6e6;*background-color:#d9d9d9}.btn:active,.btn.active{background-color:#ccc \9}.btn:first-child{*margin-left:0}.btn:hover{color:#333;text-decoration:none;background-color:#e6e6e6;*background-color:#d9d9d9;background-position:0 -15px;-webkit-transition:background-position .1s linear;-moz-transition:background-position .1s linear;-o-transition:background-position .1s linear;transition:background-position .1s linear}.btn:focus{outline:thin dotted #333;outline:5px auto -webkit-focus-ring-color;outline-offset:-2px}.btn.active,.btn:active{background-color:#e6e6e6;background-color:#d9d9d9 \9;background-image:none;outline:0;-webkit-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05)}.btn.disabled,.btn[disabled]{cursor:default;background-color:#e6e6e6;background-image:none;opacity:.65;filter:alpha(opacity=65);-webkit-box-shadow:none;-moz-box-shadow:none;box-shadow:none}.btn-large{padding:9px 14px;font-size:16px;line-height:normal;-webkit-border-radius:5px;-moz-border-radius:5px;border-radius:5px}.btn-large [class^="icon-"]{margin-top:2px}.btn-small{padding:3px 9px;font-size:12px;line-height:18px}.btn-small [class^="icon-"]{margin-top:0}.btn-mini{padding:2px 6px;font-size:11px;line-height:16px}.btn-block{display:block;width:100%;padding-right:0;padding-left:0;-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box}.btn-block+.btn-block{margin-top:5px}.btn-primary.active,.btn-warning.active,.btn-danger.active,.btn-success.active,.btn-info.active,.btn-inverse.active{color:rgba(255,255,255,0.75)}.btn{border-color:#c5c5c5;border-color:rgba(0,0,0,0.15) rgba(0,0,0,0.15) rgba(0,0,0,0.25)}.btn-primary{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#006dcc;*background-color:#04c;background-image:-webkit-gradient(linear,0 0,0 100%,from(#08c),to(#04c));background-image:-webkit-linear-gradient(top,#08c,#04c);background-image:-o-linear-gradient(top,#08c,#04c);background-image:linear-gradient(to bottom,#08c,#04c);background-image:-moz-linear-gradient(top,#08c,#04c);background-repeat:repeat-x;border-color:#04c #04c #002a80;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff0088cc',endColorstr='#ff0044cc',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-primary:hover,.btn-primary:active,.btn-primary.active,.btn-primary.disabled,.btn-primary[disabled]{color:#fff;background-color:#04c;*background-color:#003bb3}.btn-primary:active,.btn-primary.active{background-color:#039 \9}.btn-warning{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#faa732;*background-color:#f89406;background-image:-webkit-gradient(linear,0 0,0 100%,from(#fbb450),to(#f89406));background-image:-webkit-linear-gradient(top,#fbb450,#f89406);background-image:-o-linear-gradient(top,#fbb450,#f89406);background-image:linear-gradient(to bottom,#fbb450,#f89406);background-image:-moz-linear-gradient(top,#fbb450,#f89406);background-repeat:repeat-x;border-color:#f89406 #f89406 #ad6704;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fffbb450',endColorstr='#fff89406',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-warning:hover,.btn-warning:active,.btn-warning.active,.btn-warning.disabled,.btn-warning[disabled]{color:#fff;background-color:#f89406;*background-color:#df8505}.btn-warning:active,.btn-warning.active{background-color:#c67605 \9}.btn-danger{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#da4f49;*background-color:#bd362f;background-image:-webkit-gradient(linear,0 0,0 100%,from(#ee5f5b),to(#bd362f));background-image:-webkit-linear-gradient(top,#ee5f5b,#bd362f);background-image:-o-linear-gradient(top,#ee5f5b,#bd362f);background-image:linear-gradient(to bottom,#ee5f5b,#bd362f);background-image:-moz-linear-gradient(top,#ee5f5b,#bd362f);background-repeat:repeat-x;border-color:#bd362f #bd362f #802420;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffee5f5b',endColorstr='#ffbd362f',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-danger:hover,.btn-danger:active,.btn-danger.active,.btn-danger.disabled,.btn-danger[disabled]{color:#fff;background-color:#bd362f;*background-color:#a9302a}.btn-danger:active,.btn-danger.active{background-color:#942a25 \9}.btn-success{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#5bb75b;*background-color:#51a351;background-image:-webkit-gradient(linear,0 0,0 100%,from(#62c462),to(#51a351));background-image:-webkit-linear-gradient(top,#62c462,#51a351);background-image:-o-linear-gradient(top,#62c462,#51a351);background-image:linear-gradient(to bottom,#62c462,#51a351);background-image:-moz-linear-gradient(top,#62c462,#51a351);background-repeat:repeat-x;border-color:#51a351 #51a351 #387038;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff62c462',endColorstr='#ff51a351',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-success:hover,.btn-success:active,.btn-success.active,.btn-success.disabled,.btn-success[disabled]{color:#fff;background-color:#51a351;*background-color:#499249}.btn-success:active,.btn-success.active{background-color:#408140 \9}.btn-info{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#49afcd;*background-color:#2f96b4;background-image:-webkit-gradient(linear,0 0,0 100%,from(#5bc0de),to(#2f96b4));background-image:-webkit-linear-gradient(top,#5bc0de,#2f96b4);background-image:-o-linear-gradient(top,#5bc0de,#2f96b4);background-image:linear-gradient(to bottom,#5bc0de,#2f96b4);background-image:-moz-linear-gradient(top,#5bc0de,#2f96b4);background-repeat:repeat-x;border-color:#2f96b4 #2f96b4 #1f6377;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff5bc0de',endColorstr='#ff2f96b4',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-info:hover,.btn-info:active,.btn-info.active,.btn-info.disabled,.btn-info[disabled]{color:#fff;background-color:#2f96b4;*background-color:#2a85a0}.btn-info:active,.btn-info.active{background-color:#24748c \9}.btn-inverse{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#363636;*background-color:#222;background-image:-webkit-gradient(linear,0 0,0 100%,from(#444),to(#222));background-image:-webkit-linear-gradient(top,#444,#222);background-image:-o-linear-gradient(top,#444,#222);background-image:linear-gradient(to bottom,#444,#222);background-image:-moz-linear-gradient(top,#444,#222);background-repeat:repeat-x;border-color:#222 #222 #000;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff444444',endColorstr='#ff222222',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.btn-inverse:hover,.btn-inverse:active,.btn-inverse.active,.btn-inverse.disabled,.btn-inverse[disabled]{color:#fff;background-color:#222;*background-color:#151515}.btn-inverse:active,.btn-inverse.active{background-color:#080808 \9}button.btn,input[type="submit"].btn{*padding-top:3px;*padding-bottom:3px}button.btn::-moz-focus-inner,input[type="submit"].btn::-moz-focus-inner{padding:0;border:0}button.btn.btn-large,input[type="submit"].btn.btn-large{*padding-top:7px;*padding-bottom:7px}button.btn.btn-small,input[type="submit"].btn.btn-small{*padding-top:3px;*padding-bottom:3px}button.btn.btn-mini,input[type="submit"].btn.btn-mini{*padding-top:1px;*padding-bottom:1px}.btn-link,.btn-link:active{background-color:transparent;background-image:none;-webkit-box-shadow:none;-moz-box-shadow:none;box-shadow:none}.btn-link{color:#08c;cursor:pointer;border-color:transparent;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.btn-link:hover{color:#005580;text-decoration:underline;background-color:transparent}.btn-group{position:relative;*margin-left:.3em;font-size:0;white-space:nowrap}.btn-group:first-child{*margin-left:0}.btn-group+.btn-group{margin-left:5px}.btn-toolbar{margin-top:10px;margin-bottom:10px;font-size:0}.btn-toolbar .btn-group{display:inline-block;*display:inline;*zoom:1}.btn-toolbar .btn+.btn,.btn-toolbar .btn-group+.btn,.btn-toolbar .btn+.btn-group{margin-left:5px}.btn-group>.btn{position:relative;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.btn-group>.btn+.btn{margin-left:-1px}.btn-group>.btn,.btn-group>.dropdown-menu{font-size:14px}.btn-group>.btn-mini{font-size:11px}.btn-group>.btn-small{font-size:12px}.btn-group>.btn-large{font-size:16px}.btn-group>.btn:first-child{margin-left:0;-webkit-border-bottom-left-radius:4px;border-bottom-left-radius:4px;-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-bottomleft:4px;-moz-border-radius-topleft:4px}.btn-group>.btn:last-child,.btn-group>.dropdown-toggle{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-webkit-border-bottom-right-radius:4px;border-bottom-right-radius:4px;-moz-border-radius-topright:4px;-moz-border-radius-bottomright:4px}.btn-group>.btn.large:first-child{margin-left:0;-webkit-border-bottom-left-radius:6px;border-bottom-left-radius:6px;-webkit-border-top-left-radius:6px;border-top-left-radius:6px;-moz-border-radius-bottomleft:6px;-moz-border-radius-topleft:6px}.btn-group>.btn.large:last-child,.btn-group>.large.dropdown-toggle{-webkit-border-top-right-radius:6px;border-top-right-radius:6px;-webkit-border-bottom-right-radius:6px;border-bottom-right-radius:6px;-moz-border-radius-topright:6px;-moz-border-radius-bottomright:6px}.btn-group>.btn:hover,.btn-group>.btn:focus,.btn-group>.btn:active,.btn-group>.btn.active{z-index:2}.btn-group .dropdown-toggle:active,.btn-group.open .dropdown-toggle{outline:0}.btn-group>.btn+.dropdown-toggle{*padding-top:5px;padding-right:8px;*padding-bottom:5px;padding-left:8px;-webkit-box-shadow:inset 1px 0 0 rgba(255,255,255,0.125),inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 1px 0 0 rgba(255,255,255,0.125),inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 1px 0 0 rgba(255,255,255,0.125),inset 0 1px 0 rgba(255,255,255,0.2),0 1px 2px rgba(0,0,0,0.05)}.btn-group>.btn-mini+.dropdown-toggle{*padding-top:2px;padding-right:5px;*padding-bottom:2px;padding-left:5px}.btn-group>.btn-small+.dropdown-toggle{*padding-top:5px;*padding-bottom:4px}.btn-group>.btn-large+.dropdown-toggle{*padding-top:7px;padding-right:12px;*padding-bottom:7px;padding-left:12px}.btn-group.open .dropdown-toggle{background-image:none;-webkit-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05);box-shadow:inset 0 2px 4px rgba(0,0,0,0.15),0 1px 2px rgba(0,0,0,0.05)}.btn-group.open .btn.dropdown-toggle{background-color:#e6e6e6}.btn-group.open .btn-primary.dropdown-toggle{background-color:#04c}.btn-group.open .btn-warning.dropdown-toggle{background-color:#f89406}.btn-group.open .btn-danger.dropdown-toggle{background-color:#bd362f}.btn-group.open .btn-success.dropdown-toggle{background-color:#51a351}.btn-group.open .btn-info.dropdown-toggle{background-color:#2f96b4}.btn-group.open .btn-inverse.dropdown-toggle{background-color:#222}.btn .caret{margin-top:8px;margin-left:0}.btn-mini .caret,.btn-small .caret,.btn-large .caret{margin-top:6px}.btn-large .caret{border-top-width:5px;border-right-width:5px;border-left-width:5px}.dropup .btn-large .caret{border-top:0;border-bottom:5px solid #000}.btn-primary .caret,.btn-warning .caret,.btn-danger .caret,.btn-info .caret,.btn-success .caret,.btn-inverse .caret{border-top-color:#fff;border-bottom-color:#fff}.btn-group-vertical{display:inline-block;*display:inline;*zoom:1}.btn-group-vertical .btn{display:block;float:none;width:100%;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.btn-group-vertical .btn+.btn{margin-top:-1px;margin-left:0}.btn-group-vertical .btn:first-child{-webkit-border-radius:4px 4px 0 0;-moz-border-radius:4px 4px 0 0;border-radius:4px 4px 0 0}.btn-group-vertical .btn:last-child{-webkit-border-radius:0 0 4px 4px;-moz-border-radius:0 0 4px 4px;border-radius:0 0 4px 4px}.btn-group-vertical .btn-large:first-child{-webkit-border-radius:6px 6px 0 0;-moz-border-radius:6px 6px 0 0;border-radius:6px 6px 0 0}.btn-group-vertical .btn-large:last-child{-webkit-border-radius:0 0 6px 6px;-moz-border-radius:0 0 6px 6px;border-radius:0 0 6px 6px}.alert{padding:8px 35px 8px 14px;margin-bottom:20px;color:#c09853;text-shadow:0 1px 0 rgba(255,255,255,0.5);background-color:#fcf8e3;border:1px solid #fbeed5;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.alert h4{margin:0}.alert .close{position:relative;top:-2px;right:-21px;line-height:20px}.alert-success{color:#468847;background-color:#dff0d8;border-color:#d6e9c6}.alert-danger,.alert-error{color:#b94a48;background-color:#f2dede;border-color:#eed3d7}.alert-info{color:#3a87ad;background-color:#d9edf7;border-color:#bce8f1}.alert-block{padding-top:14px;padding-bottom:14px}.alert-block>p,.alert-block>ul{margin-bottom:0}.alert-block p+p{margin-top:5px}.nav{margin-bottom:20px;margin-left:0;list-style:none}.nav>li>a{display:block}.nav>li>a:hover{text-decoration:none;background-color:#eee}.nav>.pull-right{float:right}.nav-header{display:block;padding:3px 15px;font-size:11px;font-weight:bold;line-height:20px;color:#999;text-shadow:0 1px 0 rgba(255,255,255,0.5);text-transform:uppercase}.nav li+.nav-header{margin-top:9px}.nav-list{padding-right:15px;padding-left:15px;margin-bottom:0}.nav-list>li>a,.nav-list .nav-header{margin-right:-15px;margin-left:-15px;text-shadow:0 1px 0 rgba(255,255,255,0.5)}.nav-list>li>a{padding:3px 15px}.nav-list>.active>a,.nav-list>.active>a:hover{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.2);background-color:#08c}.nav-list [class^="icon-"]{margin-right:2px}.nav-list .divider{*width:100%;height:1px;margin:9px 1px;*margin:-5px 0 5px;overflow:hidden;background-color:#e5e5e5;border-bottom:1px solid #fff}.nav-tabs,.nav-pills{*zoom:1}.nav-tabs:before,.nav-pills:before,.nav-tabs:after,.nav-pills:after{display:table;line-height:0;content:""}.nav-tabs:after,.nav-pills:after{clear:both}.nav-tabs>li,.nav-pills>li{float:left}.nav-tabs>li>a,.nav-pills>li>a{padding-right:12px;padding-left:12px;margin-right:2px;line-height:14px}.nav-tabs{border-bottom:1px solid #ddd}.nav-tabs>li{margin-bottom:-1px}.nav-tabs>li>a{padding-top:8px;padding-bottom:8px;line-height:20px;border:1px solid transparent;-webkit-border-radius:4px 4px 0 0;-moz-border-radius:4px 4px 0 0;border-radius:4px 4px 0 0}.nav-tabs>li>a:hover{border-color:#eee #eee #ddd}.nav-tabs>.active>a,.nav-tabs>.active>a:hover{color:#555;cursor:default;background-color:#fff;border:1px solid #ddd;border-bottom-color:transparent}.nav-pills>li>a{padding-top:8px;padding-bottom:8px;margin-top:2px;margin-bottom:2px;-webkit-border-radius:5px;-moz-border-radius:5px;border-radius:5px}.nav-pills>.active>a,.nav-pills>.active>a:hover{color:#fff;background-color:#08c}.nav-stacked>li{float:none}.nav-stacked>li>a{margin-right:0}.nav-tabs.nav-stacked{border-bottom:0}.nav-tabs.nav-stacked>li>a{border:1px solid #ddd;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.nav-tabs.nav-stacked>li:first-child>a{-webkit-border-top-right-radius:4px;border-top-right-radius:4px;-webkit-border-top-left-radius:4px;border-top-left-radius:4px;-moz-border-radius-topright:4px;-moz-border-radius-topleft:4px}.nav-tabs.nav-stacked>li:last-child>a{-webkit-border-bottom-right-radius:4px;border-bottom-right-radius:4px;-webkit-border-bottom-left-radius:4px;border-bottom-left-radius:4px;-moz-border-radius-bottomright:4px;-moz-border-radius-bottomleft:4px}.nav-tabs.nav-stacked>li>a:hover{z-index:2;border-color:#ddd}.nav-pills.nav-stacked>li>a{margin-bottom:3px}.nav-pills.nav-stacked>li:last-child>a{margin-bottom:1px}.nav-tabs .dropdown-menu{-webkit-border-radius:0 0 6px 6px;-moz-border-radius:0 0 6px 6px;border-radius:0 0 6px 6px}.nav-pills .dropdown-menu{-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.nav .dropdown-toggle .caret{margin-top:6px;border-top-color:#08c;border-bottom-color:#08c}.nav .dropdown-toggle:hover .caret{border-top-color:#005580;border-bottom-color:#005580}.nav-tabs .dropdown-toggle .caret{margin-top:8px}.nav .active .dropdown-toggle .caret{border-top-color:#fff;border-bottom-color:#fff}.nav-tabs .active .dropdown-toggle .caret{border-top-color:#555;border-bottom-color:#555}.nav>.dropdown.active>a:hover{cursor:pointer}.nav-tabs .open .dropdown-toggle,.nav-pills .open .dropdown-toggle,.nav>li.dropdown.open.active>a:hover{color:#fff;background-color:#999;border-color:#999}.nav li.dropdown.open .caret,.nav li.dropdown.open.active .caret,.nav li.dropdown.open a:hover .caret{border-top-color:#fff;border-bottom-color:#fff;opacity:1;filter:alpha(opacity=100)}.tabs-stacked .open>a:hover{border-color:#999}.tabbable{*zoom:1}.tabbable:before,.tabbable:after{display:table;line-height:0;content:""}.tabbable:after{clear:both}.tab-content{overflow:auto}.tabs-below>.nav-tabs,.tabs-right>.nav-tabs,.tabs-left>.nav-tabs{border-bottom:0}.tab-content>.tab-pane,.pill-content>.pill-pane{display:none}.tab-content>.active,.pill-content>.active{display:block}.tabs-below>.nav-tabs{border-top:1px solid #ddd}.tabs-below>.nav-tabs>li{margin-top:-1px;margin-bottom:0}.tabs-below>.nav-tabs>li>a{-webkit-border-radius:0 0 4px 4px;-moz-border-radius:0 0 4px 4px;border-radius:0 0 4px 4px}.tabs-below>.nav-tabs>li>a:hover{border-top-color:#ddd;border-bottom-color:transparent}.tabs-below>.nav-tabs>.active>a,.tabs-below>.nav-tabs>.active>a:hover{border-color:transparent #ddd #ddd #ddd}.tabs-left>.nav-tabs>li,.tabs-right>.nav-tabs>li{float:none}.tabs-left>.nav-tabs>li>a,.tabs-right>.nav-tabs>li>a{min-width:74px;margin-right:0;margin-bottom:3px}.tabs-left>.nav-tabs{float:left;margin-right:19px;border-right:1px solid #ddd}.tabs-left>.nav-tabs>li>a{margin-right:-1px;-webkit-border-radius:4px 0 0 4px;-moz-border-radius:4px 0 0 4px;border-radius:4px 0 0 4px}.tabs-left>.nav-tabs>li>a:hover{border-color:#eee #ddd #eee #eee}.tabs-left>.nav-tabs .active>a,.tabs-left>.nav-tabs .active>a:hover{border-color:#ddd transparent #ddd #ddd;*border-right-color:#fff}.tabs-right>.nav-tabs{float:right;margin-left:19px;border-left:1px solid #ddd}.tabs-right>.nav-tabs>li>a{margin-left:-1px;-webkit-border-radius:0 4px 4px 0;-moz-border-radius:0 4px 4px 0;border-radius:0 4px 4px 0}.tabs-right>.nav-tabs>li>a:hover{border-color:#eee #eee #eee #ddd}.tabs-right>.nav-tabs .active>a,.tabs-right>.nav-tabs .active>a:hover{border-color:#ddd #ddd #ddd transparent;*border-left-color:#fff}.nav>.disabled>a{color:#999}.nav>.disabled>a:hover{text-decoration:none;cursor:default;background-color:transparent}.navbar{*position:relative;*z-index:2;margin-bottom:20px;overflow:visible;color:#555}.navbar-inner{min-height:40px;padding-right:20px;padding-left:20px;background-color:#fafafa;background-image:-moz-linear-gradient(top,#fff,#e2f2e2);background-image:-webkit-gradient(linear,0 0,0 100%,from(#fff),to(#e2f2e2));background-image:-webkit-linear-gradient(top,#fff,#e2f2e2);background-image:-o-linear-gradient(top,#fff,#e2f2e2);background-image:linear-gradient(to bottom,#fff,#e2f2e2);background-repeat:repeat-x;border:1px solid #d4d4d4;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffffddff',endColorstr='#ffe2f2e2',GradientType=0);-webkit-box-shadow:0 1px 4px rgba(0,0,0,0.065);-moz-box-shadow:0 1px 4px rgba(0,0,0,0.065);box-shadow:0 1px 4px rgba(0,0,0,0.065)}.navbar .container{width:auto}.nav-collapse.collapse{height:auto}.navbar .brand{display:block;float:left;padding:10px 20px 10px;margin-left:-20px;font-size:20px;font-weight:200;color:#555;text-shadow:0 1px 0 #fff}.navbar .brand:hover{text-decoration:none}.navbar-text{margin-bottom:0;line-height:40px}.navbar-link{color:#555}.navbar-link:hover{color:#333}.navbar .divider-vertical{height:40px;margin:0 9px;border-right:1px solid #fff;border-left:1px solid #f2f2f2}.navbar .btn,.navbar .btn-group{margin-top:6px}.navbar .btn-group .btn{margin:0}.navbar-form{margin-bottom:0;*zoom:1}.navbar-form:before,.navbar-form:after{display:table;line-height:0;content:""}.navbar-form:after{clear:both}.navbar-form input,.navbar-form select,.navbar-form .radio,.navbar-form .checkbox{margin-top:5px}.navbar-form input,.navbar-form select,.navbar-form .btn{display:inline-block;margin-bottom:0}.navbar-form input[type="image"],.navbar-form input[type="checkbox"],.navbar-form input[type="radio"]{margin-top:3px}.navbar-form .input-append,.navbar-form .input-prepend{margin-top:6px;white-space:nowrap}.navbar-form .input-append input,.navbar-form .input-prepend input{margin-top:0}.navbar-search{position:relative;float:left;margin-top:5px;margin-bottom:0}.navbar-search .search-query{padding:4px 14px;margin-bottom:0;font-family:"Helvetica Neue",Helvetica,Arial,sans-serif;font-size:13px;font-weight:normal;line-height:1;-webkit-border-radius:15px;-moz-border-radius:15px;border-radius:15px}.navbar-static-top{position:static;width:100%;margin-bottom:0}.navbar-static-top .navbar-inner{-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.navbar-fixed-top,.navbar-fixed-bottom{position:fixed;right:0;left:0;z-index:1030;margin-bottom:0}.navbar-fixed-top .navbar-inner,.navbar-fixed-bottom .navbar-inner,.navbar-static-top .navbar-inner{border:0}.navbar-fixed-top .navbar-inner,.navbar-fixed-bottom .navbar-inner{padding-right:0;padding-left:0;-webkit-border-radius:0;-moz-border-radius:0;border-radius:0}.navbar-static-top .container,.navbar-fixed-top .container,.navbar-fixed-bottom .container{width:940px}.navbar-fixed-top{top:0}.navbar-fixed-top .navbar-inner,.navbar-static-top .navbar-inner{-webkit-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.1),0 1px 10px rgba(0,0,0,0.1);-moz-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.1),0 1px 10px rgba(0,0,0,0.1);box-shadow:inset 0 -1px 0 rgba(0,0,0,0.1),0 1px 10px rgba(0,0,0,0.1)}.navbar-fixed-bottom{bottom:0}.navbar-fixed-bottom .navbar-inner{-webkit-box-shadow:inset 0 1px 0 rgba(0,0,0,0.1),0 -1px 10px rgba(0,0,0,0.1);-moz-box-shadow:inset 0 1px 0 rgba(0,0,0,0.1),0 -1px 10px rgba(0,0,0,0.1);box-shadow:inset 0 1px 0 rgba(0,0,0,0.1),0 -1px 10px rgba(0,0,0,0.1)}.navbar .nav{position:relative;left:0;display:block;float:left;margin:0 10px 0 0}.navbar .nav.pull-right{float:right}.navbar .nav>li{float:left}.navbar .nav>li>a{float:none;padding:10px 15px 10px;color:#555;text-decoration:none;text-shadow:0 1px 0 #fff}.navbar .nav .dropdown-toggle .caret{margin-top:8px}.navbar .nav>li>a:focus,.navbar .nav>li>a:hover{color:#333;text-decoration:none;background-color:transparent}.navbar .nav>.active>a,.navbar .nav>.active>a:hover,.navbar .nav>.active>a:focus{color:#555;text-decoration:none;background-color:#e5e5e5;-webkit-box-shadow:inset 0 3px 8px rgba(0,0,0,0.125);-moz-box-shadow:inset 0 3px 8px rgba(0,0,0,0.125);box-shadow:inset 0 3px 8px rgba(0,0,0,0.125)}.navbar .btn-navbar{display:none;float:right;padding:7px 10px;margin-right:5px;margin-left:5px;color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#ededed;*background-color:#e5e5e5;background-image:-webkit-gradient(linear,0 0,0 100%,from(#f2f2f2),to(#e5e5e5));background-image:-webkit-linear-gradient(top,#f2f2f2,#e5e5e5);background-image:-o-linear-gradient(top,#f2f2f2,#e5e5e5);background-image:linear-gradient(to bottom,#f2f2f2,#e5e5e5);background-image:-moz-linear-gradient(top,#f2f2f2,#e5e5e5);background-repeat:repeat-x;border-color:#e5e5e5 #e5e5e5 #bfbfbf;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fff2f2f2',endColorstr='#ffe5e5e5',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false);-webkit-box-shadow:inset 0 1px 0 rgba(255,255,255,0.1),0 1px 0 rgba(255,255,255,0.075);-moz-box-shadow:inset 0 1px 0 rgba(255,255,255,0.1),0 1px 0 rgba(255,255,255,0.075);box-shadow:inset 0 1px 0 rgba(255,255,255,0.1),0 1px 0 rgba(255,255,255,0.075)}.navbar .btn-navbar:hover,.navbar .btn-navbar:active,.navbar .btn-navbar.active,.navbar .btn-navbar.disabled,.navbar .btn-navbar[disabled]{color:#fff;background-color:#e5e5e5;*background-color:#d9d9d9}.navbar .btn-navbar:active,.navbar .btn-navbar.active{background-color:#ccc \9}.navbar .btn-navbar .icon-bar{display:block;width:18px;height:2px;background-color:#f5f5f5;-webkit-border-radius:1px;-moz-border-radius:1px;border-radius:1px;-webkit-box-shadow:0 1px 0 rgba(0,0,0,0.25);-moz-box-shadow:0 1px 0 rgba(0,0,0,0.25);box-shadow:0 1px 0 rgba(0,0,0,0.25)}.btn-navbar .icon-bar+.icon-bar{margin-top:3px}.navbar .nav>li>.dropdown-menu:before{position:absolute;top:-7px;left:9px;display:inline-block;border-right:7px solid transparent;border-bottom:7px solid #ccc;border-left:7px solid transparent;border-bottom-color:rgba(0,0,0,0.2);content:''}.navbar .nav>li>.dropdown-menu:after{position:absolute;top:-6px;left:10px;display:inline-block;border-right:6px solid transparent;border-bottom:6px solid #fff;border-left:6px solid transparent;content:''}.navbar-fixed-bottom .nav>li>.dropdown-menu:before{top:auto;bottom:-7px;border-top:7px solid #ccc;border-bottom:0;border-top-color:rgba(0,0,0,0.2)}.navbar-fixed-bottom .nav>li>.dropdown-menu:after{top:auto;bottom:-6px;border-top:6px solid #fff;border-bottom:0}.navbar .nav li.dropdown.open>.dropdown-toggle,.navbar .nav li.dropdown.active>.dropdown-toggle,.navbar .nav li.dropdown.open.active>.dropdown-toggle{color:#555;background-color:#e5e5e5}.navbar .nav li.dropdown>.dropdown-toggle .caret{border-top-color:#555;border-bottom-color:#555}.navbar .nav li.dropdown.open>.dropdown-toggle .caret,.navbar .nav li.dropdown.active>.dropdown-toggle .caret,.navbar .nav li.dropdown.open.active>.dropdown-toggle .caret{border-top-color:#555;border-bottom-color:#555}.navbar .pull-right>li>.dropdown-menu,.navbar .nav>li>.dropdown-menu.pull-right{right:0;left:auto}.navbar .pull-right>li>.dropdown-menu:before,.navbar .nav>li>.dropdown-menu.pull-right:before{right:12px;left:auto}.navbar .pull-right>li>.dropdown-menu:after,.navbar .nav>li>.dropdown-menu.pull-right:after{right:13px;left:auto}.navbar .pull-right>li>.dropdown-menu .dropdown-menu,.navbar .nav>li>.dropdown-menu.pull-right .dropdown-menu{right:100%;left:auto;margin-right:-1px;margin-left:0;-webkit-border-radius:6px 0 6px 6px;-moz-border-radius:6px 0 6px 6px;border-radius:6px 0 6px 6px}.navbar-inverse{color:#999}.navbar-inverse .navbar-inner{background-color:#1b1b1b;background-image:-moz-linear-gradient(top,#222,#111);background-image:-webkit-gradient(linear,0 0,0 100%,from(#222),to(#111));background-image:-webkit-linear-gradient(top,#222,#111);background-image:-o-linear-gradient(top,#222,#111);background-image:linear-gradient(to bottom,#222,#111);background-repeat:repeat-x;border-color:#252525;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff222222',endColorstr='#ff111111',GradientType=0)}.navbar-inverse .brand,.navbar-inverse .nav>li>a{color:#999;text-shadow:0 -1px 0 rgba(0,0,0,0.25)}.navbar-inverse .brand:hover,.navbar-inverse .nav>li>a:hover{color:#fff}.navbar-inverse .nav>li>a:focus,.navbar-inverse .nav>li>a:hover{color:#fff;background-color:transparent}.navbar-inverse .nav .active>a,.navbar-inverse .nav .active>a:hover,.navbar-inverse .nav .active>a:focus{color:#fff;background-color:#111}.navbar-inverse .navbar-link{color:#999}.navbar-inverse .navbar-link:hover{color:#fff}.navbar-inverse .divider-vertical{border-right-color:#222;border-left-color:#111}.navbar-inverse .nav li.dropdown.open>.dropdown-toggle,.navbar-inverse .nav li.dropdown.active>.dropdown-toggle,.navbar-inverse .nav li.dropdown.open.active>.dropdown-toggle{color:#fff;background-color:#111}.navbar-inverse .nav li.dropdown>.dropdown-toggle .caret{border-top-color:#999;border-bottom-color:#999}.navbar-inverse .nav li.dropdown.open>.dropdown-toggle .caret,.navbar-inverse .nav li.dropdown.active>.dropdown-toggle .caret,.navbar-inverse .nav li.dropdown.open.active>.dropdown-toggle .caret{border-top-color:#fff;border-bottom-color:#fff}.navbar-inverse .navbar-search .search-query{color:#fff;background-color:#515151;border-color:#111;-webkit-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1),0 1px 0 rgba(255,255,255,0.15);-moz-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1),0 1px 0 rgba(255,255,255,0.15);box-shadow:inset 0 1px 2px rgba(0,0,0,0.1),0 1px 0 rgba(255,255,255,0.15);-webkit-transition:none;-moz-transition:none;-o-transition:none;transition:none}.navbar-inverse .navbar-search .search-query:-moz-placeholder{color:#ccc}.navbar-inverse .navbar-search .search-query:-ms-input-placeholder{color:#ccc}.navbar-inverse .navbar-search .search-query::-webkit-input-placeholder{color:#ccc}.navbar-inverse .navbar-search .search-query:focus,.navbar-inverse .navbar-search .search-query.focused{padding:5px 15px;color:#333;text-shadow:0 1px 0 #fff;background-color:#fff;border:0;outline:0;-webkit-box-shadow:0 0 3px rgba(0,0,0,0.15);-moz-box-shadow:0 0 3px rgba(0,0,0,0.15);box-shadow:0 0 3px rgba(0,0,0,0.15)}.navbar-inverse .btn-navbar{color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#0e0e0e;*background-color:#040404;background-image:-webkit-gradient(linear,0 0,0 100%,from(#151515),to(#040404));background-image:-webkit-linear-gradient(top,#151515,#040404);background-image:-o-linear-gradient(top,#151515,#040404);background-image:linear-gradient(to bottom,#151515,#040404);background-image:-moz-linear-gradient(top,#151515,#040404);background-repeat:repeat-x;border-color:#040404 #040404 #000;border-color:rgba(0,0,0,0.1) rgba(0,0,0,0.1) rgba(0,0,0,0.25);filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff151515',endColorstr='#ff040404',GradientType=0);filter:progid:dximagetransform.microsoft.gradient(enabled=false)}.navbar-inverse .btn-navbar:hover,.navbar-inverse .btn-navbar:active,.navbar-inverse .btn-navbar.active,.navbar-inverse .btn-navbar.disabled,.navbar-inverse .btn-navbar[disabled]{color:#fff;background-color:#040404;*background-color:#000}.navbar-inverse .btn-navbar:active,.navbar-inverse .btn-navbar.active{background-color:#000 \9}.breadcrumb{padding:8px 15px;margin:0 0 20px;list-style:none;background-color:#f5f5f5;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.breadcrumb li{display:inline-block;*display:inline;text-shadow:0 1px 0 #fff;*zoom:1}.breadcrumb .divider{padding:0 5px;color:#ccc}.breadcrumb .active{color:#999}.pagination{height:40px;margin:20px 0}.pagination ul{display:inline-block;*display:inline;margin-bottom:0;margin-left:0;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px;*zoom:1;-webkit-box-shadow:0 1px 2px rgba(0,0,0,0.05);-moz-box-shadow:0 1px 2px rgba(0,0,0,0.05);box-shadow:0 1px 2px rgba(0,0,0,0.05)}.pagination li{display:inline}.pagination a,.pagination span{float:left;padding:0 14px;line-height:38px;text-decoration:none;background-color:#fff;border:1px solid #ddd;border-left-width:0}.pagination a:hover,.pagination .active a,.pagination .active span{background-color:#f5f5f5}.pagination .active a,.pagination .active span{color:#999;cursor:default}.pagination .disabled span,.pagination .disabled a,.pagination .disabled a:hover{color:#999;cursor:default;background-color:transparent}.pagination li:first-child a,.pagination li:first-child span{border-left-width:1px;-webkit-border-radius:3px 0 0 3px;-moz-border-radius:3px 0 0 3px;border-radius:3px 0 0 3px}.pagination li:last-child a,.pagination li:last-child span{-webkit-border-radius:0 3px 3px 0;-moz-border-radius:0 3px 3px 0;border-radius:0 3px 3px 0}.pagination-centered{text-align:center}.pagination-right{text-align:right}.pager{margin:20px 0;text-align:center;list-style:none;*zoom:1}.pager:before,.pager:after{display:table;line-height:0;content:""}.pager:after{clear:both}.pager li{display:inline}.pager a{display:inline-block;padding:5px 14px;background-color:#fff;border:1px solid #ddd;-webkit-border-radius:15px;-moz-border-radius:15px;border-radius:15px}.pager a:hover{text-decoration:none;background-color:#f5f5f5}.pager .next a{float:right}.pager .previous a{float:left}.pager .disabled a,.pager .disabled a:hover{color:#999;cursor:default;background-color:#fff}.modal-open .dropdown-menu{z-index:2050}.modal-open .dropdown.open{*z-index:2050}.modal-open .popover{z-index:2060}.modal-open .tooltip{z-index:2080}.modal-backdrop{position:fixed;top:0;right:0;bottom:0;left:0;z-index:1040;background-color:#000}.modal-backdrop.fade{opacity:0}.modal-backdrop,.modal-backdrop.fade.in{opacity:.8;filter:alpha(opacity=80)}.modal{position:fixed;top:50%;left:50%;z-index:1050;width:560px;margin:-250px 0 0 -280px;overflow:auto;background-color:#fff;border:1px solid #999;border:1px solid rgba(0,0,0,0.3);*border:1px solid #999;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px;-webkit-box-shadow:0 3px 7px rgba(0,0,0,0.3);-moz-box-shadow:0 3px 7px rgba(0,0,0,0.3);box-shadow:0 3px 7px rgba(0,0,0,0.3);-webkit-background-clip:padding-box;-moz-background-clip:padding-box;background-clip:padding-box}.modal.fade{top:-25%;-webkit-transition:opacity .3s linear,top .3s ease-out;-moz-transition:opacity .3s linear,top .3s ease-out;-o-transition:opacity .3s linear,top .3s ease-out;transition:opacity .3s linear,top .3s ease-out}.modal.fade.in{top:50%}.modal-header{padding:9px 15px;border-bottom:1px solid #eee}.modal-header .close{margin-top:2px}.modal-header h3{margin:0;line-height:30px}.modal-body{max-height:400px;padding:15px;overflow-y:auto}.modal-form{margin-bottom:0}.modal-footer{padding:14px 15px 15px;margin-bottom:0;text-align:right;background-color:#f5f5f5;border-top:1px solid #ddd;-webkit-border-radius:0 0 6px 6px;-moz-border-radius:0 0 6px 6px;border-radius:0 0 6px 6px;*zoom:1;-webkit-box-shadow:inset 0 1px 0 #fff;-moz-box-shadow:inset 0 1px 0 #fff;box-shadow:inset 0 1px 0 #fff}.modal-footer:before,.modal-footer:after{display:table;line-height:0;content:""}.modal-footer:after{clear:both}.modal-footer .btn+.btn{margin-bottom:0;margin-left:5px}.modal-footer .btn-group .btn+.btn{margin-left:-1px}.tooltip{position:absolute;z-index:1030;display:block;padding:5px;font-size:11px;opacity:0;filter:alpha(opacity=0);visibility:visible}.tooltip.in{opacity:.8;filter:alpha(opacity=80)}.tooltip.top{margin-top:-3px}.tooltip.right{margin-left:3px}.tooltip.bottom{margin-top:3px}.tooltip.left{margin-left:-3px}.tooltip-inner{max-width:200px;padding:3px 8px;color:#fff;text-align:center;text-decoration:none;background-color:#000;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.tooltip-arrow{position:absolute;width:0;height:0;border-color:transparent;border-style:solid}.tooltip.top .tooltip-arrow{bottom:0;left:50%;margin-left:-5px;border-top-color:#000;border-width:5px 5px 0}.tooltip.right .tooltip-arrow{top:50%;left:0;margin-top:-5px;border-right-color:#000;border-width:5px 5px 5px 0}.tooltip.left .tooltip-arrow{top:50%;right:0;margin-top:-5px;border-left-color:#000;border-width:5px 0 5px 5px}.tooltip.bottom .tooltip-arrow{top:0;left:50%;margin-left:-5px;border-bottom-color:#000;border-width:0 5px 5px}.popover{position:absolute;top:0;left:0;z-index:1010;display:none;width:236px;padding:1px;background-color:#fff;border:1px solid #ccc;border:1px solid rgba(0,0,0,0.2);-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px;-webkit-box-shadow:0 5px 10px rgba(0,0,0,0.2);-moz-box-shadow:0 5px 10px rgba(0,0,0,0.2);box-shadow:0 5px 10px rgba(0,0,0,0.2);-webkit-background-clip:padding-box;-moz-background-clip:padding;background-clip:padding-box}.popover.top{margin-bottom:10px}.popover.right{margin-left:10px}.popover.bottom{margin-top:10px}.popover.left{margin-right:10px}.popover-title{padding:8px 14px;margin:0;font-size:14px;font-weight:normal;line-height:18px;background-color:#f7f7f7;border-bottom:1px solid #ebebeb;-webkit-border-radius:5px 5px 0 0;-moz-border-radius:5px 5px 0 0;border-radius:5px 5px 0 0}.popover-content{padding:9px 14px}.popover-content p,.popover-content ul,.popover-content ol{margin-bottom:0}.popover .arrow,.popover .arrow:after{position:absolute;display:inline-block;width:0;height:0;border-color:transparent;border-style:solid}.popover .arrow:after{z-index:-1;content:""}.popover.top .arrow{bottom:-10px;left:50%;margin-left:-10px;border-top-color:#fff;border-width:10px 10px 0}.popover.top .arrow:after{bottom:-1px;left:-11px;border-top-color:rgba(0,0,0,0.25);border-width:11px 11px 0}.popover.right .arrow{top:50%;left:-10px;margin-top:-10px;border-right-color:#fff;border-width:10px 10px 10px 0}.popover.right .arrow:after{bottom:-11px;left:-1px;border-right-color:rgba(0,0,0,0.25);border-width:11px 11px 11px 0}.popover.bottom .arrow{top:-10px;left:50%;margin-left:-10px;border-bottom-color:#fff;border-width:0 10px 10px}.popover.bottom .arrow:after{top:-1px;left:-11px;border-bottom-color:rgba(0,0,0,0.25);border-width:0 11px 11px}.popover.left .arrow{top:50%;right:-10px;margin-top:-10px;border-left-color:#fff;border-width:10px 0 10px 10px}.popover.left .arrow:after{right:-1px;bottom:-11px;border-left-color:rgba(0,0,0,0.25);border-width:11px 0 11px 11px}.thumbnails{margin-left:-20px;list-style:none;*zoom:1}.thumbnails:before,.thumbnails:after{display:table;line-height:0;content:""}.thumbnails:after{clear:both}.row-fluid .thumbnails{margin-left:0}.thumbnails>li{float:left;margin-bottom:20px;margin-left:20px}.thumbnail{display:block;padding:4px;line-height:20px;border:1px solid #ddd;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;-webkit-box-shadow:0 1px 3px rgba(0,0,0,0.055);-moz-box-shadow:0 1px 3px rgba(0,0,0,0.055);box-shadow:0 1px 3px rgba(0,0,0,0.055);-webkit-transition:all .2s ease-in-out;-moz-transition:all .2s ease-in-out;-o-transition:all .2s ease-in-out;transition:all .2s ease-in-out}a.thumbnail:hover{border-color:#08c;-webkit-box-shadow:0 1px 4px rgba(0,105,214,0.25);-moz-box-shadow:0 1px 4px rgba(0,105,214,0.25);box-shadow:0 1px 4px rgba(0,105,214,0.25)}.thumbnail>img{display:block;max-width:100%;margin-right:auto;margin-left:auto}.thumbnail .caption{padding:9px;color:#555}.label,.badge{font-size:11.844px;font-weight:bold;line-height:14px;color:#fff;text-shadow:0 -1px 0 rgba(0,0,0,0.25);white-space:nowrap;vertical-align:baseline;background-color:#999}.label{padding:1px 4px 2px;-webkit-border-radius:3px;-moz-border-radius:3px;border-radius:3px}.badge{padding:1px 9px 2px;-webkit-border-radius:9px;-moz-border-radius:9px;border-radius:9px}a.label:hover,a.badge:hover{color:#fff;text-decoration:none;cursor:pointer}.label-important,.badge-important{background-color:#b94a48}.label-important[href],.badge-important[href]{background-color:#953b39}.label-warning,.badge-warning{background-color:#f89406}.label-warning[href],.badge-warning[href]{background-color:#c67605}.label-success,.badge-success{background-color:#468847}.label-success[href],.badge-success[href]{background-color:#356635}.label-info,.badge-info{background-color:#3a87ad}.label-info[href],.badge-info[href]{background-color:#2d6987}.label-inverse,.badge-inverse{background-color:#333}.label-inverse[href],.badge-inverse[href]{background-color:#1a1a1a}.btn .label,.btn .badge{position:relative;top:-1px}.btn-mini .label,.btn-mini .badge{top:0}@-webkit-keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}@-moz-keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}@-ms-keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}@-o-keyframes progress-bar-stripes{from{background-position:0 0}to{background-position:40px 0}}@keyframes progress-bar-stripes{from{background-position:40px 0}to{background-position:0 0}}.progress{height:20px;margin-bottom:20px;overflow:hidden;background-color:#f7f7f7;background-image:-moz-linear-gradient(top,#f5f5f5,#f9f9f9);background-image:-webkit-gradient(linear,0 0,0 100%,from(#f5f5f5),to(#f9f9f9));background-image:-webkit-linear-gradient(top,#f5f5f5,#f9f9f9);background-image:-o-linear-gradient(top,#f5f5f5,#f9f9f9);background-image:linear-gradient(to bottom,#f5f5f5,#f9f9f9);background-repeat:repeat-x;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fff5f5f5',endColorstr='#fff9f9f9',GradientType=0);-webkit-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1);-moz-box-shadow:inset 0 1px 2px rgba(0,0,0,0.1);box-shadow:inset 0 1px 2px rgba(0,0,0,0.1)}.progress .bar{float:left;width:0;height:100%;font-size:12px;color:#fff;text-align:center;text-shadow:0 -1px 0 rgba(0,0,0,0.25);background-color:#0e90d2;background-image:-moz-linear-gradient(top,#149bdf,#0480be);background-image:-webkit-gradient(linear,0 0,0 100%,from(#149bdf),to(#0480be));background-image:-webkit-linear-gradient(top,#149bdf,#0480be);background-image:-o-linear-gradient(top,#149bdf,#0480be);background-image:linear-gradient(to bottom,#149bdf,#0480be);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff149bdf',endColorstr='#ff0480be',GradientType=0);-webkit-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.15);-moz-box-shadow:inset 0 -1px 0 rgba(0,0,0,0.15);box-shadow:inset 0 -1px 0 rgba(0,0,0,0.15);-webkit-box-sizing:border-box;-moz-box-sizing:border-box;box-sizing:border-box;-webkit-transition:width .6s ease;-moz-transition:width .6s ease;-o-transition:width .6s ease;transition:width .6s ease}.progress .bar+.bar{-webkit-box-shadow:inset 1px 0 0 rgba(0,0,0,0.15),inset 0 -1px 0 rgba(0,0,0,0.15);-moz-box-shadow:inset 1px 0 0 rgba(0,0,0,0.15),inset 0 -1px 0 rgba(0,0,0,0.15);box-shadow:inset 1px 0 0 rgba(0,0,0,0.15),inset 0 -1px 0 rgba(0,0,0,0.15)}.progress-striped .bar{background-color:#149bdf;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);-webkit-background-size:40px 40px;-moz-background-size:40px 40px;-o-background-size:40px 40px;background-size:40px 40px}.progress.active .bar{-webkit-animation:progress-bar-stripes 2s linear infinite;-moz-animation:progress-bar-stripes 2s linear infinite;-ms-animation:progress-bar-stripes 2s linear infinite;-o-animation:progress-bar-stripes 2s linear infinite;animation:progress-bar-stripes 2s linear infinite}.progress-danger .bar,.progress .bar-danger{background-color:#dd514c;background-image:-moz-linear-gradient(top,#ee5f5b,#c43c35);background-image:-webkit-gradient(linear,0 0,0 100%,from(#ee5f5b),to(#c43c35));background-image:-webkit-linear-gradient(top,#ee5f5b,#c43c35);background-image:-o-linear-gradient(top,#ee5f5b,#c43c35);background-image:linear-gradient(to bottom,#ee5f5b,#c43c35);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ffee5f5b',endColorstr='#ffc43c35',GradientType=0)}.progress-danger.progress-striped .bar,.progress-striped .bar-danger{background-color:#ee5f5b;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.progress-success .bar,.progress .bar-success{background-color:#5eb95e;background-image:-moz-linear-gradient(top,#62c462,#57a957);background-image:-webkit-gradient(linear,0 0,0 100%,from(#62c462),to(#57a957));background-image:-webkit-linear-gradient(top,#62c462,#57a957);background-image:-o-linear-gradient(top,#62c462,#57a957);background-image:linear-gradient(to bottom,#62c462,#57a957);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff62c462',endColorstr='#ff57a957',GradientType=0)}.progress-success.progress-striped .bar,.progress-striped .bar-success{background-color:#62c462;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.progress-info .bar,.progress .bar-info{background-color:#4bb1cf;background-image:-moz-linear-gradient(top,#5bc0de,#339bb9);background-image:-webkit-gradient(linear,0 0,0 100%,from(#5bc0de),to(#339bb9));background-image:-webkit-linear-gradient(top,#5bc0de,#339bb9);background-image:-o-linear-gradient(top,#5bc0de,#339bb9);background-image:linear-gradient(to bottom,#5bc0de,#339bb9);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#ff5bc0de',endColorstr='#ff339bb9',GradientType=0)}.progress-info.progress-striped .bar,.progress-striped .bar-info{background-color:#5bc0de;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.progress-warning .bar,.progress .bar-warning{background-color:#faa732;background-image:-moz-linear-gradient(top,#fbb450,#f89406);background-image:-webkit-gradient(linear,0 0,0 100%,from(#fbb450),to(#f89406));background-image:-webkit-linear-gradient(top,#fbb450,#f89406);background-image:-o-linear-gradient(top,#fbb450,#f89406);background-image:linear-gradient(to bottom,#fbb450,#f89406);background-repeat:repeat-x;filter:progid:dximagetransform.microsoft.gradient(startColorstr='#fffbb450',endColorstr='#fff89406',GradientType=0)}.progress-warning.progress-striped .bar,.progress-striped .bar-warning{background-color:#fbb450;background-image:-webkit-gradient(linear,0 100%,100% 0,color-stop(0.25,rgba(255,255,255,0.15)),color-stop(0.25,transparent),color-stop(0.5,transparent),color-stop(0.5,rgba(255,255,255,0.15)),color-stop(0.75,rgba(255,255,255,0.15)),color-stop(0.75,transparent),to(transparent));background-image:-webkit-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-moz-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:-o-linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent);background-image:linear-gradient(45deg,rgba(255,255,255,0.15) 25%,transparent 25%,transparent 50%,rgba(255,255,255,0.15) 50%,rgba(255,255,255,0.15) 75%,transparent 75%,transparent)}.accordion{margin-bottom:20px}.accordion-group{margin-bottom:2px;border:1px solid #e5e5e5;-webkit-border-radius:4px;-moz-border-radius:4px;border-radius:4px}.accordion-heading{border-bottom:0}.accordion-heading .accordion-toggle{display:block;padding:8px 15px}.accordion-toggle{cursor:pointer}.accordion-inner{padding:9px 15px;border-top:1px solid #e5e5e5}.carousel{position:relative;margin-bottom:20px;line-height:1}.carousel-inner{position:relative;width:100%;overflow:hidden}.carousel .item{position:relative;display:none;-webkit-transition:.6s ease-in-out left;-moz-transition:.6s ease-in-out left;-o-transition:.6s ease-in-out left;transition:.6s ease-in-out left}.carousel .item>img{display:block;line-height:1}.carousel .active,.carousel .next,.carousel .prev{display:block}.carousel .active{left:0}.carousel .next,.carousel .prev{position:absolute;top:0;width:100%}.carousel .next{left:100%}.carousel .prev{left:-100%}.carousel .next.left,.carousel .prev.right{left:0}.carousel .active.left{left:-100%}.carousel .active.right{left:100%}.carousel-control{position:absolute;top:40%;left:15px;width:40px;height:40px;margin-top:-20px;font-size:60px;font-weight:100;line-height:30px;color:#fff;text-align:center;background:#222;border:3px solid #fff;-webkit-border-radius:23px;-moz-border-radius:23px;border-radius:23px;opacity:.5;filter:alpha(opacity=50)}.carousel-control.right{right:15px;left:auto}.carousel-control:hover{color:#fff;text-decoration:none;opacity:.9;filter:alpha(opacity=90)}.carousel-caption{position:absolute;right:0;bottom:0;left:0;padding:15px;background:#333;background:rgba(0,0,0,0.75)}.carousel-caption h4,.carousel-caption p{line-height:20px;color:#fff}.carousel-caption h4{margin:0 0 5px}.carousel-caption p{margin-bottom:0}.hero-unit{padding:60px;margin-bottom:30px;background-color:#eee;-webkit-border-radius:6px;-moz-border-radius:6px;border-radius:6px}.hero-unit h1{margin-bottom:0;font-size:60px;line-height:1;letter-spacing:-1px;color:inherit}.hero-unit p{font-size:18px;font-weight:200;line-height:30px;color:inherit}.pull-right{float:right}.pull-left{float:left}.hide{display:none}.show{display:block}.invisible{visibility:hidden}.affix{position:fixed} diff --git a/docs/ec2-scripts.md b/docs/ec2-scripts.md index 931b7a66bd..dc57035eba 100644 --- a/docs/ec2-scripts.md +++ b/docs/ec2-scripts.md @@ -45,9 +45,9 @@ identify machines belonging to each cluster in the Amazon EC2 Console. key pair, `<num-slaves>` is the number of slave nodes to launch (try 1 at first), and `<cluster-name>` is the name to give to your cluster. -- After everything launches, check that Mesos is up and sees all the - slaves by going to the Mesos Web UI link printed at the end of the - script (`http://<master-hostname>:8080`). +- After everything launches, check that the cluster scheduler is up and sees + all the slaves by going to its web UI, which will be printed at the end of + the script (typically `http://<master-hostname>:8080`). You can also run `./spark-ec2 --help` to see more usage options. The following options are worth pointing out: @@ -68,6 +68,9 @@ available. - `--ebs-vol-size=GB` will attach an EBS volume with a given amount of space to each node so that you can have a persistent HDFS cluster on your nodes across cluster restarts (see below). +- `--spot-price=PRICE` will launch the worker nodes as + [Spot Instances](http://aws.amazon.com/ec2/spot-instances/), + bidding for the given maximum price (in dollars). - If one of your launches fails due to e.g. not having the right permissions on your private key file, you can run `launch` with the `--resume` option to restart the setup process on an existing cluster. @@ -80,7 +83,7 @@ permissions on your private key file, you can run `launch` with the above. (This is just for convenience; you could also use the EC2 console.) - To deploy code or data within your cluster, you can log in and use the - provided script `~/mesos-ec2/copy-dir`, which, + provided script `~/spark-ec2/copy-dir`, which, given a directory path, RSYNCs it to the same location on all the slaves. - If your job needs to access large datasets, the fastest way to do that is to load them from Amazon S3 or an Amazon EBS device into an @@ -106,7 +109,7 @@ You can edit `/root/spark/conf/spark-env.sh` on each machine to set Spark config as JVM options and, most crucially, the amount of memory to use per machine (`SPARK_MEM`). This file needs to be copied to **every machine** to reflect the change. The easiest way to do this is to use a script we provide called `copy-dir`. First edit your `spark-env.sh` file on the master, -then run `~/mesos-ec2/copy-dir /root/spark/conf` to RSYNC it to all the workers. +then run `~/spark-ec2/copy-dir /root/spark/conf` to RSYNC it to all the workers. The [configuration guide](configuration.html) describes the available configuration options. @@ -152,10 +155,10 @@ If you have a patch or suggestion for one of these limitations, feel free to # Using a Newer Spark Version -The Spark EC2 machine images may not come with the latest version of Spark. To use a newer version, you can run `git pull` to pull in `/root/spark` to pull in the latest version of Spark from `git`, and build it using `sbt/sbt compile`. You will also need to copy it to all the other nodes in the cluster using `~/mesos-ec2/copy-dir /root/spark`. +The Spark EC2 machine images may not come with the latest version of Spark. To use a newer version, you can run `git pull` to pull in `/root/spark` to pull in the latest version of Spark from `git`, and build it using `sbt/sbt compile`. You will also need to copy it to all the other nodes in the cluster using `~/spark-ec2/copy-dir /root/spark`. # Accessing Data in S3 -Spark's file interface allows it to process data in Amazon S3 using the same URI formats that are supported for Hadoop. You can specify a path in S3 as input through a URI of the form `s3n://<id>:<secret>@<bucket>/path`, where `<id>` is your Amazon access key ID and `<secret>` is your Amazon secret access key. Note that you should escape any `/` characters in the secret key as `%2F`. Full instructions can be found on the [Hadoop S3 page](http://wiki.apache.org/hadoop/AmazonS3). +Spark's file interface allows it to process data in Amazon S3 using the same URI formats that are supported for Hadoop. You can specify a path in S3 as input through a URI of the form `s3n://<bucket>/path`. You will also need to set your Amazon security credentials, either by setting the environment variables `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY` before your program or through `SparkContext.hadoopConfiguration`. Full instructions on S3 access using the Hadoop input libraries can be found on the [Hadoop S3 page](http://wiki.apache.org/hadoop/AmazonS3). In addition to using a single input file, you can also use a directory of files as input by simply giving the path to the directory. diff --git a/docs/index.md b/docs/index.md index c6ef507cb0..0c4add45dc 100644 --- a/docs/index.md +++ b/docs/index.md @@ -3,15 +3,9 @@ layout: global title: Spark Overview --- -{% comment %} -TODO(andyk): Rewrite to make the Java API a first class part of the story. -{% endcomment %} - Spark is a MapReduce-like cluster computing framework designed for low-latency iterative jobs and interactive use from an interpreter. It provides clean, language-integrated APIs in [Scala](scala-programming-guide.html), [Java](java-programming-guide.html), and [Python](python-programming-guide.html), with a rich array of parallel operators. -Spark can run on top of the [Apache Mesos](http://incubator.apache.org/mesos/) cluster manager, -[Hadoop YARN](http://hadoop.apache.org/docs/r2.0.1-alpha/hadoop-yarn/hadoop-yarn-site/YARN.html), -Amazon EC2, or without an independent resource manager ("standalone mode"). +Spark can run on the Apache Mesos cluster manager, Hadoop YARN, Amazon EC2, or without an independent resource manager ("standalone mode"). # Downloading @@ -24,10 +18,12 @@ or you will need to set the `SCALA_HOME` environment variable to point to where you've installed Scala. Scala must also be accessible through one of these methods on slave nodes on your cluster. -Spark uses [Simple Build Tool](https://github.com/harrah/xsbt/wiki), which is bundled with it. To compile the code, go into the top-level Spark directory and run +Spark uses [Simple Build Tool](http://www.scala-sbt.org), which is bundled with it. To compile the code, go into the top-level Spark directory and run sbt/sbt package +Spark also supports building using Maven. If you would like to build using Maven, see the [instructions for building Spark with Maven](building-with-maven.html). + # Testing the Build Spark comes with a number of sample programs in the `examples` directory. @@ -58,9 +54,9 @@ of `project/SparkBuild.scala`, then rebuilding Spark (`sbt/sbt clean compile`). * [Quick Start](quick-start.html): a quick introduction to the Spark API; start here! * [Spark Programming Guide](scala-programming-guide.html): an overview of Spark concepts, and details on the Scala API -* [Streaming Programming Guide](streaming-programming-guide.html): an API preview of Spark Streaming * [Java Programming Guide](java-programming-guide.html): using Spark from Java * [Python Programming Guide](python-programming-guide.html): using Spark from Python +* [Spark Streaming Guide](streaming-programming-guide.html): using the alpha release of Spark Streaming **API Docs:** @@ -78,6 +74,7 @@ of `project/SparkBuild.scala`, then rebuilding Spark (`sbt/sbt clean compile`). **Other documents:** +* [Building Spark With Maven](building-with-maven.html): Build Spark using the Maven build tool * [Configuration](configuration.html): customize Spark via its configuration system * [Tuning Guide](tuning.html): best practices to optimize performance and memory use * [Bagel](bagel-programming-guide.html): an implementation of Google's Pregel on Spark @@ -92,7 +89,8 @@ of `project/SparkBuild.scala`, then rebuilding Spark (`sbt/sbt clean compile`). [slides](http://ampcamp.berkeley.edu/agenda-2012) and [exercises](http://ampcamp.berkeley.edu/exercises-2012) are available online for free. * [Code Examples](http://spark-project.org/examples.html): more are also available in the [examples subfolder](https://github.com/mesos/spark/tree/master/examples/src/main/scala/spark/examples) of Spark -* [Paper Describing the Spark System](http://www.cs.berkeley.edu/~matei/papers/2012/nsdi_spark.pdf) +* [Paper Describing Spark](http://www.cs.berkeley.edu/~matei/papers/2012/nsdi_spark.pdf) +* [Paper Describing Spark Streaming](http://www.eecs.berkeley.edu/Pubs/TechRpts/2012/EECS-2012-259.pdf) # Community diff --git a/docs/java-programming-guide.md b/docs/java-programming-guide.md index 37a906ea1c..ae8257b539 100644 --- a/docs/java-programming-guide.md +++ b/docs/java-programming-guide.md @@ -189,7 +189,7 @@ We hope to generate documentation with Java-style syntax in the future. # Where to Go from Here Spark includes several sample programs using the Java API in -`examples/src/main/java`. You can run them by passing the class name to the +[`examples/src/main/java`](https://github.com/mesos/spark/tree/master/examples/src/main/java/spark/examples). You can run them by passing the class name to the `run` script included in Spark -- for example, `./run spark.examples.JavaWordCount`. Each example program prints usage help when run without any arguments. diff --git a/docs/python-programming-guide.md b/docs/python-programming-guide.md index 4e84d23edf..3a7a8db4a6 100644 --- a/docs/python-programming-guide.md +++ b/docs/python-programming-guide.md @@ -87,7 +87,7 @@ By default, the `pyspark` shell creates SparkContext that runs jobs locally. To connect to a non-local cluster, set the `MASTER` environment variable. For example, to use the `pyspark` shell with a [standalone Spark cluster](spark-standalone.html): -{% highlight shell %} +{% highlight bash %} $ MASTER=spark://IP:PORT ./pyspark {% endhighlight %} @@ -109,9 +109,9 @@ Code dependencies can be added to an existing SparkContext using its `addPyFile( # Where to Go from Here -PySpark includes several sample programs using the Python API in `python/examples`. +PySpark includes several sample programs in the [`python/examples` folder](https://github.com/mesos/spark/tree/master/python/examples). You can run them by passing the files to the `pyspark` script -- for example `./pyspark python/examples/wordcount.py`. -Each example program prints usage help when run without any arguments. +Each program prints usage help when run without arguments. We currently provide [API documentation](api/pyspark/index.html) for the Python API as Epydoc. Many of the RDD method descriptions contain [doctests](http://docs.python.org/2/library/doctest.html) that provide additional usage examples. diff --git a/docs/quick-start.md b/docs/quick-start.md index a4c4c9a8fb..335643536a 100644 --- a/docs/quick-start.md +++ b/docs/quick-start.md @@ -53,8 +53,8 @@ scala> textFile.filter(line => line.contains("Spark")).count() // How many lines res3: Long = 15 {% endhighlight %} -## Transformations -RDD transformations can be used for more complex computations. Let's say we want to find the line with the most words: +## More On RDD Operations +RDD actions and transformations can be used for more complex computations. Let's say we want to find the line with the most words: {% highlight scala %} scala> textFile.map(line => line.split(" ").size).reduce((a, b) => if (a > b) a else b) @@ -111,18 +111,20 @@ We'll create a very simple Spark job in Scala. So simple, in fact, that it's nam import spark.SparkContext import SparkContext._ -object SimpleJob extends Application { - val logFile = "/var/log/syslog" // Should be some file on your system - val sc = new SparkContext("local", "Simple Job", "$YOUR_SPARK_HOME", - List("target/scala-{{site.SCALA_VERSION}}/simple-project_{{site.SCALA_VERSION}}-1.0.jar")) - val logData = sc.textFile(logFile, 2).cache() - val numAs = logData.filter(line => line.contains("a")).count() - val numBs = logData.filter(line => line.contains("b")).count() - println("Lines with a: %s, Lines with b: %s".format(numAs, numBs)) +object SimpleJob { + def main(args: Array[String]) { + val logFile = "$YOUR_SPARK_HOME/README.md" // Should be some file on your system + val sc = new SparkContext("local", "Simple Job", "YOUR_SPARK_HOME", + List("target/scala-{{site.SCALA_VERSION}}/simple-project_{{site.SCALA_VERSION}}-1.0.jar")) + val logData = sc.textFile(logFile, 2).cache() + val numAs = logData.filter(line => line.contains("a")).count() + val numBs = logData.filter(line => line.contains("b")).count() + println("Lines with a: %s, Lines with b: %s".format(numAs, numBs)) + } } {% endhighlight %} -This job simply counts the number of lines containing 'a' and the number containing 'b' in a system log file. Unlike the earlier examples with the Spark shell, which initializes its own SparkContext, we initialize a SparkContext as part of the job. We pass the SparkContext constructor four arguments, the type of scheduler we want to use (in this case, a local scheduler), a name for the job, the directory where Spark is installed, and a name for the jar file containing the job's sources. The final two arguments are needed in a distributed setting, where Spark is running across several nodes, so we include them for completeness. Spark will automatically ship the jar files you list to slave nodes. +This job simply counts the number of lines containing 'a' and the number containing 'b' in the Spark README. Note that you'll need to replace $YOUR_SPARK_HOME with the location where Spark is installed. Unlike the earlier examples with the Spark shell, which initializes its own SparkContext, we initialize a SparkContext as part of the job. We pass the SparkContext constructor four arguments, the type of scheduler we want to use (in this case, a local scheduler), a name for the job, the directory where Spark is installed, and a name for the jar file containing the job's sources. The final two arguments are needed in a distributed setting, where Spark is running across several nodes, so we include them for completeness. Spark will automatically ship the jar files you list to slave nodes. This file depends on the Spark API, so we'll also include an sbt configuration file, `simple.sbt` which explains that Spark is a dependency. This file also adds two repositories which host Spark dependencies: @@ -136,7 +138,7 @@ scalaVersion := "{{site.SCALA_VERSION}}" libraryDependencies += "org.spark-project" %% "spark-core" % "{{site.SPARK_VERSION}}" resolvers ++= Seq( - "Typesafe Repository" at "http://repo.typesafe.com/typesafe/releases/", + "Akka Repository" at "http://repo.akka.io/releases/", "Spray Repository" at "http://repo.spray.cc/") {% endhighlight %} @@ -154,7 +156,7 @@ $ find . $ sbt package $ sbt run ... -Lines with a: 8422, Lines with b: 1836 +Lines with a: 46, Lines with b: 23 {% endhighlight %} This example only runs the job locally; for a tutorial on running jobs across several machines, see the [Standalone Mode](spark-standalone.html) documentation, and consider using a distributed input source, such as HDFS. @@ -171,7 +173,7 @@ import spark.api.java.function.Function; public class SimpleJob { public static void main(String[] args) { - String logFile = "/var/log/syslog"; // Should be some file on your system + String logFile = "$YOUR_SPARK_HOME/README.md"; // Should be some file on your system JavaSparkContext sc = new JavaSparkContext("local", "Simple Job", "$YOUR_SPARK_HOME", new String[]{"target/simple-project-1.0.jar"}); JavaRDD<String> logData = sc.textFile(logFile).cache(); @@ -189,7 +191,7 @@ public class SimpleJob { } {% endhighlight %} -This job simply counts the number of lines containing 'a' and the number containing 'b' in a system log file. Note that like in the Scala example, we initialize a SparkContext, though we use the special `JavaSparkContext` class to get a Java-friendly one. We also create RDDs (represented by `JavaRDD`) and run transformations on them. Finally, we pass functions to Spark by creating classes that extend `spark.api.java.function.Function`. The [Java programming guide]("java-programming-guide") describes these differences in more detail. +This job simply counts the number of lines containing 'a' and the number containing 'b' in a system log file. Note that you'll need to replace $YOUR_SPARK_HOME with the location where Spark is installed. As with the Scala example, we initialize a SparkContext, though we use the special `JavaSparkContext` class to get a Java-friendly one. We also create RDDs (represented by `JavaRDD`) and run transformations on them. Finally, we pass functions to Spark by creating classes that extend `spark.api.java.function.Function`. The [Java programming guide](java-programming-guide.html) describes these differences in more detail. To build the job, we also write a Maven `pom.xml` file that lists Spark as a dependency. Note that Spark artifacts are tagged with a Scala version. @@ -207,8 +209,8 @@ To build the job, we also write a Maven `pom.xml` file that lists Spark as a dep <url>http://repo.spray.cc</url> </repository> <repository> - <id>Typesafe repository</id> - <url>http://repo.typesafe.com/typesafe/releases</url> + <id>Akka repository</id> + <url>http://repo.akka.io/releases</url> </repository> </repositories> <dependencies> @@ -237,7 +239,7 @@ Now, we can execute the job using Maven: $ mvn package $ mvn exec:java -Dexec.mainClass="SimpleJob" ... -Lines with a: 8422, Lines with b: 1836 +Lines with a: 46, Lines with b: 23 {% endhighlight %} This example only runs the job locally; for a tutorial on running jobs across several machines, see the [Standalone Mode](spark-standalone.html) documentation, and consider using a distributed input source, such as HDFS. @@ -251,7 +253,7 @@ As an example, we'll create a simple Spark job, `SimpleJob.py`: """SimpleJob.py""" from pyspark import SparkContext -logFile = "/var/log/syslog" # Should be some file on your system +logFile = "$YOUR_SPARK_HOME/README.md" # Should be some file on your system sc = SparkContext("local", "Simple job") logData = sc.textFile(logFile).cache() @@ -263,9 +265,10 @@ print "Lines with a: %i, lines with b: %i" % (numAs, numBs) This job simply counts the number of lines containing 'a' and the number containing 'b' in a system log file. -Like in the Scala and Java examples, we use a SparkContext to create RDDs. +Note that you'll need to replace $YOUR_SPARK_HOME with the location where Spark is installed. +As with the Scala and Java examples, we use a SparkContext to create RDDs. We can pass Python functions to Spark, which are automatically serialized along with any variables that they reference. -For jobs that use custom classes or third-party libraries, we can add those code dependencies to SparkContext to ensure that they will be available on remote machines; this is described in more detail in the [Python programming guide](python-programming-guide). +For jobs that use custom classes or third-party libraries, we can add those code dependencies to SparkContext to ensure that they will be available on remote machines; this is described in more detail in the [Python programming guide](python-programming-guide.html). `SimpleJob` is simple enough that we do not need to specify any code dependencies. We can run this job using the `pyspark` script: @@ -274,7 +277,7 @@ We can run this job using the `pyspark` script: $ cd $SPARK_HOME $ ./pyspark SimpleJob.py ... -Lines with a: 8422, Lines with b: 1836 +Lines with a: 46, Lines with b: 23 {% endhighlight python %} This example only runs the job locally; for a tutorial on running jobs across several machines, see the [Standalone Mode](spark-standalone.html) documentation, and consider using a distributed input source, such as HDFS. diff --git a/docs/running-on-yarn.md b/docs/running-on-yarn.md index 6fb81b6004..66fb8d73e8 100644 --- a/docs/running-on-yarn.md +++ b/docs/running-on-yarn.md @@ -4,25 +4,55 @@ title: Launching Spark on YARN --- Experimental support for running over a [YARN (Hadoop -NextGen)](http://hadoop.apache.org/docs/r2.0.1-alpha/hadoop-yarn/hadoop-yarn-site/YARN.html) -cluster was added to Spark in version 0.6.0. Because YARN depends on version -2.0 of the Hadoop libraries, this currently requires checking out a separate -branch of Spark, called `yarn`, which you can do as follows: +NextGen)](http://hadoop.apache.org/docs/r2.0.2-alpha/hadoop-yarn/hadoop-yarn-site/YARN.html) +cluster was added to Spark in version 0.6.0. This was merged into master as part of 0.7 effort. +To build spark core with YARN support, please use the hadoop2-yarn profile. +Ex: mvn -Phadoop2-yarn clean install - git clone git://github.com/mesos/spark - cd spark - git checkout -b yarn --track origin/yarn +# Building spark core consolidated jar. + +We need a consolidated spark core jar (which bundles all the required dependencies) to run Spark jobs on a yarn cluster. +This can be built either through sbt or via maven. + +- Building spark assembled jar via sbt. + It is a manual process of enabling it in project/SparkBuild.scala. +Please comment out the + HADOOP_VERSION, HADOOP_MAJOR_VERSION and HADOOP_YARN +variables before the line 'For Hadoop 2 YARN support' +Next, uncomment the subsequent 3 variable declaration lines (for these three variables) which enable hadoop yarn support. + +Assembly of the jar Ex: + + ./sbt/sbt clean assembly + +The assembled jar would typically be something like : +`./core/target/spark-core-assembly-0.8.0-SNAPSHOT.jar` + + +- Building spark assembled jar via Maven. + Use the hadoop2-yarn profile and execute the package target. + +Something like this. Ex: + + mvn -Phadoop2-yarn clean package -DskipTests=true + + +This will build the shaded (consolidated) jar. Typically something like : +`./repl-bin/target/spark-repl-bin-<VERSION>-shaded-hadoop2-yarn.jar` # Preparations -- In order to distribute Spark within the cluster, it must be packaged into a single JAR file. This can be done by running `sbt/sbt assembly` +- Building spark core assembled jar (see above). - Your application code must be packaged into a separate JAR file. If you want to test out the YARN deployment mode, you can use the current Spark examples. A `spark-examples_{{site.SCALA_VERSION}}-{{site.SPARK_VERSION}}` file can be generated by running `sbt/sbt package`. NOTE: since the documentation you're reading is for Spark version {{site.SPARK_VERSION}}, we are assuming here that you have downloaded Spark {{site.SPARK_VERSION}} or checked it out of source control. If you are using a different version of Spark, the version numbers in the jar generated by the sbt package command will obviously be different. # Launching Spark on YARN +Ensure that HADOOP_CONF_DIR or YARN_CONF_DIR points to the directory which contains the (client side) configuration files for the hadoop cluster. +This would be used to connect to the cluster, write to the dfs and submit jobs to the resource manager. + The command to launch the YARN Client is as follows: SPARK_JAR=<SPARK_YAR_FILE> ./run spark.deploy.yarn.Client \ @@ -30,22 +60,28 @@ The command to launch the YARN Client is as follows: --class <APP_MAIN_CLASS> \ --args <APP_MAIN_ARGUMENTS> \ --num-workers <NUMBER_OF_WORKER_MACHINES> \ + --master-memory <MEMORY_FOR_MASTER> \ --worker-memory <MEMORY_PER_WORKER> \ - --worker-cores <CORES_PER_WORKER> + --worker-cores <CORES_PER_WORKER> \ + --user <hadoop_user> \ + --queue <queue_name> For example: SPARK_JAR=./core/target/spark-core-assembly-{{site.SPARK_VERSION}}.jar ./run spark.deploy.yarn.Client \ --jar examples/target/scala-{{site.SCALA_VERSION}}/spark-examples_{{site.SCALA_VERSION}}-{{site.SPARK_VERSION}}.jar \ --class spark.examples.SparkPi \ - --args standalone \ + --args yarn-standalone \ --num-workers 3 \ + --master-memory 4g \ --worker-memory 2g \ - --worker-cores 2 + --worker-cores 1 The above starts a YARN Client programs which periodically polls the Application Master for status updates and displays them in the console. The client will exit once your application has finished running. # Important Notes -- When your application instantiates a Spark context it must use a special "standalone" master url. This starts the scheduler without forcing it to connect to a cluster. A good way to handle this is to pass "standalone" as an argument to your program, as shown in the example above. -- YARN does not support requesting container resources based on the number of cores. Thus the numbers of cores given via command line arguments cannot be guaranteed. +- When your application instantiates a Spark context it must use a special "yarn-standalone" master url. This starts the scheduler without forcing it to connect to a cluster. A good way to handle this is to pass "yarn-standalone" as an argument to your program, as shown in the example above. +- We do not requesting container resources based on the number of cores. Thus the numbers of cores given via command line arguments cannot be guaranteed. +- Currently, we have not yet integrated with hadoop security. If --user is present, the hadoop_user specified will be used to run the tasks on the cluster. If unspecified, current user will be used (which should be valid in cluster). + Once hadoop security support is added, and if hadoop cluster is enabled with security, additional restrictions would apply via delegation tokens passed. diff --git a/docs/scala-programming-guide.md b/docs/scala-programming-guide.md index b98718a553..e9cf9ef36f 100644 --- a/docs/scala-programming-guide.md +++ b/docs/scala-programming-guide.md @@ -38,17 +38,23 @@ The first thing a Spark program must do is to create a `SparkContext` object, wh This is done through the following constructor: {% highlight scala %} -new SparkContext(master, jobName, [sparkHome], [jars]) +new SparkContext(master, appName, [sparkHome], [jars]) {% endhighlight %} -The `master` parameter is a string specifying a [Mesos](running-on-mesos.html) cluster to connect to, or a special "local" string to run in local mode, as described below. `jobName` is a name for your job, which will be shown in the Mesos web UI when running on a cluster. Finally, the last two parameters are needed to deploy your code to a cluster if running in distributed mode, as described later. +The `master` parameter is a string specifying a [Spark or Mesos cluster URL](#master-urls) to connect to, or a special "local" string to run in local mode, as described below. `appName` is a name for your application, which will be shown in the cluster web UI. Finally, the last two parameters are needed to deploy your code to a cluster if running in distributed mode, as described later. -In the Spark shell, a special interpreter-aware SparkContext is already created for you, in the variable called `sc`. Making your own SparkContext will not work. You can set which master the context connects to using the `MASTER` environment variable. For example, to run on four cores, use +In the Spark shell, a special interpreter-aware SparkContext is already created for you, in the variable called `sc`. Making your own SparkContext will not work. You can set which master the context connects to using the `MASTER` environment variable, and you can add JARs to the classpath with the `ADD_JARS` variable. For example, to run `spark-shell` on four cores, use {% highlight bash %} $ MASTER=local[4] ./spark-shell {% endhighlight %} +Or, to also add `code.jar` to its classpath, use: + +{% highlight bash %} +$ MASTER=local[4] ADD_JARS=code.jar ./spark-shell +{% endhighlight %} + ### Master URLs The master URL passed to Spark can be in one of the following formats: @@ -67,6 +73,8 @@ The master URL passed to Spark can be in one of the following formats: </td></tr> </table> +If no master URL is specified, the spark shell defaults to "local". + For running on YARN, Spark launches an instance of the standalone deploy cluster within YARN; see [running on YARN](running-on-yarn.html) for details. ### Deploying Code on a Cluster @@ -76,7 +84,7 @@ If you want to run your job on a cluster, you will need to specify the two optio * `sparkHome`: The path at which Spark is installed on your worker machines (it should be the same on all of them). * `jars`: A list of JAR files on the local machine containing your job's code and any dependencies, which Spark will deploy to all the worker nodes. You'll need to package your job into a set of JARs using your build system. For example, if you're using SBT, the [sbt-assembly](https://github.com/sbt/sbt-assembly) plugin is a good way to make a single JAR with your code and dependencies. -If you run `spark-shell` on a cluster, any classes you define in the shell will automatically be distributed. +If you run `spark-shell` on a cluster, you can add JARs to it by specifying the `ADD_JARS` environment variable before you launch it. This variable should contain a comma-separated list of JARs. For example, `ADD_JARS=a.jar,b.jar ./spark-shell` will launch a shell with `a.jar` and `b.jar` on its classpath. In addition, any new classes you define in the shell will automatically be distributed. # Resilient Distributed Datasets (RDDs) diff --git a/docs/streaming-custom-receivers.md b/docs/streaming-custom-receivers.md new file mode 100644 index 0000000000..5476c00d02 --- /dev/null +++ b/docs/streaming-custom-receivers.md @@ -0,0 +1,101 @@ +--- +layout: global +title: Tutorial - Spark Streaming, Plugging in a custom receiver. +--- + +A "Spark Streaming" receiver can be a simple network stream, streams of messages from a message queue, files etc. A receiver can also assume roles more than just receiving data like filtering, preprocessing, to name a few of the possibilities. The api to plug-in any user defined custom receiver is thus provided to encourage development of receivers which may be well suited to ones specific need. + +This guide shows the programming model and features by walking through a simple sample receiver and corresponding Spark Streaming application. + + +## A quick and naive walk-through + +### Write a simple receiver + +This starts with implementing [Actor](#References) + +Following is a simple socket text-stream receiver, which is appearently overly simplified using Akka's socket.io api. + +{% highlight scala %} + + class SocketTextStreamReceiver (host:String, + port:Int, + bytesToString: ByteString => String) extends Actor with Receiver { + + override def preStart = IOManager(context.system).connect(host, port) + + def receive = { + case IO.Read(socket, bytes) => pushBlock(bytesToString(bytes)) + } + + } + + +{% endhighlight %} + +All we did here is mixed in trait Receiver and called pushBlock api method to push our blocks of data. Please refer to scala-docs of Receiver for more details. + +### A sample spark application + +* First create a Spark streaming context with master url and batchduration. + +{% highlight scala %} + + val ssc = new StreamingContext(master, "WordCountCustomStreamSource", + Seconds(batchDuration)) + +{% endhighlight %} + +* Plug-in the actor configuration into the spark streaming context and create a DStream. + +{% highlight scala %} + + val lines = ssc.actorStream[String](Props(new SocketTextStreamReceiver( + "localhost",8445, z => z.utf8String)),"SocketReceiver") + +{% endhighlight %} + +* Process it. + +{% highlight scala %} + + val words = lines.flatMap(_.split(" ")) + val wordCounts = words.map(x => (x, 1)).reduceByKey(_ + _) + + wordCounts.print() + ssc.start() + + +{% endhighlight %} + +* After processing it, stream can be tested using the netcat utility. + + $ nc -l localhost 8445 + hello world + hello hello + + +## Multiple homogeneous/heterogeneous receivers. + +A DStream union operation is provided for taking union on multiple input streams. + +{% highlight scala %} + + val lines = ssc.actorStream[String](Props(new SocketTextStreamReceiver( + "localhost",8445, z => z.utf8String)),"SocketReceiver") + + // Another socket stream receiver + val lines2 = ssc.actorStream[String](Props(new SocketTextStreamReceiver( + "localhost",8446, z => z.utf8String)),"SocketReceiver") + + val union = lines.union(lines2) + +{% endhighlight %} + +Above stream can be easily process as described earlier. + +_A more comprehensive example is provided in the spark streaming examples_ + +## References + +1.[Akka Actor documentation](http://doc.akka.io/docs/akka/2.0.5/scala/actors.html) diff --git a/docs/streaming-programming-guide.md b/docs/streaming-programming-guide.md index b6da7af654..f5788dc467 100644 --- a/docs/streaming-programming-guide.md +++ b/docs/streaming-programming-guide.md @@ -7,9 +7,9 @@ title: Spark Streaming Programming Guide {:toc} # Overview -A Spark Streaming application is very similar to a Spark application; it consists of a *driver program* that runs the user's `main` function and continuous executes various *parallel operations* on input streams of data. The main abstraction Spark Streaming provides is a *discretized stream* (DStream), which is a continuous sequence of RDDs (distributed collection of elements) representing a continuous stream of data. DStreams can created from live incoming data (such as data from a socket, Kafka, etc.) or it can be generated by transformation of existing DStreams using parallel operators like map, reduce, and window. The basic processing model is as follows: -(i) While a Spark Streaming driver program is running, the system receives data from various sources and and divides the data into batches. Each batch of data is treated as a RDD, that is a immutable and parallel collection of data. These input data RDDs are automatically persisted in memory (serialized by default) and replicated to two nodes for fault-tolerance. This sequence of RDDs is collectively referred to as an InputDStream. -(ii) Data received by InputDStreams are processed processed using DStream operations. Since all data is represented as RDDs and all DStream operations as RDD operations, data is automatically recovered in the event of node failures. +A Spark Streaming application is very similar to a Spark application; it consists of a *driver program* that runs the user's `main` function and continuous executes various *parallel operations* on input streams of data. The main abstraction Spark Streaming provides is a *discretized stream* (DStream), which is a continuous sequence of RDDs (distributed collections of elements) representing a continuous stream of data. DStreams can be created from live incoming data (such as data from a socket, Kafka, etc.) or can be generated by transformong existing DStreams using parallel operators like `map`, `reduce`, and `window`. The basic processing model is as follows: +(i) While a Spark Streaming driver program is running, the system receives data from various sources and and divides it into batches. Each batch of data is treated as an RDD, that is, an immutable parallel collection of data. These input RDDs are saved in memory and replicated to two nodes for fault-tolerance. This sequence of RDDs is collectively called an InputDStream. +(ii) Data received by InputDStreams are processed using DStream operations. Since all data is represented as RDDs and all DStream operations as RDD operations, data is automatically recovered in the event of node failures. This guide shows some how to start programming with DStreams. @@ -17,16 +17,12 @@ This guide shows some how to start programming with DStreams. The first thing a Spark Streaming program must do is create a `StreamingContext` object, which tells Spark how to access a cluster. A `StreamingContext` can be created by using {% highlight scala %} -new StreamingContext(master, jobName, batchDuration) +new StreamingContext(master, appName, batchDuration, [sparkHome], [jars]) {% endhighlight %} -The `master` parameter is either the [Mesos master URL](running-on-mesos.html) (for running on a cluster)or the special "local" string (for local mode) that is used to create a Spark Context. For more information about this please refer to the [Spark programming guide](scala-programming-guide.html). The `jobName` is the name of the streaming job, which is the same as the jobName used in SparkContext. It is used to identify this job in the Mesos web UI. The `batchDuration` is the size of the batches (as explained earlier). This must be set carefully such the cluster can keep up with the processing of the data streams. Starting with something conservative like 5 seconds maybe a good start. See [Performance Tuning](#setting-the-right-batch-size) section for a detailed discussion. - -This constructor creates a SparkContext object using the given `master` and `jobName` parameters. However, if you already have a SparkContext or you need to create a custom SparkContext by specifying list of JARs, then a StreamingContext can be created from the existing SparkContext, by using -{% highlight scala %} -new StreamingContext(sparkContext, batchDuration) -{% endhighlight %} +The `master` parameter is a standard [Spark cluster URL](scala-programming-guide.html#master-urls) and can be "local" for local testing. The `appName` is a name of your program, which will be shown on your cluster's web UI. The `batchDuration` is the size of the batches (as explained earlier). This must be set carefully such the cluster can keep up with the processing of the data streams. Start with something conservative like 5 seconds. See the [Performance Tuning](#setting-the-right-batch-size) section for a detailed discussion. Finally, `sparkHome` and `jars` are necessary when running on a cluster to specify the location of your code, as described in the [Spark programming guide](scala-programming-guide.html#deploying-code-on-a-cluster). +This constructor creates a SparkContext for your job as well, which can be accessed with `streamingContext.sparkContext`. # Attaching Input Sources - InputDStreams @@ -34,34 +30,34 @@ The StreamingContext is used to creating InputDStreams from input sources: {% highlight scala %} // Assuming ssc is the StreamingContext -ssc.networkStream(hostname, port) // Creates a stream that uses a TCP socket to read data from hostname:port -ssc.textFileStream(directory) // Creates a stream by monitoring and processing new files in a HDFS directory +ssc.textFileStream(directory) // Creates a stream by monitoring and processing new files in a HDFS directory +ssc.socketStream(hostname, port) // Creates a stream that uses a TCP socket to read data from hostname:port {% endhighlight %} -A complete list of input sources is available in the [StreamingContext API documentation](api/streaming/index.html#spark.streaming.StreamingContext). Data received from these sources can be processed using DStream operations, which are explained next. +We also provide a input streams for Kafka, Flume, Akka actor, etc. For a complete list of input streams, take a look at the [StreamingContext API documentation](api/streaming/index.html#spark.streaming.StreamingContext). # DStream Operations -Once an input DStream has been created, you can transform it using _DStream operators_. Most of these operators return new DStreams which you can further transform. Eventually, you'll need to call an _output operator_, which forces evaluation of the DStream by writing data out to an external source. +Data received from the input streams can be processed using _DStream operations_. There are two kinds of operations - _transformations_ and _output operations_. Similar to RDD transformations, DStream transformations operate on one or more DStreams to create new DStreams with transformed data. After applying a sequence of transformations to the input streams, you'll need to call the output operations, which writies data out to an external source. ## Transformations DStreams support many of the transformations available on normal Spark RDD's: <table class="table"> -<tr><th style="width:25%">Transformation</th><th>Meaning</th></tr> +<tr><th style="width:30%">Transformation</th><th>Meaning</th></tr> <tr> <td> <b>map</b>(<i>func</i>) </td> - <td> Returns a new DStream formed by passing each element of the source through a function <i>func</i>. </td> + <td> Returns a new DStream formed by passing each element of the source DStream through a function <i>func</i>. </td> </tr> <tr> <td> <b>filter</b>(<i>func</i>) </td> - <td> Returns a new stream formed by selecting those elements of the source on which <i>func</i> returns true. </td> + <td> Returns a new DStream formed by selecting those elements of the source DStream on which <i>func</i> returns true. </td> </tr> <tr> <td> <b>flatMap</b>(<i>func</i>) </td> - <td> Similar to map, but each input item can be mapped to 0 or more output items (so <i>func</i> should return a Seq rather than a single item). </td> + <td> Similar to map, but each input item can be mapped to 0 or more output items (so <i>func</i> should return a <code>Seq</code> rather than a single item). </td> </tr> <tr> <td> <b>mapPartitions</b>(<i>func</i>) </td> @@ -70,73 +66,92 @@ DStreams support many of the transformations available on normal Spark RDD's: </tr> <tr> <td> <b>union</b>(<i>otherStream</i>) </td> - <td> Return a new stream that contains the union of the elements in the source stream and the argument. </td> + <td> Return a new DStream that contains the union of the elements in the source DStream and the argument DStream. </td> +</tr> +<tr> + <td> <b>count</b>() </td> + <td> Returns a new DStream of single-element RDDs by counting the number of elements in each RDD of the source DStream. </td> +</tr> +<tr> + <td> <b>reduce</b>(<i>func</i>) </td> + <td> Returns a new DStream of single-element RDDs by aggregating the elements in each RDD of the source DStream using a function <i>func</i> (which takes two arguments and returns one). The function should be associative so that it can be computed in parallel. </td> +</tr> +<tr> + <td> <b>countByValue</b>() </td> + <td> When called on a DStream of elements of type K, returns a new DStream of (K, Long) pairs where the value of each key is its frequency in each RDD of the source DStream. </td> </tr> <tr> <td> <b>groupByKey</b>([<i>numTasks</i>]) </td> - <td> When called on a stream of (K, V) pairs, returns a stream of (K, Seq[V]) pairs. <br /> -<b>Note:</b> By default, this uses only 8 parallel tasks to do the grouping. You can pass an optional <code>numTasks</code> argument to set a different number of tasks. + <td> When called on a DStream of (K, V) pairs, returns a new DStream of (K, Seq[V]) pairs by grouping together all the values of each key in the RDDs of the source DStream. <br /> + <b>Note:</b> By default, this uses Spark's default number of parallel tasks (2 for local machine, 8 for a cluster) to do the grouping. You can pass an optional <code>numTasks</code> argument to set a different number of tasks. </td> </tr> <tr> <td> <b>reduceByKey</b>(<i>func</i>, [<i>numTasks</i>]) </td> - <td> When called on a stream of (K, V) pairs, returns a stream of (K, V) pairs where the values for each key are aggregated using the given reduce function. Like in <code>groupByKey</code>, the number of reduce tasks is configurable through an optional second argument. </td> + <td> When called on a DStream of (K, V) pairs, returns a new DStream of (K, V) pairs where the values for each key are aggregated using the given reduce function. Like in <code>groupByKey</code>, the number of reduce tasks is configurable through an optional second argument. </td> </tr> <tr> <td> <b>join</b>(<i>otherStream</i>, [<i>numTasks</i>]) </td> - <td> When called on streams of type (K, V) and (K, W), returns a stream of (K, (V, W)) pairs with all pairs of elements for each key. </td> + <td> When called on two DStreams of (K, V) and (K, W) pairs, returns a new DStream of (K, (V, W)) pairs with all pairs of elements for each key. </td> </tr> <tr> <td> <b>cogroup</b>(<i>otherStream</i>, [<i>numTasks</i>]) </td> - <td> When called on DStream of type (K, V) and (K, W), returns a DStream of (K, Seq[V], Seq[W]) tuples.</td> -</tr> -<tr> - <td> <b>reduce</b>(<i>func</i>) </td> - <td> Returns a new DStream of single-element RDDs by aggregating the elements of the stream using a function func (which takes two arguments and returns one). The function should be associative so that it can be computed correctly in parallel. </td> + <td> When called on DStream of (K, V) and (K, W) pairs, returns a new DStream of (K, Seq[V], Seq[W]) tuples.</td> </tr> <tr> <td> <b>transform</b>(<i>func</i>) </td> <td> Returns a new DStream by applying func (a RDD-to-RDD function) to every RDD of the stream. This can be used to do arbitrary RDD operations on the DStream. </td> </tr> +<tr> + <td> <b>updateStateByKey</b>(<i>func</i>) </td> + <td> Return a new "state" DStream where the state for each key is updated by applying the given function on the previous state of the key and the new values of each key. This can be used to track session state by using the session-id as the key and updating the session state as new data is received.</td> +</tr> + </table> -Spark Streaming features windowed computations, which allow you to report statistics over a sliding window of data. All window functions take a <i>windowDuration</i>, which represents the width of the window and a <i>slideTime</i>, which represents the frequency during which the window is calculated. +Spark Streaming features windowed computations, which allow you to apply transformations over a sliding window of data. All window functions take a <i>windowDuration</i>, which represents the width of the window and a <i>slideTime</i>, which represents the frequency during which the window is calculated. <table class="table"> -<tr><th style="width:25%">Transformation</th><th>Meaning</th></tr> +<tr><th style="width:30%">Transformation</th><th>Meaning</th></tr> <tr> - <td> <b>window</b>(<i>windowDuration</i>, </i>slideTime</i>) </td> - <td> Return a new stream which is computed based on windowed batches of the source stream. <i>windowDuration</i> is the width of the window and <i>slideTime</i> is the frequency during which the window is calculated. Both times must be multiples of the batch interval. + <td> <b>window</b>(<i>windowDuration</i>, </i>slideDuration</i>) </td> + <td> Return a new DStream which is computed based on windowed batches of the source DStream. <i>windowDuration</i> is the width of the window and <i>slideTime</i> is the frequency during which the window is calculated. Both times must be multiples of the batch interval. </td> </tr> <tr> - <td> <b>countByWindow</b>(<i>windowDuration</i>, </i>slideTime</i>) </td> + <td> <b>countByWindow</b>(<i>windowDuration</i>, </i>slideDuration</i>) </td> <td> Return a sliding count of elements in the stream. <i>windowDuration</i> and <i>slideDuration</i> are exactly as defined in <code>window()</code>. </td> </tr> <tr> - <td> <b>reduceByWindow</b>(<i>func</i>, <i>windowDuration</i>, </i>slideDuration</i>) </td> + <td> <b>reduceByWindow</b>(<i>func</i>, <i>windowDuration</i>, <i>slideDuration</i>) </td> <td> Return a new single-element stream, created by aggregating elements in the stream over a sliding interval using <i>func</i>. The function should be associative so that it can be computed correctly in parallel. <i>windowDuration</i> and <i>slideDuration</i> are exactly as defined in <code>window()</code>. </td> </tr> <tr> - <td> <b>groupByKeyAndWindow</b>(windowDuration, slideDuration, [<i>numTasks</i>]) + <td> <b>groupByKeyAndWindow</b>(<i>windowDuration</i>, <i>slideDuration</i>, [<i>numTasks</i>]) </td> - <td> When called on a stream of (K, V) pairs, returns a stream of (K, Seq[V]) pairs over a sliding window. <br /> -<b>Note:</b> By default, this uses only 8 parallel tasks to do the grouping. You can pass an optional <code>numTasks</code> argument to set a different number of tasks. <i>windowDuration</i> and <i>slideDuration</i> are exactly as defined in <code>window()</code>. -</td> + <td> When called on a DStream of (K, V) pairs, returns a new DStream of (K, Seq[V]) pairs by grouping together values of each key over batches in a sliding window. <br /> +<b>Note:</b> By default, this uses Spark's default number of parallel tasks (2 for local machine, 8 for a cluster) to do the grouping. You can pass an optional <code>numTasks</code> argument to set a different number of tasks.</td> </tr> <tr> - <td> <b>reduceByKeyAndWindow</b>(<i>func</i>, [<i>numTasks</i>]) </td> - <td> When called on a stream of (K, V) pairs, returns a stream of (K, V) pairs where the values for each key are aggregated using the given reduce function over batches within a sliding window. Like in <code>groupByKeyAndWindow</code>, the number of reduce tasks is configurable through an optional second argument. + <td> <b>reduceByKeyAndWindow</b>(<i>func</i>, <i>windowDuration</i>, <i>slideDuration</i>, [<i>numTasks</i>]) </td> + <td> When called on a DStream of (K, V) pairs, returns a new DStream of (K, V) pairs where the values for each key are aggregated using the given reduce function <i>func</i> over batches in a sliding window. Like in <code>groupByKeyAndWindow</code>, the number of reduce tasks is configurable through an optional second argument. <i>windowDuration</i> and <i>slideDuration</i> are exactly as defined in <code>window()</code>. </td> </tr> <tr> - <td> <b>countByKeyAndWindow</b>([<i>numTasks</i>]) </td> - <td> When called on a stream of (K, V) pairs, returns a stream of (K, Int) pairs where the values for each key are the count within a sliding window. Like in <code>countByKeyAndWindow</code>, the number of reduce tasks is configurable through an optional second argument. + <td> <b>reduceByKeyAndWindow</b>(<i>func</i>, <i>invFunc</i>, <i>windowDuration</i>, <i>slideDuration</i>, [<i>numTasks</i>]) </td> + <td> A more efficient version of the above <code>reduceByKeyAndWindow()</code> where the reduce value of each window is calculated + incrementally using the reduce values of the previous window. This is done by reducing the new data that enter the sliding window, and "inverse reducing" the old data that leave the window. An example would be that of "adding" and "subtracting" counts of keys as the window slides. However, it is applicable to only "invertible reduce functions", that is, those reduce functions which have a corresponding "inverse reduce" function (taken as parameter <i>invFunc</i>. Like in <code>groupByKeyAndWindow</code>, the number of reduce tasks is configurable through an optional second argument. <i>windowDuration</i> and <i>slideDuration</i> are exactly as defined in <code>window()</code>. -</td> +</td> +</tr> +<tr> + <td> <b>countByValueAndWindow</b>(<i>windowDuration</i>, <i>slideDuration</i>, [<i>numTasks</i>]) </td> + <td> When called on a DStream of (K, V) pairs, returns a new DStream of (K, Long) pairs where the value of each key is its frequency within a sliding window. Like in <code>groupByKeyAndWindow</code>, the number of reduce tasks is configurable through an optional second argument. + <i>windowDuration</i> and <i>slideDuration</i> are exactly as defined in <code>window()</code>. +</td> </tr> </table> @@ -147,7 +162,7 @@ A complete list of DStream operations is available in the API documentation of [ When an output operator is called, it triggers the computation of a stream. Currently the following output operators are defined: <table class="table"> -<tr><th style="width:25%">Operator</th><th>Meaning</th></tr> +<tr><th style="width:30%">Operator</th><th>Meaning</th></tr> <tr> <td> <b>foreach</b>(<i>func</i>) </td> <td> The fundamental output operator. Applies a function, <i>func</i>, to each RDD generated from the stream. This function should have side effects, such as printing output, saving the RDD to external files, or writing it over the network to an external system. </td> @@ -176,11 +191,6 @@ When an output operator is called, it triggers the computation of a stream. Curr </table> -## DStream Persistence -Similar to RDDs, DStreams also allow developers to persist the stream's data in memory. That is, using `persist()` method on a DStream would automatically persist every RDD of that DStream in memory. This is useful if the data in the DStream will be computed multiple times (e.g., multiple DStream operations on the same data). For window-based operations like `reduceByWindow` and `reduceByKeyAndWindow` and state-based operations like `updateStateByKey`, this is implicitly true. Hence, DStreams generated by window-based operations are automatically persisted in memory, without the developer calling `persist()`. - -Note that, unlike RDDs, the default persistence level of DStreams keeps the data serialized in memory. This is further discussed in the [Performance Tuning](#memory-tuning) section. More information on different persistence levels can be found in [Spark Programming Guide](scala-programming-guide.html#rdd-persistence). - # Starting the Streaming computation All the above DStream operations are completely lazy, that is, the operations will start executing only after the context is started by using {% highlight scala %} @@ -192,8 +202,8 @@ Conversely, the computation can be stopped by using ssc.stop() {% endhighlight %} -# Example - NetworkWordCount.scala -A good example to start off is the spark.streaming.examples.NetworkWordCount. This example counts the words received from a network server every second. Given below is the relevant sections of the source code. You can find the full source code in <Spark repo>/streaming/src/main/scala/spark/streaming/examples/WordCountNetwork.scala. +# Example +A simple example to start off is the [NetworkWordCount](https://github.com/mesos/spark/tree/master/examples/src/main/scala/spark/streaming/examples/NetworkWordCount.scala). This example counts the words received from a network server every second. Given below is the relevant sections of the source code. You can find the full source code in `<Spark repo>/streaming/src/main/scala/spark/streaming/examples/NetworkWordCount.scala` . {% highlight scala %} import spark.streaming.{Seconds, StreamingContext} @@ -202,7 +212,7 @@ import spark.streaming.StreamingContext._ // Create the context and set up a network input stream to receive from a host:port val ssc = new StreamingContext(args(0), "NetworkWordCount", Seconds(1)) -val lines = ssc.networkTextStream(args(1), args(2).toInt) +val lines = ssc.socketTextStream(args(1), args(2).toInt) // Split the lines into words, count them, and print some of the counts on the master val words = lines.flatMap(_.split(" ")) @@ -213,6 +223,8 @@ wordCounts.print() ssc.start() {% endhighlight %} +The `socketTextStream` returns a DStream of lines received from a TCP socket-based source. The `lines` DStream is _transformed_ into a DStream using the `flatMap` operation, where each line is split into words. This `words` DStream is then mapped to a DStream of `(word, 1)` pairs, which is finally reduced to get the word counts. `wordCounts.print()` will print 10 of the counts generated every second. + To run this example on your local machine, you need to first run a Netcat server by using {% highlight bash %} @@ -260,6 +272,33 @@ Time: 1357008430000 ms </td> </table> +You can find more examples in `<Spark repo>/streaming/src/main/scala/spark/streaming/examples/`. They can be run in the similar manner using `./run spark.streaming.examples....` . Executing without any parameter would give the required parameter list. Further explanation to run them can be found in comments in the files. + +# DStream Persistence +Similar to RDDs, DStreams also allow developers to persist the stream's data in memory. That is, using `persist()` method on a DStream would automatically persist every RDD of that DStream in memory. This is useful if the data in the DStream will be computed multiple times (e.g., multiple operations on the same data). For window-based operations like `reduceByWindow` and `reduceByKeyAndWindow` and state-based operations like `updateStateByKey`, this is implicitly true. Hence, DStreams generated by window-based operations are automatically persisted in memory, without the developer calling `persist()`. + +For input streams that receive data from the network (that is, subclasses of NetworkInputDStream like FlumeInputDStream and KafkaInputDStream), the default persistence level is set to replicate the data to two nodes for fault-tolerance. + +Note that, unlike RDDs, the default persistence level of DStreams keeps the data serialized in memory. This is further discussed in the [Performance Tuning](#memory-tuning) section. More information on different persistence levels can be found in [Spark Programming Guide](scala-programming-guide.html#rdd-persistence). + +# RDD Checkpointing within DStreams +A _stateful operation_ is one which operates over multiple batches of data. This includes all window-based operations and the `updateStateByKey` operation. + +Because stateful operations have a dependency on previous batches of data, they continuously accumulate metadata over time. To clear this metadata, streaming supports periodic _checkpointing_ by saving intermediate data to HDFS. Note that checkpointing also incurs the cost of saving to HDFS which may cause the corresponding batch to take longer to process. Hence, the interval of checkpointing needs to be set carefully. At small batch sizes (say 1 second), checkpointing every batch may significantly reduce operation throughput. Conversely, checkpointing too slowly causes the lineage and task sizes to grow which may have detrimental effects. Typically, a checkpoint interval of 5 - 10 times of sliding interval of a DStream is good setting to try. + +To enable checkpointing, the developer has to provide the HDFS path to which RDD will be saved. This is done by using + +{% highlight scala %} +ssc.checkpoint(hdfsPath) // assuming ssc is the StreamingContext +{% endhighlight %} + +The interval of checkpointing of a DStream can be set by using + +{% highlight scala %} +dstream.checkpoint(checkpointInterval) // checkpointInterval must be a multiple of slide duration of dstream +{% endhighlight %} + +For DStreams that must be checkpointed (that is, DStreams created by `updateStateByKey` and `reduceByKeyAndWindow` with inverse function), the checkpoint interval of the DStream is by default set to a multiple of the DStream's sliding interval such that its at least 10 seconds. # Performance Tuning @@ -273,17 +312,21 @@ Getting the best performance of a Spark Streaming application on a cluster requi There are a number of optimizations that can be done in Spark to minimize the processing time of each batch. These have been discussed in detail in [Tuning Guide](tuning.html). This section highlights some of the most important ones. ### Level of Parallelism -Cluster resources maybe underutilized if the number of parallel tasks used in any stage of the computation is not high enough. For example, for distributed reduce operations like `reduceByKey` and `reduceByKeyAndWindow`, the default number of parallel tasks is 8. You can pass the level of parallelism as an argument (see the [`spark.PairDStreamFunctions`](api/streaming/index.html#spark.PairDStreamFunctions) documentation), or set the system property `spark.default.parallelism` to change the default. +Cluster resources maybe under-utilized if the number of parallel tasks used in any stage of the computation is not high enough. For example, for distributed reduce operations like `reduceByKey` and `reduceByKeyAndWindow`, the default number of parallel tasks is 8. You can pass the level of parallelism as an argument (see the [`spark.PairDStreamFunctions`](api/streaming/index.html#spark.PairDStreamFunctions) documentation), or set the system property `spark.default.parallelism` to change the default. ### Data Serialization The overhead of data serialization can be significant, especially when sub-second batch sizes are to be achieved. There are two aspects to it. -* Serialization of RDD data in Spark: Please refer to the detailed discussion on data serialization in the [Tuning Guide](tuning.html). However, note that unlike Spark, by default RDDs are persisted as serialized byte arrays to minimize pauses related to GC. -* Serialization of input data: To ingest external data into Spark, data received as bytes (say, from the network) needs to deserialized from bytes and re-serialized into Spark's serialization format. Hence, the deserialization overhead of input data may be a bottleneck. + +* **Serialization of RDD data in Spark**: Please refer to the detailed discussion on data serialization in the [Tuning Guide](tuning.html). However, note that unlike Spark, by default RDDs are persisted as serialized byte arrays to minimize pauses related to GC. + +* **Serialization of input data**: To ingest external data into Spark, data received as bytes (say, from the network) needs to deserialized from bytes and re-serialized into Spark's serialization format. Hence, the deserialization overhead of input data may be a bottleneck. ### Task Launching Overheads If the number of tasks launched per second is high (say, 50 or more per second), then the overhead of sending out tasks to the slaves maybe significant and will make it hard to achieve sub-second latencies. The overhead can be reduced by the following changes: -* Task Serialization: Using Kryo serialization for serializing tasks can reduced the task sizes, and therefore reduce the time taken to send them to the slaves. -* Execution mode: Running Spark in Standalone mode or coarse-grained Mesos mode leads to better task launch times than the fine-grained Mesos mode. Please refer to the [Running on Mesos guide](running-on-mesos.html) for more details. + +* **Task Serialization**: Using Kryo serialization for serializing tasks can reduced the task sizes, and therefore reduce the time taken to send them to the slaves. + +* **Execution mode**: Running Spark in Standalone mode or coarse-grained Mesos mode leads to better task launch times than the fine-grained Mesos mode. Please refer to the [Running on Mesos guide](running-on-mesos.html) for more details. These changes may reduce batch processing time by 100s of milliseconds, thus allowing sub-second batch size to be viable. ## Setting the Right Batch Size @@ -292,22 +335,183 @@ For a Spark Streaming application running on a cluster to be stable, the process A good approach to figure out the right batch size for your application is to test it with a conservative batch size (say, 5-10 seconds) and a low data rate. To verify whether the system is able to keep up with data rate, you can check the value of the end-to-end delay experienced by each processed batch (in the Spark master logs, find the line having the phrase "Total delay"). If the delay is maintained to be less than the batch size, then system is stable. Otherwise, if the delay is continuously increasing, it means that the system is unable to keep up and it therefore unstable. Once you have an idea of a stable configuration, you can try increasing the data rate and/or reducing the batch size. Note that momentary increase in the delay due to temporary data rate increases maybe fine as long as the delay reduces back to a low value (i.e., less than batch size). ## 24/7 Operation -By default, Spark does not forget any of the metadata (RDDs generated, stages processed, etc.). But for a Spark Streaming application to operate 24/7, it is necessary for Spark to do periodic cleanup of it metadata. This can be enabled by setting the Java system property `spark.cleaner.delay` to the number of minutes you want any metadata to persist. For example, setting `spark.cleaner.delay` to 10 would cause Spark periodically cleanup all metadata and persisted RDDs that are older than 10 minutes. Note, that this property needs to be set before the SparkContext is created. +By default, Spark does not forget any of the metadata (RDDs generated, stages processed, etc.). But for a Spark Streaming application to operate 24/7, it is necessary for Spark to do periodic cleanup of it metadata. This can be enabled by setting the Java system property `spark.cleaner.ttl` to the number of seconds you want any metadata to persist. For example, setting `spark.cleaner.ttl` to 600 would cause Spark periodically cleanup all metadata and persisted RDDs that are older than 10 minutes. Note, that this property needs to be set before the SparkContext is created. This value is closely tied with any window operation that is being used. Any window operation would require the input data to be persisted in memory for at least the duration of the window. Hence it is necessary to set the delay to at least the value of the largest window operation used in the Spark Streaming application. If this delay is set too low, the application will throw an exception saying so. ## Memory Tuning Tuning the memory usage and GC behavior of Spark applications have been discussed in great detail in the [Tuning Guide](tuning.html). It is recommended that you read that. In this section, we highlight a few customizations that are strongly recommended to minimize GC related pauses in Spark Streaming applications and achieving more consistent batch processing times. -* <b>Default persistence level of DStreams</b>: Unlike RDDs, the default persistence level of DStreams serializes the data in memory (that is, [StorageLevel.MEMORY_ONLY_SER](api/core/index.html#spark.storage.StorageLevel$) for DStream compared to [StorageLevel.MEMORY_ONLY](api/core/index.html#spark.storage.StorageLevel$) for RDDs). Even though keeping the data serialized incurs a higher serialization overheads, it significantly reduces GC pauses. +* **Default persistence level of DStreams**: Unlike RDDs, the default persistence level of DStreams serializes the data in memory (that is, [StorageLevel.MEMORY_ONLY_SER](api/core/index.html#spark.storage.StorageLevel$) for DStream compared to [StorageLevel.MEMORY_ONLY](api/core/index.html#spark.storage.StorageLevel$) for RDDs). Even though keeping the data serialized incurs a higher serialization overheads, it significantly reduces GC pauses. + +* **Concurrent garbage collector**: Using the concurrent mark-and-sweep GC further minimizes the variability of GC pauses. Even though concurrent GC is known to reduce the overall processing throughput of the system, its use is still recommended to achieve more consistent batch processing times. + +# Fault-tolerance Properties +In this section, we are going to discuss the behavior of Spark Streaming application in the event of a node failure. To understand this, let us remember the basic fault-tolerance properties of Spark's RDDs. + + 1. An RDD is an immutable, and deterministically re-computable, distributed dataset. Each RDD remembers the lineage of deterministic operations that were used on a fault-tolerant input dataset to create it. + 1. If any partition of an RDD is lost due to a worker node failure, then that partition can be re-computed from the original fault-tolerant dataset using the lineage of operations. + +Since all data transformations in Spark Streaming are based on RDD operations, as long as the input dataset is present, all intermediate data can recomputed. Keeping these properties in mind, we are going to discuss the failure semantics in more detail. + +## Failure of a Worker Node + +There are two failure behaviors based on which input sources are used. + +1. _Using HDFS files as input source_ - Since the data is reliably stored on HDFS, all data can re-computed and therefore no data will be lost due to any failure. +1. _Using any input source that receives data through a network_ - For network-based data sources like Kafka and Flume, the received input data is replicated in memory between nodes of the cluster (default replication factor is 2). So if a worker node fails, then the system can recompute the lost from the the left over copy of the input data. However, if the worker node where a network receiver was running fails, then a tiny bit of data may be lost, that is, the data received by the system but not yet replicated to other node(s). The receiver will be started on a different node and it will continue to receive data. + +Since all data is modeled as RDDs with their lineage of deterministic operations, any recomputation always leads to the same result. As a result, all DStream transformations are guaranteed to have _exactly-once_ semantics. That is, the final transformed result will be same even if there were was a worker node failure. However, output operations (like `foreach`) have _at-least once_ semantics, that is, the transformed data may get written to an external entity more than once in the event of a worker failure. While this is acceptable for saving to HDFS using the `saveAs*Files` operations (as the file will simply get over-written by the same data), additional transactions-like mechanisms may be necessary to achieve exactly-once semantics for output operations. + +## Failure of the Driver Node +A system that is required to operate 24/7 needs to be able tolerate the failure of the driver node as well. Spark Streaming does this by saving the state of the DStream computation periodically to a HDFS file, that can be used to restart the streaming computation in the event of a failure of the driver node. This checkpointing is enabled by setting a HDFS directory for checkpointing using `ssc.checkpoint(<checkpoint directory>)` as described [earlier](#rdd-checkpointing-within-dstreams). To elaborate, the following state is periodically saved to a file. + +1. The DStream operator graph (input streams, output streams, etc.) +1. The configuration of each DStream (checkpoint interval, etc.) +1. The RDD checkpoint files of each DStream + +All this is periodically saved in the file `<checkpoint directory>/graph`. To recover, a new Streaming Context can be created with this directory by using + +{% highlight scala %} +val ssc = new StreamingContext(checkpointDirectory) +{% endhighlight %} + +On calling `ssc.start()` on this new context, the following steps are taken by the system -* <b>Concurrent garbage collector</b>: Using the concurrent mark-and-sweep GC further minimizes the variability of GC pauses. Even though concurrent GC is known to reduce the overall processing throughput of the system, its use is still recommended to achieve more consistent batch processing times. +1. Schedule the transformations and output operations for all the time steps between the time when the driver failed and when it last checkpointed. This is also done for those time steps that were previously scheduled but not processed due to the failure. This will make the system recompute all the intermediate data from the checkpointed RDD files, etc. +1. Restart the network receivers, if any, and continue receiving new data. -# Master Fault-tolerance (Alpha) -TODO +In the current _alpha_ release, there are two different failure behaviors based on which input sources are used. + +1. _Using HDFS files as input source_ - Since the data is reliably stored on HDFS, all data can re-computed and therefore no data will be lost due to any failure. +1. _Using any input source that receives data through a network_ - The received input data is replicated in memory to multiple nodes. Since, all the data in the Spark worker's memory is lost when the Spark driver fails, the past input data will not be accessible and driver recovers. Hence, if stateful and window-based operations are used (like `updateStateByKey`, `window`, `countByValueAndWindow`, etc.), then the intermediate state will not be recovered completely. + +In future releases, we will support full recoverability for all input sources. Note that for non-stateful transformations like `map`, `count`, and `reduceByKey`, with _all_ input streams, the system, upon restarting, will continue to receive and process new data. + +To better understand the behavior of the system under driver failure with a HDFS source, lets consider what will happen with a file input stream Specifically, in the case of the file input stream, it will correctly identify new files that were created while the driver was down and process them in the same way as it would have if the driver had not failed. To explain further in the case of file input stream, we shall use an example. Lets say, files are being generated every second, and a Spark Streaming program reads every new file and output the number of lines in the file. This is what the sequence of outputs would be with and without a driver failure. + +<table class="table"> + <!-- Results table headers --> + <tr> + <th> Time </th> + <th> Number of lines in input file </th> + <th> Output without driver failure </th> + <th> Output with driver failure </th> + </tr> + <tr> + <td>1</td> + <td>10</td> + <td>10</td> + <td>10</td> + </tr> + <tr> + <td>2</td> + <td>20</td> + <td>20</td> + <td>20</td> + </tr> + <tr> + <td>3</td> + <td>30</td> + <td>30</td> + <td>30</td> + </tr> + <tr> + <td>4</td> + <td>40</td> + <td>40</td> + <td>[DRIVER FAILS]<br />no output</td> + </tr> + <tr> + <td>5</td> + <td>50</td> + <td>50</td> + <td>no output</td> + </tr> + <tr> + <td>6</td> + <td>60</td> + <td>60</td> + <td>no output</td> + </tr> + <tr> + <td>7</td> + <td>70</td> + <td>70</td> + <td>[DRIVER RECOVERS]<br />40, 50, 60, 70</td> + </tr> + <tr> + <td>8</td> + <td>80</td> + <td>80</td> + <td>80</td> + </tr> + <tr> + <td>9</td> + <td>90</td> + <td>90</td> + <td>90</td> + </tr> + <tr> + <td>10</td> + <td>100</td> + <td>100</td> + <td>100</td> + </tr> +</table> + +If the driver had crashed in the middle of the processing of time 3, then it will process time 3 and output 30 after recovery. + +# Java API + +Similar to [Spark's Java API](java-programming-guide.html), we also provide a Java API for Spark Streaming which allows all its features to be accessible from a Java program. This is defined in [spark.streaming.api.java] (api/streaming/index.html#spark.streaming.api.java.package) package and includes [JavaStreamingContext](api/streaming/index.html#spark.streaming.api.java.JavaStreamingContext) and [JavaDStream](api/streaming/index.html#spark.streaming.api.java.JavaDStream) classes that provide the same methods as their Scala counterparts, but take Java functions (that is, Function, and Function2) and return Java data and collection types. Some of the key points to note are: + +1. Functions for transformations must be implemented as subclasses of [Function](api/core/index.html#spark.api.java.function.Function) and [Function2](api/core/index.html#spark.api.java.function.Function2) +1. Unlike the Scala API, the Java API handles DStreams for key-value pairs using a separate [JavaPairDStream](api/streaming/index.html#spark.streaming.api.java.JavaPairDStream) class(similar to [JavaRDD and JavaPairRDD](java-programming-guide.html#rdd-classes). DStream functions like `map` and `filter` are implemented separately by JavaDStreams and JavaPairDStream to return DStreams of appropriate types. + +Spark's [Java Programming Guide](java-programming-guide.html) gives more ideas about using the Java API. To extends the ideas presented for the RDDs to DStreams, we present parts of the Java version of the same NetworkWordCount example presented above. The full source code is given at `<spark repo>/examples/src/main/java/spark/streaming/examples/JavaNetworkWordCount.java` + +The streaming context and the socket stream from input source is started by using a `JavaStreamingContext`, that has the same parameters and provides the same input streams as its Scala counterpart. + +{% highlight java %} +JavaStreamingContext ssc = new JavaStreamingContext(mesosUrl, "NetworkWordCount", Seconds(1)); +JavaDStream<String> lines = ssc.socketTextStream(ip, port); +{% endhighlight %} + + +Then the `lines` are split into words by using the `flatMap` function and [FlatMapFunction](api/core/index.html#spark.api.java.function.FlatMapFunction). + +{% highlight java %} +JavaDStream<String> words = lines.flatMap( + new FlatMapFunction<String, String>() { + @Override + public Iterable<String> call(String x) { + return Lists.newArrayList(x.split(" ")); + } + }); +{% endhighlight %} + +The `words` is then mapped to a [JavaPairDStream](api/streaming/index.html#spark.streaming.api.java.JavaPairDStream) of `(word, 1)` pairs using `map` and [PairFunction](api/core/index.html#spark.api.java.function.PairFunction). This is reduced by using `reduceByKey` and [Function2](api/core/index.html#spark.api.java.function.Function2). + +{% highlight java %} +JavaPairDStream<String, Integer> wordCounts = words.map( + new PairFunction<String, String, Integer>() { + @Override + public Tuple2<String, Integer> call(String s) throws Exception { + return new Tuple2<String, Integer>(s, 1); + } + }).reduceByKey( + new Function2<Integer, Integer, Integer>() { + @Override + public Integer call(Integer i1, Integer i2) throws Exception { + return i1 + i2; + } + }); +{% endhighlight %} -* Checkpointing of DStream graph -* Recovery from master faults -* Current state and future directions
\ No newline at end of file +# Where to Go from Here +* API docs - [Scala](api/streaming/index.html#spark.streaming.package) and [Java](api/streaming/index.html#spark.streaming.api.java.package) +* More examples - [Scala](https://github.com/mesos/spark/tree/master/examples/src/main/scala/spark/streaming/examples) and [Java](https://github.com/mesos/spark/tree/master/examples/src/main/java/spark/streaming/examples) +* [Paper describing Spark Streaming](http://www.eecs.berkeley.edu/Pubs/TechRpts/2012/EECS-2012-259.pdf) diff --git a/docs/tuning.md b/docs/tuning.md index 738c530458..32c7ab86e9 100644 --- a/docs/tuning.md +++ b/docs/tuning.md @@ -49,7 +49,7 @@ Finally, to register your classes with Kryo, create a public class that extends {% highlight scala %} import com.esotericsoftware.kryo.Kryo -class MyRegistrator extends KryoRegistrator { +class MyRegistrator extends spark.KryoRegistrator { override def registerClasses(kryo: Kryo) { kryo.register(classOf[MyClass1]) kryo.register(classOf[MyClass2]) @@ -213,10 +213,10 @@ but at a high level, managing how frequently full GC takes place can help in red Clusters will not be fully utilized unless you set the level of parallelism for each operation high enough. Spark automatically sets the number of "map" tasks to run on each file according to its size -(though you can control it through optional parameters to `SparkContext.textFile`, etc), but for -distributed "reduce" operations, such as `groupByKey` and `reduceByKey`, it uses a default value of 8. -You can pass the level of parallelism as a second argument (see the -[`spark.PairRDDFunctions`](api/core/index.html#spark.PairRDDFunctions) documentation), +(though you can control it through optional parameters to `SparkContext.textFile`, etc), and for +distributed "reduce" operations, such as `groupByKey` and `reduceByKey`, it uses the largest +parent RDD's number of partitions. You can pass the level of parallelism as a second argument +(see the [`spark.PairRDDFunctions`](api/core/index.html#spark.PairRDDFunctions) documentation), or set the system property `spark.default.parallelism` to change the default. In general, we recommend 2-3 tasks per CPU core in your cluster. diff --git a/ec2/README b/ec2/README index 58dc087a81..0add81312c 100644 --- a/ec2/README +++ b/ec2/README @@ -1,4 +1,4 @@ This folder contains a script, spark-ec2, for launching Spark clusters on Amazon EC2. Usage instructions are available online at: -https://github.com/mesos/spark/wiki/Running-Spark-on-Amazon-EC2 +http://spark-project.org/docs/latest/ec2-scripts.html diff --git a/ec2/spark_ec2.py b/ec2/spark_ec2.py index 66b1faf2cd..7affe6fffc 100755 --- a/ec2/spark_ec2.py +++ b/ec2/spark_ec2.py @@ -35,7 +35,7 @@ from boto.ec2.blockdevicemapping import BlockDeviceMapping, EBSBlockDeviceType from boto import ec2 # A static URL from which to figure out the latest Mesos EC2 AMI -LATEST_AMI_URL = "https://s3.amazonaws.com/mesos-images/ids/latest-spark-0.6" +LATEST_AMI_URL = "https://s3.amazonaws.com/mesos-images/ids/latest-spark-0.7" # Configure and parse our command-line arguments @@ -83,16 +83,16 @@ def parse_args(): help="If specified, launch slaves as spot instances with the given " + "maximum price (in dollars)") parser.add_option("--cluster-type", type="choice", metavar="TYPE", - choices=["mesos", "standalone"], default="mesos", + choices=["mesos", "standalone"], default="standalone", help="'mesos' for a Mesos cluster, 'standalone' for a standalone " + - "Spark cluster (default: mesos)") + "Spark cluster (default: standalone)") parser.add_option("--ganglia", action="store_true", default=True, help="Setup Ganglia monitoring on cluster (default: on). NOTE: " + "the Ganglia page will be publicly accessible") parser.add_option("--no-ganglia", action="store_false", dest="ganglia", help="Disable Ganglia monitoring for the cluster") - parser.add_option("--new-scripts", action="store_true", default=False, - help="Use new spark-ec2 scripts, for Spark >= 0.7 AMIs") + parser.add_option("--old-scripts", action="store_true", default=False, + help="Use old mesos-ec2 scripts, for Spark <= 0.6 AMIs") parser.add_option("-u", "--user", default="root", help="The SSH user you want to connect as (default: root)") parser.add_option("--delete-groups", action="store_true", default=False, @@ -103,7 +103,7 @@ def parse_args(): parser.print_help() sys.exit(1) (action, cluster_name) = args - if opts.identity_file == None and action in ['launch', 'login']: + if opts.identity_file == None and action in ['launch', 'login', 'start']: print >> stderr, ("ERROR: The -i or --identity-file argument is " + "required for " + action) sys.exit(1) @@ -383,7 +383,7 @@ def setup_cluster(conn, master_nodes, slave_nodes, zoo_nodes, opts, deploy_ssh_k if opts.ganglia: modules.append('ganglia') - if opts.new_scripts: + if not opts.old_scripts: # NOTE: We should clone the repository before running deploy_files to # prevent ec2-variables.sh from being overwritten ssh(master, opts, "rm -rf spark-ec2 && git clone https://github.com/mesos/spark-ec2.git") @@ -393,7 +393,7 @@ def setup_cluster(conn, master_nodes, slave_nodes, zoo_nodes, opts, deploy_ssh_k zoo_nodes, modules) print "Running setup on master..." - if not opts.new_scripts: + if opts.old_scripts: if opts.cluster_type == "mesos": setup_mesos_cluster(master, opts) elif opts.cluster_type == "standalone": @@ -540,11 +540,24 @@ def scp(host, opts, local_file, dest_file): (opts.identity_file, local_file, opts.user, host, dest_file), shell=True) -# Run a command on a host through ssh, throwing an exception if ssh fails +# Run a command on a host through ssh, retrying up to two times +# and then throwing an exception if ssh continues to fail. def ssh(host, opts, command): - subprocess.check_call( - "ssh -t -o StrictHostKeyChecking=no -i %s %s@%s '%s'" % - (opts.identity_file, opts.user, host, command), shell=True) + tries = 0 + while True: + try: + return subprocess.check_call( + "ssh -t -o StrictHostKeyChecking=no -i %s %s@%s '%s'" % + (opts.identity_file, opts.user, host, command), shell=True) + except subprocess.CalledProcessError as e: + if (tries > 2): + raise e + print "Error connecting to host {0}, sleeping 30".format(e) + time.sleep(30) + tries = tries + 1 + + + # Gets a list of zones to launch instances in diff --git a/examples/pom.xml b/examples/pom.xml index f43af670c6..3e5271ec2f 100644 --- a/examples/pom.xml +++ b/examples/pom.xml @@ -3,8 +3,8 @@ <modelVersion>4.0.0</modelVersion> <parent> <groupId>org.spark-project</groupId> - <artifactId>parent</artifactId> - <version>0.7.0-SNAPSHOT</version> + <artifactId>spark-parent</artifactId> + <version>0.8.0-SNAPSHOT</version> <relativePath>../pom.xml</relativePath> </parent> @@ -20,11 +20,10 @@ <artifactId>jetty-server</artifactId> </dependency> <dependency> - <groupId>org.twitter4j</groupId> - <artifactId>twitter4j-stream</artifactId> - <version>3.0.3</version> + <groupId>com.twitter</groupId> + <artifactId>algebird-core_2.9.2</artifactId> + <version>0.1.11</version> </dependency> - <dependency> <groupId>org.scalatest</groupId> <artifactId>scalatest_${scala.version}</artifactId> @@ -35,6 +34,41 @@ <artifactId>scalacheck_${scala.version}</artifactId> <scope>test</scope> </dependency> + <dependency> + <groupId>org.apache.cassandra</groupId> + <artifactId>cassandra-all</artifactId> + <version>1.2.5</version> + <exclusions> + <exclusion> + <groupId>com.google.guava</groupId> + <artifactId>guava</artifactId> + </exclusion> + <exclusion> + <groupId>com.googlecode.concurrentlinkedhashmap</groupId> + <artifactId>concurrentlinkedhashmap-lru</artifactId> + </exclusion> + <exclusion> + <groupId>com.ning</groupId> + <artifactId>compress-lzf</artifactId> + </exclusion> + <exclusion> + <groupId>io.netty</groupId> + <artifactId>netty</artifactId> + </exclusion> + <exclusion> + <groupId>jline</groupId> + <artifactId>jline</artifactId> + </exclusion> + <exclusion> + <groupId>log4j</groupId> + <artifactId>log4j</artifactId> + </exclusion> + <exclusion> + <groupId>org.apache.cassandra.deps</groupId> + <artifactId>avro</artifactId> + </exclusion> + </exclusions> + </dependency> </dependencies> <build> <outputDirectory>target/scala-${scala.version}/classes</outputDirectory> @@ -68,6 +102,11 @@ <artifactId>hadoop-core</artifactId> <scope>provided</scope> </dependency> + <dependency> + <groupId>org.apache.hbase</groupId> + <artifactId>hbase</artifactId> + <version>0.94.6</version> + </dependency> </dependencies> <build> <plugins> @@ -106,6 +145,11 @@ <artifactId>hadoop-client</artifactId> <scope>provided</scope> </dependency> + <dependency> + <groupId>org.apache.hbase</groupId> + <artifactId>hbase</artifactId> + <version>0.94.6</version> + </dependency> </dependencies> <build> <plugins> @@ -119,5 +163,48 @@ </plugins> </build> </profile> + <profile> + <id>hadoop2-yarn</id> + <dependencies> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-core</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + </dependency> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-streaming</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-api</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> + <scope>provided</scope> + </dependency> + </dependencies> + <build> + <plugins> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-jar-plugin</artifactId> + <configuration> + <classifier>hadoop2-yarn</classifier> + </configuration> + </plugin> + </plugins> + </build> + </profile> </profiles> </project> diff --git a/examples/src/main/java/spark/examples/JavaHdfsLR.java b/examples/src/main/java/spark/examples/JavaHdfsLR.java index 29839d5668..8b0a9b6808 100644 --- a/examples/src/main/java/spark/examples/JavaHdfsLR.java +++ b/examples/src/main/java/spark/examples/JavaHdfsLR.java @@ -10,6 +10,9 @@ import java.util.Arrays; import java.util.StringTokenizer; import java.util.Random; +/** + * Logistic regression based classification. + */ public class JavaHdfsLR { static int D = 10; // Number of dimensions @@ -85,7 +88,8 @@ public class JavaHdfsLR { System.exit(1); } - JavaSparkContext sc = new JavaSparkContext(args[0], "JavaHdfsLR"); + JavaSparkContext sc = new JavaSparkContext(args[0], "JavaHdfsLR", + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); JavaRDD<String> lines = sc.textFile(args[1]); JavaRDD<DataPoint> points = lines.map(new ParsePoint()).cache(); int ITERATIONS = Integer.parseInt(args[2]); diff --git a/examples/src/main/java/spark/examples/JavaKMeans.java b/examples/src/main/java/spark/examples/JavaKMeans.java new file mode 100644 index 0000000000..626034eb0d --- /dev/null +++ b/examples/src/main/java/spark/examples/JavaKMeans.java @@ -0,0 +1,114 @@ +package spark.examples; + +import scala.Tuple2; +import spark.api.java.JavaPairRDD; +import spark.api.java.JavaRDD; +import spark.api.java.JavaSparkContext; +import spark.api.java.function.Function; +import spark.api.java.function.PairFunction; +import spark.util.Vector; + +import java.util.List; +import java.util.Map; + +/** + * K-means clustering using Java API. + */ +public class JavaKMeans { + + /** Parses numbers split by whitespace to a vector */ + static Vector parseVector(String line) { + String[] splits = line.split(" "); + double[] data = new double[splits.length]; + int i = 0; + for (String s : splits) + data[i] = Double.parseDouble(splits[i++]); + return new Vector(data); + } + + /** Computes the vector to which the input vector is closest using squared distance */ + static int closestPoint(Vector p, List<Vector> centers) { + int bestIndex = 0; + double closest = Double.POSITIVE_INFINITY; + for (int i = 0; i < centers.size(); i++) { + double tempDist = p.squaredDist(centers.get(i)); + if (tempDist < closest) { + closest = tempDist; + bestIndex = i; + } + } + return bestIndex; + } + + /** Computes the mean across all vectors in the input set of vectors */ + static Vector average(List<Vector> ps) { + int numVectors = ps.size(); + Vector out = new Vector(ps.get(0).elements()); + // start from i = 1 since we already copied index 0 above + for (int i = 1; i < numVectors; i++) { + out.addInPlace(ps.get(i)); + } + return out.divide(numVectors); + } + + public static void main(String[] args) throws Exception { + if (args.length < 4) { + System.err.println("Usage: JavaKMeans <master> <file> <k> <convergeDist>"); + System.exit(1); + } + JavaSparkContext sc = new JavaSparkContext(args[0], "JavaKMeans", + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); + String path = args[1]; + int K = Integer.parseInt(args[2]); + double convergeDist = Double.parseDouble(args[3]); + + JavaRDD<Vector> data = sc.textFile(path).map( + new Function<String, Vector>() { + @Override + public Vector call(String line) throws Exception { + return parseVector(line); + } + } + ).cache(); + + final List<Vector> centroids = data.takeSample(false, K, 42); + + double tempDist; + do { + // allocate each vector to closest centroid + JavaPairRDD<Integer, Vector> closest = data.map( + new PairFunction<Vector, Integer, Vector>() { + @Override + public Tuple2<Integer, Vector> call(Vector vector) throws Exception { + return new Tuple2<Integer, Vector>( + closestPoint(vector, centroids), vector); + } + } + ); + + // group by cluster id and average the vectors within each cluster to compute centroids + JavaPairRDD<Integer, List<Vector>> pointsGroup = closest.groupByKey(); + Map<Integer, Vector> newCentroids = pointsGroup.mapValues( + new Function<List<Vector>, Vector>() { + public Vector call(List<Vector> ps) throws Exception { + return average(ps); + } + }).collectAsMap(); + tempDist = 0.0; + for (int i = 0; i < K; i++) { + tempDist += centroids.get(i).squaredDist(newCentroids.get(i)); + } + for (Map.Entry<Integer, Vector> t: newCentroids.entrySet()) { + centroids.set(t.getKey(), t.getValue()); + } + System.out.println("Finished iteration (delta = " + tempDist + ")"); + } while (tempDist > convergeDist); + + System.out.println("Final centers:"); + for (Vector c : centroids) + System.out.println(c); + + System.exit(0); + + } +} diff --git a/examples/src/main/java/spark/examples/JavaLogQuery.java b/examples/src/main/java/spark/examples/JavaLogQuery.java new file mode 100644 index 0000000000..6b22e7120c --- /dev/null +++ b/examples/src/main/java/spark/examples/JavaLogQuery.java @@ -0,0 +1,114 @@ +package spark.examples; + +import com.google.common.collect.Lists; +import scala.Tuple2; +import scala.Tuple3; +import spark.api.java.JavaPairRDD; +import spark.api.java.JavaRDD; +import spark.api.java.JavaSparkContext; +import spark.api.java.function.Function2; +import spark.api.java.function.PairFunction; + +import java.io.Serializable; +import java.util.Collections; +import java.util.List; +import java.util.regex.Matcher; +import java.util.regex.Pattern; + +/** + * Executes a roll up-style query against Apache logs. + */ +public class JavaLogQuery { + + public static List<String> exampleApacheLogs = Lists.newArrayList( + "10.10.10.10 - \"FRED\" [18/Jan/2013:17:56:07 +1100] \"GET http://images.com/2013/Generic.jpg " + + "HTTP/1.1\" 304 315 \"http://referall.com/\" \"Mozilla/4.0 (compatible; MSIE 7.0; " + + "Windows NT 5.1; GTB7.4; .NET CLR 2.0.50727; .NET CLR 3.0.04506.30; .NET CLR 3.0.04506.648; " + + ".NET CLR 3.5.21022; .NET CLR 3.0.4506.2152; .NET CLR 1.0.3705; .NET CLR 1.1.4322; .NET CLR " + + "3.5.30729; Release=ARP)\" \"UD-1\" - \"image/jpeg\" \"whatever\" 0.350 \"-\" - \"\" 265 923 934 \"\" " + + "62.24.11.25 images.com 1358492167 - Whatup", + "10.10.10.10 - \"FRED\" [18/Jan/2013:18:02:37 +1100] \"GET http://images.com/2013/Generic.jpg " + + "HTTP/1.1\" 304 306 \"http:/referall.com\" \"Mozilla/4.0 (compatible; MSIE 7.0; Windows NT 5.1; " + + "GTB7.4; .NET CLR 2.0.50727; .NET CLR 3.0.04506.30; .NET CLR 3.0.04506.648; .NET CLR " + + "3.5.21022; .NET CLR 3.0.4506.2152; .NET CLR 1.0.3705; .NET CLR 1.1.4322; .NET CLR " + + "3.5.30729; Release=ARP)\" \"UD-1\" - \"image/jpeg\" \"whatever\" 0.352 \"-\" - \"\" 256 977 988 \"\" " + + "0 73.23.2.15 images.com 1358492557 - Whatup"); + + public static Pattern apacheLogRegex = Pattern.compile( + "^([\\d.]+) (\\S+) (\\S+) \\[([\\w\\d:/]+\\s[+\\-]\\d{4})\\] \"(.+?)\" (\\d{3}) ([\\d\\-]+) \"([^\"]+)\" \"([^\"]+)\".*"); + + /** Tracks the total query count and number of aggregate bytes for a particular group. */ + public static class Stats implements Serializable { + + private int count; + private int numBytes; + + public Stats(int count, int numBytes) { + this.count = count; + this.numBytes = numBytes; + } + public Stats merge(Stats other) { + return new Stats(count + other.count, numBytes + other.numBytes); + } + + public String toString() { + return String.format("bytes=%s\tn=%s", numBytes, count); + } + } + + public static Tuple3<String, String, String> extractKey(String line) { + Matcher m = apacheLogRegex.matcher(line); + List<String> key = Collections.emptyList(); + if (m.find()) { + String ip = m.group(1); + String user = m.group(3); + String query = m.group(5); + if (!user.equalsIgnoreCase("-")) { + return new Tuple3<String, String, String>(ip, user, query); + } + } + return new Tuple3<String, String, String>(null, null, null); + } + + public static Stats extractStats(String line) { + Matcher m = apacheLogRegex.matcher(line); + if (m.find()) { + int bytes = Integer.parseInt(m.group(7)); + return new Stats(1, bytes); + } + else + return new Stats(1, 0); + } + + public static void main(String[] args) throws Exception { + if (args.length == 0) { + System.err.println("Usage: JavaLogQuery <master> [logFile]"); + System.exit(1); + } + + JavaSparkContext jsc = new JavaSparkContext(args[0], "JavaLogQuery", + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); + + JavaRDD<String> dataSet = (args.length == 2) ? jsc.textFile(args[1]) : jsc.parallelize(exampleApacheLogs); + + JavaPairRDD<Tuple3<String, String, String>, Stats> extracted = dataSet.map(new PairFunction<String, Tuple3<String, String, String>, Stats>() { + @Override + public Tuple2<Tuple3<String, String, String>, Stats> call(String s) throws Exception { + return new Tuple2<Tuple3<String, String, String>, Stats>(extractKey(s), extractStats(s)); + } + }); + + JavaPairRDD<Tuple3<String, String, String>, Stats> counts = extracted.reduceByKey(new Function2<Stats, Stats, Stats>() { + @Override + public Stats call(Stats stats, Stats stats2) throws Exception { + return stats.merge(stats2); + } + }); + + List<Tuple2<Tuple3<String, String, String>, Stats>> output = counts.collect(); + for (Tuple2 t : output) { + System.out.println(t._1 + "\t" + t._2); + } + System.exit(0); + } +} diff --git a/examples/src/main/java/spark/examples/JavaSparkPi.java b/examples/src/main/java/spark/examples/JavaSparkPi.java new file mode 100644 index 0000000000..a15a967de8 --- /dev/null +++ b/examples/src/main/java/spark/examples/JavaSparkPi.java @@ -0,0 +1,48 @@ +package spark.examples; + +import spark.api.java.JavaRDD; +import spark.api.java.JavaSparkContext; +import spark.api.java.function.Function; +import spark.api.java.function.Function2; + +import java.util.ArrayList; +import java.util.List; + +/** Computes an approximation to pi */ +public class JavaSparkPi { + + + public static void main(String[] args) throws Exception { + if (args.length == 0) { + System.err.println("Usage: JavaLogQuery <master> [slices]"); + System.exit(1); + } + + JavaSparkContext jsc = new JavaSparkContext(args[0], "JavaLogQuery", + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); + + int slices = (args.length == 2) ? Integer.parseInt(args[1]) : 2; + int n = 100000 * slices; + List<Integer> l = new ArrayList<Integer>(n); + for (int i = 0; i < n; i++) + l.add(i); + + JavaRDD<Integer> dataSet = jsc.parallelize(l, slices); + + int count = dataSet.map(new Function<Integer, Integer>() { + @Override + public Integer call(Integer integer) throws Exception { + double x = Math.random() * 2 - 1; + double y = Math.random() * 2 - 1; + return (x * x + y * y < 1) ? 1 : 0; + } + }).reduce(new Function2<Integer, Integer, Integer>() { + @Override + public Integer call(Integer integer, Integer integer2) throws Exception { + return integer + integer2; + } + }); + + System.out.println("Pi is roughly " + 4.0 * count / n); + } +} diff --git a/examples/src/main/java/spark/examples/JavaTC.java b/examples/src/main/java/spark/examples/JavaTC.java index e3bd881b8f..b319bdab44 100644 --- a/examples/src/main/java/spark/examples/JavaTC.java +++ b/examples/src/main/java/spark/examples/JavaTC.java @@ -28,7 +28,7 @@ public class JavaTC { Tuple2<Integer, Integer> e = new Tuple2<Integer, Integer>(from, to); if (from != to) edges.add(e); } - return new ArrayList(edges); + return new ArrayList<Tuple2<Integer, Integer>>(edges); } static class ProjectFn extends PairFunction<Tuple2<Integer, Tuple2<Integer, Integer>>, @@ -46,7 +46,8 @@ public class JavaTC { System.exit(1); } - JavaSparkContext sc = new JavaSparkContext(args[0], "JavaTC"); + JavaSparkContext sc = new JavaSparkContext(args[0], "JavaTC", + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); Integer slices = (args.length > 1) ? Integer.parseInt(args[1]): 2; JavaPairRDD<Integer, Integer> tc = sc.parallelizePairs(generateGraph(), slices).cache(); diff --git a/examples/src/main/java/spark/examples/JavaWordCount.java b/examples/src/main/java/spark/examples/JavaWordCount.java index a44cf8a120..9d4c7a252d 100644 --- a/examples/src/main/java/spark/examples/JavaWordCount.java +++ b/examples/src/main/java/spark/examples/JavaWordCount.java @@ -18,7 +18,8 @@ public class JavaWordCount { System.exit(1); } - JavaSparkContext ctx = new JavaSparkContext(args[0], "JavaWordCount"); + JavaSparkContext ctx = new JavaSparkContext(args[0], "JavaWordCount", + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); JavaRDD<String> lines = ctx.textFile(args[1], 1); JavaRDD<String> words = lines.flatMap(new FlatMapFunction<String, String>() { @@ -29,7 +30,7 @@ public class JavaWordCount { JavaPairRDD<String, Integer> ones = words.map(new PairFunction<String, String, Integer>() { public Tuple2<String, Integer> call(String s) { - return new Tuple2(s, 1); + return new Tuple2<String, Integer>(s, 1); } }); diff --git a/examples/src/main/scala/spark/streaming/examples/JavaFlumeEventCount.java b/examples/src/main/java/spark/streaming/examples/JavaFlumeEventCount.java index cddce16e39..e24c6ddaa7 100644 --- a/examples/src/main/scala/spark/streaming/examples/JavaFlumeEventCount.java +++ b/examples/src/main/java/spark/streaming/examples/JavaFlumeEventCount.java @@ -32,7 +32,8 @@ public class JavaFlumeEventCount { Duration batchInterval = new Duration(2000); - JavaStreamingContext sc = new JavaStreamingContext(master, "FlumeEventCount", batchInterval); + JavaStreamingContext sc = new JavaStreamingContext(master, "FlumeEventCount", batchInterval, + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); JavaDStream<SparkFlumeEvent> flumeStream = sc.flumeStream("localhost", port); diff --git a/examples/src/main/scala/spark/streaming/examples/JavaNetworkWordCount.java b/examples/src/main/java/spark/streaming/examples/JavaNetworkWordCount.java index 4299febfd6..3e57580fd4 100644 --- a/examples/src/main/scala/spark/streaming/examples/JavaNetworkWordCount.java +++ b/examples/src/main/java/spark/streaming/examples/JavaNetworkWordCount.java @@ -23,19 +23,19 @@ import spark.streaming.api.java.JavaStreamingContext; */ public class JavaNetworkWordCount { public static void main(String[] args) { - if (args.length < 2) { + if (args.length < 3) { System.err.println("Usage: NetworkWordCount <master> <hostname> <port>\n" + "In local mode, <master> should be 'local[n]' with n > 1"); System.exit(1); } // Create the context with a 1 second batch size - JavaStreamingContext ssc = new JavaStreamingContext( - args[0], "NetworkWordCount", new Duration(1000)); + JavaStreamingContext ssc = new JavaStreamingContext(args[0], "NetworkWordCount", + new Duration(1000), System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); // Create a NetworkInputDStream on target ip:port and count the // words in input stream of \n delimited test (eg. generated by 'nc') - JavaDStream<String> lines = ssc.networkTextStream(args[1], Integer.parseInt(args[2])); + JavaDStream<String> lines = ssc.socketTextStream(args[1], Integer.parseInt(args[2])); JavaDStream<String> words = lines.flatMap(new FlatMapFunction<String, String>() { @Override public Iterable<String> call(String x) { diff --git a/examples/src/main/scala/spark/streaming/examples/JavaQueueStream.java b/examples/src/main/java/spark/streaming/examples/JavaQueueStream.java index 43c3cd4dfa..15b82c8da1 100644 --- a/examples/src/main/scala/spark/streaming/examples/JavaQueueStream.java +++ b/examples/src/main/java/spark/streaming/examples/JavaQueueStream.java @@ -22,7 +22,8 @@ public class JavaQueueStream { } // Create the context - JavaStreamingContext ssc = new JavaStreamingContext(args[0], "QueueStream", new Duration(1000)); + JavaStreamingContext ssc = new JavaStreamingContext(args[0], "QueueStream", new Duration(1000), + System.getenv("SPARK_HOME"), System.getenv("SPARK_EXAMPLES_JAR")); // Create the queue through which RDDs can be pushed to // a QueueInputDStream diff --git a/examples/src/main/scala/spark/examples/BroadcastTest.scala b/examples/src/main/scala/spark/examples/BroadcastTest.scala index 230097c7db..ba59be1687 100644 --- a/examples/src/main/scala/spark/examples/BroadcastTest.scala +++ b/examples/src/main/scala/spark/examples/BroadcastTest.scala @@ -9,19 +9,21 @@ object BroadcastTest { System.exit(1) } - val spark = new SparkContext(args(0), "Broadcast Test") + val sc = new SparkContext(args(0), "Broadcast Test", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val slices = if (args.length > 1) args(1).toInt else 2 val num = if (args.length > 2) args(2).toInt else 1000000 var arr1 = new Array[Int](num) - for (i <- 0 until arr1.length) + for (i <- 0 until arr1.length) { arr1(i) = i + } for (i <- 0 until 2) { println("Iteration " + i) println("===========") - val barr1 = spark.broadcast(arr1) - spark.parallelize(1 to 10, slices).foreach { + val barr1 = sc.broadcast(arr1) + sc.parallelize(1 to 10, slices).foreach { i => println(barr1.value.size) } } diff --git a/examples/src/main/scala/spark/examples/CassandraTest.scala b/examples/src/main/scala/spark/examples/CassandraTest.scala new file mode 100644 index 0000000000..0fe1833e83 --- /dev/null +++ b/examples/src/main/scala/spark/examples/CassandraTest.scala @@ -0,0 +1,196 @@ +package spark.examples + +import org.apache.hadoop.mapreduce.Job +import org.apache.cassandra.hadoop.ColumnFamilyOutputFormat +import org.apache.cassandra.hadoop.ConfigHelper +import org.apache.cassandra.hadoop.ColumnFamilyInputFormat +import org.apache.cassandra.thrift._ +import spark.SparkContext +import spark.SparkContext._ +import java.nio.ByteBuffer +import java.util.SortedMap +import org.apache.cassandra.db.IColumn +import org.apache.cassandra.utils.ByteBufferUtil +import scala.collection.JavaConversions._ + + +/* + * This example demonstrates using Spark with Cassandra with the New Hadoop API and Cassandra + * support for Hadoop. + * + * To run this example, run this file with the following command params - + * <spark_master> <cassandra_node> <cassandra_port> + * + * So if you want to run this on localhost this will be, + * local[3] localhost 9160 + * + * The example makes some assumptions: + * 1. You have already created a keyspace called casDemo and it has a column family named Words + * 2. There are column family has a column named "para" which has test content. + * + * You can create the content by running the following script at the bottom of this file with + * cassandra-cli. + * + */ +object CassandraTest { + + def main(args: Array[String]) { + + // Get a SparkContext + val sc = new SparkContext(args(0), "casDemo") + + // Build the job configuration with ConfigHelper provided by Cassandra + val job = new Job() + job.setInputFormatClass(classOf[ColumnFamilyInputFormat]) + + val host: String = args(1) + val port: String = args(2) + + ConfigHelper.setInputInitialAddress(job.getConfiguration(), host) + ConfigHelper.setInputRpcPort(job.getConfiguration(), port) + ConfigHelper.setOutputInitialAddress(job.getConfiguration(), host) + ConfigHelper.setOutputRpcPort(job.getConfiguration(), port) + ConfigHelper.setInputColumnFamily(job.getConfiguration(), "casDemo", "Words") + ConfigHelper.setOutputColumnFamily(job.getConfiguration(), "casDemo", "WordCount") + + val predicate = new SlicePredicate() + val sliceRange = new SliceRange() + sliceRange.setStart(Array.empty[Byte]) + sliceRange.setFinish(Array.empty[Byte]) + predicate.setSlice_range(sliceRange) + ConfigHelper.setInputSlicePredicate(job.getConfiguration(), predicate) + + ConfigHelper.setInputPartitioner(job.getConfiguration(), "Murmur3Partitioner") + ConfigHelper.setOutputPartitioner(job.getConfiguration(), "Murmur3Partitioner") + + // Make a new Hadoop RDD + val casRdd = sc.newAPIHadoopRDD( + job.getConfiguration(), + classOf[ColumnFamilyInputFormat], + classOf[ByteBuffer], + classOf[SortedMap[ByteBuffer, IColumn]]) + + // Let us first get all the paragraphs from the retrieved rows + val paraRdd = casRdd.map { + case (key, value) => { + ByteBufferUtil.string(value.get(ByteBufferUtil.bytes("para")).value()) + } + } + + // Lets get the word count in paras + val counts = paraRdd.flatMap(p => p.split(" ")).map(word => (word, 1)).reduceByKey(_ + _) + + counts.collect().foreach { + case (word, count) => println(word + ":" + count) + } + + counts.map { + case (word, count) => { + val colWord = new org.apache.cassandra.thrift.Column() + colWord.setName(ByteBufferUtil.bytes("word")) + colWord.setValue(ByteBufferUtil.bytes(word)) + colWord.setTimestamp(System.currentTimeMillis) + + val colCount = new org.apache.cassandra.thrift.Column() + colCount.setName(ByteBufferUtil.bytes("wcount")) + colCount.setValue(ByteBufferUtil.bytes(count.toLong)) + colCount.setTimestamp(System.currentTimeMillis) + + val outputkey = ByteBufferUtil.bytes(word + "-COUNT-" + System.currentTimeMillis) + + val mutations: java.util.List[Mutation] = new Mutation() :: new Mutation() :: Nil + mutations.get(0).setColumn_or_supercolumn(new ColumnOrSuperColumn()) + mutations.get(0).column_or_supercolumn.setColumn(colWord) + mutations.get(1).setColumn_or_supercolumn(new ColumnOrSuperColumn()) + mutations.get(1).column_or_supercolumn.setColumn(colCount) + (outputkey, mutations) + } + }.saveAsNewAPIHadoopFile("casDemo", classOf[ByteBuffer], classOf[List[Mutation]], + classOf[ColumnFamilyOutputFormat], job.getConfiguration) + } +} + +/* +create keyspace casDemo; +use casDemo; + +create column family WordCount with comparator = UTF8Type; +update column family WordCount with column_metadata = + [{column_name: word, validation_class: UTF8Type}, + {column_name: wcount, validation_class: LongType}]; + +create column family Words with comparator = UTF8Type; +update column family Words with column_metadata = + [{column_name: book, validation_class: UTF8Type}, + {column_name: para, validation_class: UTF8Type}]; + +assume Words keys as utf8; + +set Words['3musk001']['book'] = 'The Three Musketeers'; +set Words['3musk001']['para'] = 'On the first Monday of the month of April, 1625, the market + town of Meung, in which the author of ROMANCE OF THE ROSE was born, appeared to + be in as perfect a state of revolution as if the Huguenots had just made + a second La Rochelle of it. Many citizens, seeing the women flying + toward the High Street, leaving their children crying at the open doors, + hastened to don the cuirass, and supporting their somewhat uncertain + courage with a musket or a partisan, directed their steps toward the + hostelry of the Jolly Miller, before which was gathered, increasing + every minute, a compact group, vociferous and full of curiosity.'; + +set Words['3musk002']['book'] = 'The Three Musketeers'; +set Words['3musk002']['para'] = 'In those times panics were common, and few days passed without + some city or other registering in its archives an event of this kind. There were + nobles, who made war against each other; there was the king, who made + war against the cardinal; there was Spain, which made war against the + king. Then, in addition to these concealed or public, secret or open + wars, there were robbers, mendicants, Huguenots, wolves, and scoundrels, + who made war upon everybody. The citizens always took up arms readily + against thieves, wolves or scoundrels, often against nobles or + Huguenots, sometimes against the king, but never against cardinal or + Spain. It resulted, then, from this habit that on the said first Monday + of April, 1625, the citizens, on hearing the clamor, and seeing neither + the red-and-yellow standard nor the livery of the Duc de Richelieu, + rushed toward the hostel of the Jolly Miller. When arrived there, the + cause of the hubbub was apparent to all'; + +set Words['3musk003']['book'] = 'The Three Musketeers'; +set Words['3musk003']['para'] = 'You ought, I say, then, to husband the means you have, however + large the sum may be; but you ought also to endeavor to perfect yourself in + the exercises becoming a gentleman. I will write a letter today to the + Director of the Royal Academy, and tomorrow he will admit you without + any expense to yourself. Do not refuse this little service. Our + best-born and richest gentlemen sometimes solicit it without being able + to obtain it. You will learn horsemanship, swordsmanship in all its + branches, and dancing. You will make some desirable acquaintances; and + from time to time you can call upon me, just to tell me how you are + getting on, and to say whether I can be of further service to you.'; + + +set Words['thelostworld001']['book'] = 'The Lost World'; +set Words['thelostworld001']['para'] = 'She sat with that proud, delicate profile of hers outlined + against the red curtain. How beautiful she was! And yet how aloof! We had been + friends, quite good friends; but never could I get beyond the same + comradeship which I might have established with one of my + fellow-reporters upon the Gazette,--perfectly frank, perfectly kindly, + and perfectly unsexual. My instincts are all against a woman being too + frank and at her ease with me. It is no compliment to a man. Where + the real sex feeling begins, timidity and distrust are its companions, + heritage from old wicked days when love and violence went often hand in + hand. The bent head, the averted eye, the faltering voice, the wincing + figure--these, and not the unshrinking gaze and frank reply, are the + true signals of passion. Even in my short life I had learned as much + as that--or had inherited it in that race memory which we call instinct.'; + +set Words['thelostworld002']['book'] = 'The Lost World'; +set Words['thelostworld002']['para'] = 'I always liked McArdle, the crabbed, old, round-backed, + red-headed news editor, and I rather hoped that he liked me. Of course, Beaumont was + the real boss; but he lived in the rarefied atmosphere of some Olympian + height from which he could distinguish nothing smaller than an + international crisis or a split in the Cabinet. Sometimes we saw him + passing in lonely majesty to his inner sanctum, with his eyes staring + vaguely and his mind hovering over the Balkans or the Persian Gulf. He + was above and beyond us. But McArdle was his first lieutenant, and it + was he that we knew. The old man nodded as I entered the room, and he + pushed his spectacles far up on his bald forehead.'; + +*/ diff --git a/examples/src/main/scala/spark/examples/ExceptionHandlingTest.scala b/examples/src/main/scala/spark/examples/ExceptionHandlingTest.scala index c89f3dac0c..21a90f2e5a 100644 --- a/examples/src/main/scala/spark/examples/ExceptionHandlingTest.scala +++ b/examples/src/main/scala/spark/examples/ExceptionHandlingTest.scala @@ -9,9 +9,10 @@ object ExceptionHandlingTest { System.exit(1) } - val sc = new SparkContext(args(0), "ExceptionHandlingTest") + val sc = new SparkContext(args(0), "ExceptionHandlingTest", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) sc.parallelize(0 until sc.defaultParallelism).foreach { i => - if (Math.random > 0.75) + if (math.random > 0.75) throw new Exception("Testing exception handling") } diff --git a/examples/src/main/scala/spark/examples/GroupByTest.scala b/examples/src/main/scala/spark/examples/GroupByTest.scala index 86dfba3a40..a6603653f1 100644 --- a/examples/src/main/scala/spark/examples/GroupByTest.scala +++ b/examples/src/main/scala/spark/examples/GroupByTest.scala @@ -9,14 +9,15 @@ object GroupByTest { if (args.length == 0) { System.err.println("Usage: GroupByTest <master> [numMappers] [numKVPairs] [KeySize] [numReducers]") System.exit(1) - } + } var numMappers = if (args.length > 1) args(1).toInt else 2 var numKVPairs = if (args.length > 2) args(2).toInt else 1000 var valSize = if (args.length > 3) args(3).toInt else 1000 var numReducers = if (args.length > 4) args(4).toInt else numMappers - val sc = new SparkContext(args(0), "GroupBy Test") + val sc = new SparkContext(args(0), "GroupBy Test", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val pairs1 = sc.parallelize(0 until numMappers, numMappers).flatMap { p => val ranGen = new Random diff --git a/examples/src/main/scala/spark/examples/HBaseTest.scala b/examples/src/main/scala/spark/examples/HBaseTest.scala new file mode 100644 index 0000000000..6e910154d4 --- /dev/null +++ b/examples/src/main/scala/spark/examples/HBaseTest.scala @@ -0,0 +1,35 @@ +package spark.examples + +import spark._ +import spark.rdd.NewHadoopRDD +import org.apache.hadoop.hbase.{HBaseConfiguration, HTableDescriptor} +import org.apache.hadoop.hbase.client.HBaseAdmin +import org.apache.hadoop.hbase.mapreduce.TableInputFormat + +object HBaseTest { + def main(args: Array[String]) { + val sc = new SparkContext(args(0), "HBaseTest", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + + val conf = HBaseConfiguration.create() + + // Other options for configuring scan behavior are available. More information available at + // http://hbase.apache.org/apidocs/org/apache/hadoop/hbase/mapreduce/TableInputFormat.html + conf.set(TableInputFormat.INPUT_TABLE, args(1)) + + // Initialize hBase table if necessary + val admin = new HBaseAdmin(conf) + if(!admin.isTableAvailable(args(1))) { + val tableDesc = new HTableDescriptor(args(1)) + admin.createTable(tableDesc) + } + + val hBaseRDD = sc.newAPIHadoopRDD(conf, classOf[TableInputFormat], + classOf[org.apache.hadoop.hbase.io.ImmutableBytesWritable], + classOf[org.apache.hadoop.hbase.client.Result]) + + hBaseRDD.count() + + System.exit(0) + } +}
\ No newline at end of file diff --git a/examples/src/main/scala/spark/examples/HdfsTest.scala b/examples/src/main/scala/spark/examples/HdfsTest.scala index 7a4530609d..dd61c467f7 100644 --- a/examples/src/main/scala/spark/examples/HdfsTest.scala +++ b/examples/src/main/scala/spark/examples/HdfsTest.scala @@ -4,7 +4,8 @@ import spark._ object HdfsTest { def main(args: Array[String]) { - val sc = new SparkContext(args(0), "HdfsTest") + val sc = new SparkContext(args(0), "HdfsTest", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val file = sc.textFile(args(1)) val mapped = file.map(s => s.length).cache() for (iter <- 1 to 10) { diff --git a/examples/src/main/scala/spark/examples/LocalALS.scala b/examples/src/main/scala/spark/examples/LocalALS.scala index 10e03359c9..2de810e062 100644 --- a/examples/src/main/scala/spark/examples/LocalALS.scala +++ b/examples/src/main/scala/spark/examples/LocalALS.scala @@ -1,11 +1,13 @@ package spark.examples -import java.util.Random import scala.math.sqrt import cern.jet.math._ import cern.colt.matrix._ import cern.colt.matrix.linalg._ +/** + * Alternating least squares matrix factorization. + */ object LocalALS { // Parameters set through command line arguments var M = 0 // Number of movies diff --git a/examples/src/main/scala/spark/examples/LocalKMeans.scala b/examples/src/main/scala/spark/examples/LocalKMeans.scala index b442c604cd..4849f216fb 100644 --- a/examples/src/main/scala/spark/examples/LocalKMeans.scala +++ b/examples/src/main/scala/spark/examples/LocalKMeans.scala @@ -6,74 +6,77 @@ import spark.SparkContext._ import scala.collection.mutable.HashMap import scala.collection.mutable.HashSet +/** + * K-means clustering. + */ object LocalKMeans { - val N = 1000 - val R = 1000 // Scaling factor - val D = 10 - val K = 10 - val convergeDist = 0.001 - val rand = new Random(42) - - def generateData = { - def generatePoint(i: Int) = { - Vector(D, _ => rand.nextDouble * R) - } - Array.tabulate(N)(generatePoint) - } - - def closestPoint(p: Vector, centers: HashMap[Int, Vector]): Int = { - var index = 0 - var bestIndex = 0 - var closest = Double.PositiveInfinity - - for (i <- 1 to centers.size) { - val vCurr = centers.get(i).get - val tempDist = p.squaredDist(vCurr) - if (tempDist < closest) { - closest = tempDist - bestIndex = i - } - } - - return bestIndex - } - - def main(args: Array[String]) { - val data = generateData - var points = new HashSet[Vector] - var kPoints = new HashMap[Int, Vector] - var tempDist = 1.0 - - while (points.size < K) { - points.add(data(rand.nextInt(N))) - } - - val iter = points.iterator - for (i <- 1 to points.size) { - kPoints.put(i, iter.next()) - } - - println("Initial centers: " + kPoints) - - while(tempDist > convergeDist) { - var closest = data.map (p => (closestPoint(p, kPoints), (p, 1))) - - var mappings = closest.groupBy[Int] (x => x._1) - - var pointStats = mappings.map(pair => pair._2.reduceLeft [(Int, (Vector, Int))] {case ((id1, (x1, y1)), (id2, (x2, y2))) => (id1, (x1 + x2, y1+y2))}) - - var newPoints = pointStats.map {mapping => (mapping._1, mapping._2._1/mapping._2._2)} - - tempDist = 0.0 - for (mapping <- newPoints) { - tempDist += kPoints.get(mapping._1).get.squaredDist(mapping._2) - } - - for (newP <- newPoints) { - kPoints.put(newP._1, newP._2) - } - } - - println("Final centers: " + kPoints) - } + val N = 1000 + val R = 1000 // Scaling factor + val D = 10 + val K = 10 + val convergeDist = 0.001 + val rand = new Random(42) + + def generateData = { + def generatePoint(i: Int) = { + Vector(D, _ => rand.nextDouble * R) + } + Array.tabulate(N)(generatePoint) + } + + def closestPoint(p: Vector, centers: HashMap[Int, Vector]): Int = { + var index = 0 + var bestIndex = 0 + var closest = Double.PositiveInfinity + + for (i <- 1 to centers.size) { + val vCurr = centers.get(i).get + val tempDist = p.squaredDist(vCurr) + if (tempDist < closest) { + closest = tempDist + bestIndex = i + } + } + + return bestIndex + } + + def main(args: Array[String]) { + val data = generateData + var points = new HashSet[Vector] + var kPoints = new HashMap[Int, Vector] + var tempDist = 1.0 + + while (points.size < K) { + points.add(data(rand.nextInt(N))) + } + + val iter = points.iterator + for (i <- 1 to points.size) { + kPoints.put(i, iter.next()) + } + + println("Initial centers: " + kPoints) + + while(tempDist > convergeDist) { + var closest = data.map (p => (closestPoint(p, kPoints), (p, 1))) + + var mappings = closest.groupBy[Int] (x => x._1) + + var pointStats = mappings.map(pair => pair._2.reduceLeft [(Int, (Vector, Int))] {case ((id1, (x1, y1)), (id2, (x2, y2))) => (id1, (x1 + x2, y1+y2))}) + + var newPoints = pointStats.map {mapping => (mapping._1, mapping._2._1/mapping._2._2)} + + tempDist = 0.0 + for (mapping <- newPoints) { + tempDist += kPoints.get(mapping._1).get.squaredDist(mapping._2) + } + + for (newP <- newPoints) { + kPoints.put(newP._1, newP._2) + } + } + + println("Final centers: " + kPoints) + } } diff --git a/examples/src/main/scala/spark/examples/LocalLR.scala b/examples/src/main/scala/spark/examples/LocalLR.scala index 9553162004..cd73f553d6 100644 --- a/examples/src/main/scala/spark/examples/LocalLR.scala +++ b/examples/src/main/scala/spark/examples/LocalLR.scala @@ -3,6 +3,9 @@ package spark.examples import java.util.Random import spark.util.Vector +/** + * Logistic regression based classification. + */ object LocalLR { val N = 10000 // Number of data points val D = 10 // Number of dimensions diff --git a/examples/src/main/scala/spark/examples/LogQuery.scala b/examples/src/main/scala/spark/examples/LogQuery.scala index 5330b8da94..6497596d35 100644 --- a/examples/src/main/scala/spark/examples/LogQuery.scala +++ b/examples/src/main/scala/spark/examples/LogQuery.scala @@ -26,7 +26,9 @@ object LogQuery { System.err.println("Usage: LogQuery <master> [logFile]") System.exit(1) } - val sc = new SparkContext(args(0), "Log Query") + + val sc = new SparkContext(args(0), "Log Query", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val dataSet = if (args.length == 2) sc.textFile(args(1)) diff --git a/examples/src/main/scala/spark/examples/MultiBroadcastTest.scala b/examples/src/main/scala/spark/examples/MultiBroadcastTest.scala index 83ae014e94..a0aaf60918 100644 --- a/examples/src/main/scala/spark/examples/MultiBroadcastTest.scala +++ b/examples/src/main/scala/spark/examples/MultiBroadcastTest.scala @@ -8,25 +8,29 @@ object MultiBroadcastTest { System.err.println("Usage: BroadcastTest <master> [<slices>] [numElem]") System.exit(1) } - - val spark = new SparkContext(args(0), "Broadcast Test") + + val sc = new SparkContext(args(0), "Broadcast Test", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + val slices = if (args.length > 1) args(1).toInt else 2 val num = if (args.length > 2) args(2).toInt else 1000000 var arr1 = new Array[Int](num) - for (i <- 0 until arr1.length) + for (i <- 0 until arr1.length) { arr1(i) = i - + } + var arr2 = new Array[Int](num) - for (i <- 0 until arr2.length) + for (i <- 0 until arr2.length) { arr2(i) = i + } - val barr1 = spark.broadcast(arr1) - val barr2 = spark.broadcast(arr2) - spark.parallelize(1 to 10, slices).foreach { + val barr1 = sc.broadcast(arr1) + val barr2 = sc.broadcast(arr2) + sc.parallelize(1 to 10, slices).foreach { i => println(barr1.value.size + barr2.value.size) } - + System.exit(0) } } diff --git a/examples/src/main/scala/spark/examples/SimpleSkewedGroupByTest.scala b/examples/src/main/scala/spark/examples/SimpleSkewedGroupByTest.scala index 50b3a263b4..461b84a2c6 100644 --- a/examples/src/main/scala/spark/examples/SimpleSkewedGroupByTest.scala +++ b/examples/src/main/scala/spark/examples/SimpleSkewedGroupByTest.scala @@ -11,15 +11,16 @@ object SimpleSkewedGroupByTest { "[numMappers] [numKVPairs] [valSize] [numReducers] [ratio]") System.exit(1) } - + var numMappers = if (args.length > 1) args(1).toInt else 2 var numKVPairs = if (args.length > 2) args(2).toInt else 1000 var valSize = if (args.length > 3) args(3).toInt else 1000 var numReducers = if (args.length > 4) args(4).toInt else numMappers var ratio = if (args.length > 5) args(5).toInt else 5.0 - val sc = new SparkContext(args(0), "GroupBy Test") - + val sc = new SparkContext(args(0), "GroupBy Test", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + val pairs1 = sc.parallelize(0 until numMappers, numMappers).flatMap { p => val ranGen = new Random var result = new Array[(Int, Array[Byte])](numKVPairs) diff --git a/examples/src/main/scala/spark/examples/SkewedGroupByTest.scala b/examples/src/main/scala/spark/examples/SkewedGroupByTest.scala index d2117a263e..435675f9de 100644 --- a/examples/src/main/scala/spark/examples/SkewedGroupByTest.scala +++ b/examples/src/main/scala/spark/examples/SkewedGroupByTest.scala @@ -10,19 +10,20 @@ object SkewedGroupByTest { System.err.println("Usage: GroupByTest <master> [numMappers] [numKVPairs] [KeySize] [numReducers]") System.exit(1) } - + var numMappers = if (args.length > 1) args(1).toInt else 2 var numKVPairs = if (args.length > 2) args(2).toInt else 1000 var valSize = if (args.length > 3) args(3).toInt else 1000 var numReducers = if (args.length > 4) args(4).toInt else numMappers - val sc = new SparkContext(args(0), "GroupBy Test") - + val sc = new SparkContext(args(0), "GroupBy Test", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + val pairs1 = sc.parallelize(0 until numMappers, numMappers).flatMap { p => val ranGen = new Random // map output sizes lineraly increase from the 1st to the last - numKVPairs = (1. * (p + 1) / numMappers * numKVPairs).toInt + numKVPairs = (1.0 * (p + 1) / numMappers * numKVPairs).toInt var arr1 = new Array[(Int, Array[Byte])](numKVPairs) for (i <- 0 until numKVPairs) { @@ -31,11 +32,11 @@ object SkewedGroupByTest { arr1(i) = (ranGen.nextInt(Int.MaxValue), byteArr) } arr1 - }.cache + }.cache() // Enforce that everything has been calculated and in cache - pairs1.count + pairs1.count() - println(pairs1.groupByKey(numReducers).count) + println(pairs1.groupByKey(numReducers).count()) System.exit(0) } diff --git a/examples/src/main/scala/spark/examples/SparkALS.scala b/examples/src/main/scala/spark/examples/SparkALS.scala index 5e01885dbb..8fb3b0fb2a 100644 --- a/examples/src/main/scala/spark/examples/SparkALS.scala +++ b/examples/src/main/scala/spark/examples/SparkALS.scala @@ -1,14 +1,14 @@ package spark.examples -import java.io.Serializable -import java.util.Random import scala.math.sqrt import cern.jet.math._ import cern.colt.matrix._ import cern.colt.matrix.linalg._ import spark._ -import scala.Option +/** + * Alternating least squares matrix factorization. + */ object SparkALS { // Parameters set through command line arguments var M = 0 // Number of movies @@ -70,30 +70,32 @@ object SparkALS { } def main(args: Array[String]) { + if (args.length == 0) { + System.err.println("Usage: SparkALS <master> [<M> <U> <F> <iters> <slices>]") + System.exit(1) + } + var host = "" var slices = 0 - (0 to 5).map(i => { - i match { - case a if a < args.length => Some(args(a)) - case _ => None - } - }).toArray match { - case Array(host_, m, u, f, iters, slices_) => { - host = host_ getOrElse "local" - M = (m getOrElse "100").toInt - U = (u getOrElse "500").toInt - F = (f getOrElse "10").toInt - ITERATIONS = (iters getOrElse "5").toInt - slices = (slices_ getOrElse "2").toInt - } - case _ => { - System.err.println("Usage: SparkALS [<master> <M> <U> <F> <iters> <slices>]") + val options = (0 to 5).map(i => if (i < args.length) Some(args(i)) else None) + + options.toArray match { + case Array(host_, m, u, f, iters, slices_) => + host = host_.get + M = m.getOrElse("100").toInt + U = u.getOrElse("500").toInt + F = f.getOrElse("10").toInt + ITERATIONS = iters.getOrElse("5").toInt + slices = slices_.getOrElse("2").toInt + case _ => + System.err.println("Usage: SparkALS <master> [<M> <U> <F> <iters> <slices>]") System.exit(1) - } } printf("Running with M=%d, U=%d, F=%d, iters=%d\n", M, U, F, ITERATIONS) - val spark = new SparkContext(host, "SparkALS") + + val sc = new SparkContext(host, "SparkALS", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val R = generateR() @@ -102,19 +104,19 @@ object SparkALS { var us = Array.fill(U)(factory1D.random(F)) // Iteratively update movies then users - val Rc = spark.broadcast(R) - var msc = spark.broadcast(ms) - var usc = spark.broadcast(us) + val Rc = sc.broadcast(R) + var msb = sc.broadcast(ms) + var usb = sc.broadcast(us) for (iter <- 1 to ITERATIONS) { println("Iteration " + iter + ":") - ms = spark.parallelize(0 until M, slices) - .map(i => update(i, msc.value(i), usc.value, Rc.value)) + ms = sc.parallelize(0 until M, slices) + .map(i => update(i, msb.value(i), usb.value, Rc.value)) .toArray - msc = spark.broadcast(ms) // Re-broadcast ms because it was updated - us = spark.parallelize(0 until U, slices) - .map(i => update(i, usc.value(i), msc.value, algebra.transpose(Rc.value))) + msb = sc.broadcast(ms) // Re-broadcast ms because it was updated + us = sc.parallelize(0 until U, slices) + .map(i => update(i, usb.value(i), msb.value, algebra.transpose(Rc.value))) .toArray - usc = spark.broadcast(us) // Re-broadcast us because it was updated + usb = sc.broadcast(us) // Re-broadcast us because it was updated println("RMSE = " + rmse(R, ms, us)) println() } diff --git a/examples/src/main/scala/spark/examples/SparkHdfsLR.scala b/examples/src/main/scala/spark/examples/SparkHdfsLR.scala index 5b2bc84d69..3d080a0257 100644 --- a/examples/src/main/scala/spark/examples/SparkHdfsLR.scala +++ b/examples/src/main/scala/spark/examples/SparkHdfsLR.scala @@ -4,7 +4,12 @@ import java.util.Random import scala.math.exp import spark.util.Vector import spark._ +import spark.deploy.SparkHadoopUtil +import spark.scheduler.InputFormatInfo +/** + * Logistic regression based classification. + */ object SparkHdfsLR { val D = 10 // Numer of dimensions val rand = new Random(42) @@ -29,8 +34,13 @@ object SparkHdfsLR { System.err.println("Usage: SparkHdfsLR <master> <file> <iters>") System.exit(1) } - val sc = new SparkContext(args(0), "SparkHdfsLR") - val lines = sc.textFile(args(1)) + val inputPath = args(1) + val conf = SparkHadoopUtil.newConfiguration() + val sc = new SparkContext(args(0), "SparkHdfsLR", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR")), Map(), + InputFormatInfo.computePreferredLocations( + Seq(new InputFormatInfo(conf, classOf[org.apache.hadoop.mapred.TextInputFormat], inputPath)))) + val lines = sc.textFile(inputPath) val points = lines.map(parsePoint _).cache() val ITERATIONS = args(2).toInt diff --git a/examples/src/main/scala/spark/examples/SparkKMeans.scala b/examples/src/main/scala/spark/examples/SparkKMeans.scala index 6375961390..4161c59fea 100644 --- a/examples/src/main/scala/spark/examples/SparkKMeans.scala +++ b/examples/src/main/scala/spark/examples/SparkKMeans.scala @@ -7,6 +7,9 @@ import spark.SparkContext._ import scala.collection.mutable.HashMap import scala.collection.mutable.HashSet +/** + * K-means clustering. + */ object SparkKMeans { val R = 1000 // Scaling factor val rand = new Random(42) @@ -36,7 +39,8 @@ object SparkKMeans { System.err.println("Usage: SparkLocalKMeans <master> <file> <k> <convergeDist>") System.exit(1) } - val sc = new SparkContext(args(0), "SparkLocalKMeans") + val sc = new SparkContext(args(0), "SparkLocalKMeans", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val lines = sc.textFile(args(1)) val data = lines.map(parseVector _).cache() val K = args(2).toInt @@ -60,6 +64,7 @@ object SparkKMeans { for (newP <- newPoints) { kPoints(newP._1) = newP._2 } + println("Finished iteration (delta = " + tempDist + ")") } println("Final centers:") diff --git a/examples/src/main/scala/spark/examples/SparkLR.scala b/examples/src/main/scala/spark/examples/SparkLR.scala index aaaf062c8f..2f41aeb376 100644 --- a/examples/src/main/scala/spark/examples/SparkLR.scala +++ b/examples/src/main/scala/spark/examples/SparkLR.scala @@ -5,6 +5,9 @@ import scala.math.exp import spark.util.Vector import spark._ +/** + * Logistic regression based classification. + */ object SparkLR { val N = 10000 // Number of data points val D = 10 // Numer of dimensions @@ -28,7 +31,8 @@ object SparkLR { System.err.println("Usage: SparkLR <master> [<slices>]") System.exit(1) } - val sc = new SparkContext(args(0), "SparkLR") + val sc = new SparkContext(args(0), "SparkLR", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val numSlices = if (args.length > 1) args(1).toInt else 2 val points = sc.parallelize(generateData, numSlices).cache() diff --git a/examples/src/main/scala/spark/examples/SparkPi.scala b/examples/src/main/scala/spark/examples/SparkPi.scala index 2f226f1380..f598d2ff9c 100644 --- a/examples/src/main/scala/spark/examples/SparkPi.scala +++ b/examples/src/main/scala/spark/examples/SparkPi.scala @@ -4,13 +4,15 @@ import scala.math.random import spark._ import SparkContext._ +/** Computes an approximation to pi */ object SparkPi { def main(args: Array[String]) { if (args.length == 0) { System.err.println("Usage: SparkPi <master> [<slices>]") System.exit(1) } - val spark = new SparkContext(args(0), "SparkPi") + val spark = new SparkContext(args(0), "SparkPi", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val slices = if (args.length > 1) args(1).toInt else 2 val n = 100000 * slices val count = spark.parallelize(1 to n, slices).map { i => diff --git a/examples/src/main/scala/spark/examples/SparkTC.scala b/examples/src/main/scala/spark/examples/SparkTC.scala index 90bae011ad..911ae8f168 100644 --- a/examples/src/main/scala/spark/examples/SparkTC.scala +++ b/examples/src/main/scala/spark/examples/SparkTC.scala @@ -9,7 +9,6 @@ import scala.collection.mutable * Transitive closure on a graph. */ object SparkTC { - val numEdges = 200 val numVertices = 100 val rand = new Random(42) @@ -29,7 +28,8 @@ object SparkTC { System.err.println("Usage: SparkTC <master> [<slices>]") System.exit(1) } - val spark = new SparkContext(args(0), "SparkTC") + val spark = new SparkContext(args(0), "SparkTC", + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) val slices = if (args.length > 1) args(1).toInt else 2 var tc = spark.parallelize(generateGraph, slices).cache() diff --git a/examples/src/main/scala/spark/streaming/examples/ActorWordCount.scala b/examples/src/main/scala/spark/streaming/examples/ActorWordCount.scala new file mode 100644 index 0000000000..3b847fe603 --- /dev/null +++ b/examples/src/main/scala/spark/streaming/examples/ActorWordCount.scala @@ -0,0 +1,158 @@ +package spark.streaming.examples + +import scala.collection.mutable.LinkedList +import scala.util.Random + +import akka.actor.Actor +import akka.actor.ActorRef +import akka.actor.Props +import akka.actor.actorRef2Scala + +import spark.streaming.Seconds +import spark.streaming.StreamingContext +import spark.streaming.StreamingContext.toPairDStreamFunctions +import spark.streaming.receivers.Receiver +import spark.util.AkkaUtils + +case class SubscribeReceiver(receiverActor: ActorRef) +case class UnsubscribeReceiver(receiverActor: ActorRef) + +/** + * Sends the random content to every receiver subscribed with 1/2 + * second delay. + */ +class FeederActor extends Actor { + + val rand = new Random() + var receivers: LinkedList[ActorRef] = new LinkedList[ActorRef]() + + val strings: Array[String] = Array("words ", "may ", "count ") + + def makeMessage(): String = { + val x = rand.nextInt(3) + strings(x) + strings(2 - x) + } + + /* + * A thread to generate random messages + */ + new Thread() { + override def run() { + while (true) { + Thread.sleep(500) + receivers.foreach(_ ! makeMessage) + } + } + }.start() + + def receive: Receive = { + + case SubscribeReceiver(receiverActor: ActorRef) => + println("received subscribe from %s".format(receiverActor.toString)) + receivers = LinkedList(receiverActor) ++ receivers + + case UnsubscribeReceiver(receiverActor: ActorRef) => + println("received unsubscribe from %s".format(receiverActor.toString)) + receivers = receivers.dropWhile(x => x eq receiverActor) + + } +} + +/** + * A sample actor as receiver, is also simplest. This receiver actor + * goes and subscribe to a typical publisher/feeder actor and receives + * data. + * + * @see [[spark.streaming.examples.FeederActor]] + */ +class SampleActorReceiver[T: ClassManifest](urlOfPublisher: String) +extends Actor with Receiver { + + lazy private val remotePublisher = context.actorFor(urlOfPublisher) + + override def preStart = remotePublisher ! SubscribeReceiver(context.self) + + def receive = { + case msg ⇒ context.parent ! pushBlock(msg.asInstanceOf[T]) + } + + override def postStop() = remotePublisher ! UnsubscribeReceiver(context.self) + +} + +/** + * A sample feeder actor + * + * Usage: FeederActor <hostname> <port> + * <hostname> and <port> describe the AkkaSystem that Spark Sample feeder would start on. + */ +object FeederActor { + + def main(args: Array[String]) { + if(args.length < 2){ + System.err.println( + "Usage: FeederActor <hostname> <port>\n" + ) + System.exit(1) + } + val Seq(host, port) = args.toSeq + + + val actorSystem = AkkaUtils.createActorSystem("test", host, port.toInt)._1 + val feeder = actorSystem.actorOf(Props[FeederActor], "FeederActor") + + println("Feeder started as:" + feeder) + + actorSystem.awaitTermination(); + } +} + +/** + * A sample word count program demonstrating the use of plugging in + * Actor as Receiver + * Usage: ActorWordCount <master> <hostname> <port> + * <master> is the Spark master URL. In local mode, <master> should be 'local[n]' with n > 1. + * <hostname> and <port> describe the AkkaSystem that Spark Sample feeder is running on. + * + * To run this example locally, you may run Feeder Actor as + * `$ ./run spark.streaming.examples.FeederActor 127.0.1.1 9999` + * and then run the example + * `$ ./run spark.streaming.examples.ActorWordCount local[2] 127.0.1.1 9999` + */ +object ActorWordCount { + def main(args: Array[String]) { + if (args.length < 3) { + System.err.println( + "Usage: ActorWordCount <master> <hostname> <port>" + + "In local mode, <master> should be 'local[n]' with n > 1") + System.exit(1) + } + + val Seq(master, host, port) = args.toSeq + + // Create the context and set the batch size + val ssc = new StreamingContext(master, "ActorWordCount", Seconds(2), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + + /* + * Following is the use of actorStream to plug in custom actor as receiver + * + * An important point to note: + * Since Actor may exist outside the spark framework, It is thus user's responsibility + * to ensure the type safety, i.e type of data received and InputDstream + * should be same. + * + * For example: Both actorStream and SampleActorReceiver are parameterized + * to same type to ensure type safety. + */ + + val lines = ssc.actorStream[String]( + Props(new SampleActorReceiver[String]("akka://test@%s:%s/user/FeederActor".format( + host, port.toInt))), "SampleReceiver") + + //compute wordcount + lines.flatMap(_.split("\\s+")).map(x => (x, 1)).reduceByKey(_ + _).print() + + ssc.start() + } +} diff --git a/examples/src/main/scala/spark/streaming/examples/FlumeEventCount.scala b/examples/src/main/scala/spark/streaming/examples/FlumeEventCount.scala index 461929fba2..39c76fd98a 100644 --- a/examples/src/main/scala/spark/streaming/examples/FlumeEventCount.scala +++ b/examples/src/main/scala/spark/streaming/examples/FlumeEventCount.scala @@ -30,7 +30,8 @@ object FlumeEventCount { val batchInterval = Milliseconds(2000) // Create the context and set the batch size - val ssc = new StreamingContext(master, "FlumeEventCount", batchInterval) + val ssc = new StreamingContext(master, "FlumeEventCount", batchInterval, + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) // Create a flume stream val stream = ssc.flumeStream(host,port,StorageLevel.MEMORY_ONLY) diff --git a/examples/src/main/scala/spark/streaming/examples/HdfsWordCount.scala b/examples/src/main/scala/spark/streaming/examples/HdfsWordCount.scala index 8530f5c175..9389f8a38d 100644 --- a/examples/src/main/scala/spark/streaming/examples/HdfsWordCount.scala +++ b/examples/src/main/scala/spark/streaming/examples/HdfsWordCount.scala @@ -22,7 +22,8 @@ object HdfsWordCount { } // Create the context - val ssc = new StreamingContext(args(0), "HdfsWordCount", Seconds(2)) + val ssc = new StreamingContext(args(0), "HdfsWordCount", Seconds(2), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) // Create the FileInputDStream on the directory and use the // stream to count words in new files created diff --git a/examples/src/main/scala/spark/streaming/examples/KafkaWordCount.scala b/examples/src/main/scala/spark/streaming/examples/KafkaWordCount.scala index fe55db6e2c..9202e65e09 100644 --- a/examples/src/main/scala/spark/streaming/examples/KafkaWordCount.scala +++ b/examples/src/main/scala/spark/streaming/examples/KafkaWordCount.scala @@ -10,22 +10,34 @@ import spark.streaming.StreamingContext._ import spark.storage.StorageLevel import spark.streaming.util.RawTextHelper._ +/** + * Consumes messages from one or more topics in Kafka and does wordcount. + * Usage: KafkaWordCount <master> <zkQuorum> <group> <topics> <numThreads> + * <master> is the Spark master URL. In local mode, <master> should be 'local[n]' with n > 1. + * <zkQuorum> is a list of one or more zookeeper servers that make quorum + * <group> is the name of kafka consumer group + * <topics> is a list of one or more kafka topics to consume from + * <numThreads> is the number of threads the kafka consumer should use + * + * Example: + * `./run spark.streaming.examples.KafkaWordCount local[2] zoo01,zoo02,zoo03 my-consumer-group topic1,topic2 1` + */ object KafkaWordCount { def main(args: Array[String]) { - if (args.length < 6) { - System.err.println("Usage: KafkaWordCount <master> <hostname> <port> <group> <topics> <numThreads>") + if (args.length < 5) { + System.err.println("Usage: KafkaWordCount <master> <zkQuorum> <group> <topics> <numThreads>") System.exit(1) } - val Array(master, hostname, port, group, topics, numThreads) = args + val Array(master, zkQuorum, group, topics, numThreads) = args - val sc = new SparkContext(master, "KafkaWordCount") - val ssc = new StreamingContext(sc, Seconds(2)) + val ssc = new StreamingContext(master, "KafkaWordCount", Seconds(2), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) ssc.checkpoint("checkpoint") val topicpMap = topics.split(",").map((_,numThreads.toInt)).toMap - val lines = ssc.kafkaStream[String](hostname, port.toInt, group, topicpMap) + val lines = ssc.kafkaStream(zkQuorum, group, topicpMap) val words = lines.flatMap(_.split(" ")) val wordCounts = words.map(x => (x, 1l)).reduceByKeyAndWindow(add _, subtract _, Minutes(10), Seconds(2), 2) wordCounts.print() @@ -38,16 +50,16 @@ object KafkaWordCount { object KafkaWordCountProducer { def main(args: Array[String]) { - if (args.length < 3) { - System.err.println("Usage: KafkaWordCountProducer <hostname> <port> <topic> <messagesPerSec> <wordsPerMessage>") + if (args.length < 2) { + System.err.println("Usage: KafkaWordCountProducer <zkQuorum> <topic> <messagesPerSec> <wordsPerMessage>") System.exit(1) } - val Array(hostname, port, topic, messagesPerSec, wordsPerMessage) = args + val Array(zkQuorum, topic, messagesPerSec, wordsPerMessage) = args // Zookeper connection properties val props = new Properties() - props.put("zk.connect", hostname + ":" + port) + props.put("zk.connect", zkQuorum) props.put("serializer.class", "kafka.serializer.StringEncoder") val config = new ProducerConfig(props) diff --git a/examples/src/main/scala/spark/streaming/examples/NetworkWordCount.scala b/examples/src/main/scala/spark/streaming/examples/NetworkWordCount.scala index 32f7d57bea..704540c2bf 100644 --- a/examples/src/main/scala/spark/streaming/examples/NetworkWordCount.scala +++ b/examples/src/main/scala/spark/streaming/examples/NetworkWordCount.scala @@ -16,18 +16,19 @@ import spark.streaming.StreamingContext._ */ object NetworkWordCount { def main(args: Array[String]) { - if (args.length < 2) { + if (args.length < 3) { System.err.println("Usage: NetworkWordCount <master> <hostname> <port>\n" + "In local mode, <master> should be 'local[n]' with n > 1") System.exit(1) } // Create the context with a 1 second batch size - val ssc = new StreamingContext(args(0), "NetworkWordCount", Seconds(1)) + val ssc = new StreamingContext(args(0), "NetworkWordCount", Seconds(1), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) // Create a NetworkInputDStream on target ip:port and count the // words in input stream of \n delimited test (eg. generated by 'nc') - val lines = ssc.networkTextStream(args(1), args(2).toInt) + val lines = ssc.socketTextStream(args(1), args(2).toInt) val words = lines.flatMap(_.split(" ")) val wordCounts = words.map(x => (x, 1)).reduceByKey(_ + _) wordCounts.print() diff --git a/examples/src/main/scala/spark/streaming/examples/QueueStream.scala b/examples/src/main/scala/spark/streaming/examples/QueueStream.scala index 2a265d021d..f450e21040 100644 --- a/examples/src/main/scala/spark/streaming/examples/QueueStream.scala +++ b/examples/src/main/scala/spark/streaming/examples/QueueStream.scala @@ -15,7 +15,8 @@ object QueueStream { } // Create the context - val ssc = new StreamingContext(args(0), "QueueStream", Seconds(1)) + val ssc = new StreamingContext(args(0), "QueueStream", Seconds(1), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) // Create the queue through which RDDs can be pushed to // a QueueInputDStream @@ -30,10 +31,10 @@ object QueueStream { // Create and push some RDDs into for (i <- 1 to 30) { - rddQueue += ssc.sc.makeRDD(1 to 1000, 10) + rddQueue += ssc.sparkContext.makeRDD(1 to 1000, 10) Thread.sleep(1000) } ssc.stop() System.exit(0) } -}
\ No newline at end of file +} diff --git a/examples/src/main/scala/spark/streaming/examples/RawNetworkGrep.scala b/examples/src/main/scala/spark/streaming/examples/RawNetworkGrep.scala index 2eec777c54..175281e095 100644 --- a/examples/src/main/scala/spark/streaming/examples/RawNetworkGrep.scala +++ b/examples/src/main/scala/spark/streaming/examples/RawNetworkGrep.scala @@ -31,13 +31,14 @@ object RawNetworkGrep { val Array(master, IntParam(numStreams), host, IntParam(port), IntParam(batchMillis)) = args // Create the context - val ssc = new StreamingContext(master, "RawNetworkGrep", Milliseconds(batchMillis)) + val ssc = new StreamingContext(master, "RawNetworkGrep", Milliseconds(batchMillis), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) // Warm up the JVMs on master and slave for JIT compilation to kick in - RawTextHelper.warmUp(ssc.sc) + RawTextHelper.warmUp(ssc.sparkContext) val rawStreams = (1 to numStreams).map(_ => - ssc.rawNetworkStream[String](host, port, StorageLevel.MEMORY_ONLY_SER_2)).toArray + ssc.rawSocketStream[String](host, port, StorageLevel.MEMORY_ONLY_SER_2)).toArray val union = ssc.union(rawStreams) union.filter(_.contains("the")).count().foreach(r => println("Grep count: " + r.collect().mkString)) diff --git a/examples/src/main/scala/spark/streaming/examples/TwitterAlgebirdCMS.scala b/examples/src/main/scala/spark/streaming/examples/TwitterAlgebirdCMS.scala new file mode 100644 index 0000000000..a9642100e3 --- /dev/null +++ b/examples/src/main/scala/spark/streaming/examples/TwitterAlgebirdCMS.scala @@ -0,0 +1,94 @@ +package spark.streaming.examples + +import spark.streaming.{Seconds, StreamingContext} +import spark.storage.StorageLevel +import com.twitter.algebird._ +import spark.streaming.StreamingContext._ +import spark.SparkContext._ + +/** + * Illustrates the use of the Count-Min Sketch, from Twitter's Algebird library, to compute + * windowed and global Top-K estimates of user IDs occurring in a Twitter stream. + * <br> + * <strong>Note</strong> that since Algebird's implementation currently only supports Long inputs, + * the example operates on Long IDs. Once the implementation supports other inputs (such as String), + * the same approach could be used for computing popular topics for example. + * <p> + * <p> + * <a href="http://highlyscalable.wordpress.com/2012/05/01/probabilistic-structures-web-analytics-data-mining/"> + * This blog post</a> has a good overview of the Count-Min Sketch (CMS). The CMS is a datastructure + * for approximate frequency estimation in data streams (e.g. Top-K elements, frequency of any given element, etc), + * that uses space sub-linear in the number of elements in the stream. Once elements are added to the CMS, the + * estimated count of an element can be computed, as well as "heavy-hitters" that occur more than a threshold + * percentage of the overall total count. + * <p><p> + * Algebird's implementation is a monoid, so we can succinctly merge two CMS instances in the reduce operation. + */ +object TwitterAlgebirdCMS { + def main(args: Array[String]) { + if (args.length < 3) { + System.err.println("Usage: TwitterAlgebirdCMS <master> <twitter_username> <twitter_password>" + + " [filter1] [filter2] ... [filter n]") + System.exit(1) + } + + // CMS parameters + val DELTA = 1E-3 + val EPS = 0.01 + val SEED = 1 + val PERC = 0.001 + // K highest frequency elements to take + val TOPK = 10 + + val Array(master, username, password) = args.slice(0, 3) + val filters = args.slice(3, args.length) + + val ssc = new StreamingContext(master, "TwitterAlgebirdCMS", Seconds(10), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + val stream = ssc.twitterStream(username, password, filters, StorageLevel.MEMORY_ONLY_SER) + + val users = stream.map(status => status.getUser.getId) + + val cms = new CountMinSketchMonoid(EPS, DELTA, SEED, PERC) + var globalCMS = cms.zero + val mm = new MapMonoid[Long, Int]() + var globalExact = Map[Long, Int]() + + val approxTopUsers = users.mapPartitions(ids => { + ids.map(id => cms.create(id)) + }).reduce(_ ++ _) + + val exactTopUsers = users.map(id => (id, 1)) + .reduceByKey((a, b) => a + b) + + approxTopUsers.foreach(rdd => { + if (rdd.count() != 0) { + val partial = rdd.first() + val partialTopK = partial.heavyHitters.map(id => + (id, partial.frequency(id).estimate)).toSeq.sortBy(_._2).reverse.slice(0, TOPK) + globalCMS ++= partial + val globalTopK = globalCMS.heavyHitters.map(id => + (id, globalCMS.frequency(id).estimate)).toSeq.sortBy(_._2).reverse.slice(0, TOPK) + println("Approx heavy hitters at %2.2f%% threshold this batch: %s".format(PERC, + partialTopK.mkString("[", ",", "]"))) + println("Approx heavy hitters at %2.2f%% threshold overall: %s".format(PERC, + globalTopK.mkString("[", ",", "]"))) + } + }) + + exactTopUsers.foreach(rdd => { + if (rdd.count() != 0) { + val partialMap = rdd.collect().toMap + val partialTopK = rdd.map( + {case (id, count) => (count, id)}) + .sortByKey(ascending = false).take(TOPK) + globalExact = mm.plus(globalExact.toMap, partialMap) + val globalTopK = globalExact.toSeq.sortBy(_._2).reverse.slice(0, TOPK) + println("Exact heavy hitters this batch: %s".format(partialTopK.mkString("[", ",", "]"))) + println("Exact heavy hitters overall: %s".format(globalTopK.mkString("[", ",", "]"))) + } + }) + + ssc.start() + } +} diff --git a/examples/src/main/scala/spark/streaming/examples/TwitterAlgebirdHLL.scala b/examples/src/main/scala/spark/streaming/examples/TwitterAlgebirdHLL.scala new file mode 100644 index 0000000000..f3288bfb85 --- /dev/null +++ b/examples/src/main/scala/spark/streaming/examples/TwitterAlgebirdHLL.scala @@ -0,0 +1,72 @@ +package spark.streaming.examples + +import spark.streaming.{Seconds, StreamingContext} +import spark.storage.StorageLevel +import com.twitter.algebird.HyperLogLog._ +import com.twitter.algebird.HyperLogLogMonoid +import spark.streaming.dstream.TwitterInputDStream + +/** + * Illustrates the use of the HyperLogLog algorithm, from Twitter's Algebird library, to compute + * a windowed and global estimate of the unique user IDs occurring in a Twitter stream. + * <p> + * <p> + * This <a href="http://highlyscalable.wordpress.com/2012/05/01/probabilistic-structures-web-analytics-data-mining/"> + * blog post</a> and this + * <a href="http://highscalability.com/blog/2012/4/5/big-data-counting-how-to-count-a-billion-distinct-objects-us.html">blog post</a> + * have good overviews of HyperLogLog (HLL). HLL is a memory-efficient datastructure for estimating + * the cardinality of a data stream, i.e. the number of unique elements. + * <p><p> + * Algebird's implementation is a monoid, so we can succinctly merge two HLL instances in the reduce operation. + */ +object TwitterAlgebirdHLL { + def main(args: Array[String]) { + if (args.length < 3) { + System.err.println("Usage: TwitterAlgebirdHLL <master> <twitter_username> <twitter_password>" + + " [filter1] [filter2] ... [filter n]") + System.exit(1) + } + + /** Bit size parameter for HyperLogLog, trades off accuracy vs size */ + val BIT_SIZE = 12 + val Array(master, username, password) = args.slice(0, 3) + val filters = args.slice(3, args.length) + + val ssc = new StreamingContext(master, "TwitterAlgebirdHLL", Seconds(5), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + val stream = ssc.twitterStream(username, password, filters, StorageLevel.MEMORY_ONLY_SER) + + val users = stream.map(status => status.getUser.getId) + + val hll = new HyperLogLogMonoid(BIT_SIZE) + var globalHll = hll.zero + var userSet: Set[Long] = Set() + + val approxUsers = users.mapPartitions(ids => { + ids.map(id => hll(id)) + }).reduce(_ + _) + + val exactUsers = users.map(id => Set(id)).reduce(_ ++ _) + + approxUsers.foreach(rdd => { + if (rdd.count() != 0) { + val partial = rdd.first() + globalHll += partial + println("Approx distinct users this batch: %d".format(partial.estimatedSize.toInt)) + println("Approx distinct users overall: %d".format(globalHll.estimatedSize.toInt)) + } + }) + + exactUsers.foreach(rdd => { + if (rdd.count() != 0) { + val partial = rdd.first() + userSet ++= partial + println("Exact distinct users this batch: %d".format(partial.size)) + println("Exact distinct users overall: %d".format(userSet.size)) + println("Error rate: %2.5f%%".format(((globalHll.estimatedSize / userSet.size.toDouble) - 1) * 100)) + } + }) + + ssc.start() + } +} diff --git a/examples/src/main/scala/spark/streaming/examples/twitter/TwitterBasic.scala b/examples/src/main/scala/spark/streaming/examples/TwitterPopularTags.scala index 377bc0c98e..9d4494c6f2 100644 --- a/examples/src/main/scala/spark/streaming/examples/twitter/TwitterBasic.scala +++ b/examples/src/main/scala/spark/streaming/examples/TwitterPopularTags.scala @@ -1,19 +1,19 @@ -package spark.streaming.examples.twitter +package spark.streaming.examples -import spark.streaming.StreamingContext._ import spark.streaming.{Seconds, StreamingContext} +import StreamingContext._ import spark.SparkContext._ -import spark.storage.StorageLevel /** * Calculates popular hashtags (topics) over sliding 10 and 60 second windows from a Twitter * stream. The stream is instantiated with credentials and optionally filters supplied by the * command line arguments. + * */ -object TwitterBasic { +object TwitterPopularTags { def main(args: Array[String]) { if (args.length < 3) { - System.err.println("Usage: TwitterBasic <master> <twitter_username> <twitter_password>" + + System.err.println("Usage: TwitterPopularTags <master> <twitter_username> <twitter_password>" + " [filter1] [filter2] ... [filter n]") System.exit(1) } @@ -21,10 +21,9 @@ object TwitterBasic { val Array(master, username, password) = args.slice(0, 3) val filters = args.slice(3, args.length) - val ssc = new StreamingContext(master, "TwitterBasic", Seconds(2)) - val stream = new TwitterInputDStream(ssc, username, password, filters, - StorageLevel.MEMORY_ONLY_SER) - ssc.registerInputStream(stream) + val ssc = new StreamingContext(master, "TwitterPopularTags", Seconds(2), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + val stream = ssc.twitterStream(username, password, filters) val hashTags = stream.flatMap(status => status.getText.split(" ").filter(_.startsWith("#"))) @@ -39,22 +38,17 @@ object TwitterBasic { // Print popular hashtags topCounts60.foreach(rdd => { - if (rdd.count() != 0) { - val topList = rdd.take(5) - println("\nPopular topics in last 60 seconds (%s total):".format(rdd.count())) - topList.foreach{case (count, tag) => println("%s (%s tweets)".format(tag, count))} - } + val topList = rdd.take(5) + println("\nPopular topics in last 60 seconds (%s total):".format(rdd.count())) + topList.foreach{case (count, tag) => println("%s (%s tweets)".format(tag, count))} }) topCounts10.foreach(rdd => { - if (rdd.count() != 0) { - val topList = rdd.take(5) - println("\nPopular topics in last 10 seconds (%s total):".format(rdd.count())) - topList.foreach{case (count, tag) => println("%s (%s tweets)".format(tag, count))} - } + val topList = rdd.take(5) + println("\nPopular topics in last 10 seconds (%s total):".format(rdd.count())) + topList.foreach{case (count, tag) => println("%s (%s tweets)".format(tag, count))} }) ssc.start() } - } diff --git a/examples/src/main/scala/spark/streaming/examples/ZeroMQWordCount.scala b/examples/src/main/scala/spark/streaming/examples/ZeroMQWordCount.scala new file mode 100644 index 0000000000..74d0d338b7 --- /dev/null +++ b/examples/src/main/scala/spark/streaming/examples/ZeroMQWordCount.scala @@ -0,0 +1,74 @@ +package spark.streaming.examples + +import akka.actor.ActorSystem +import akka.actor.actorRef2Scala +import akka.zeromq._ +import spark.streaming.{ Seconds, StreamingContext } +import spark.streaming.StreamingContext._ +import akka.zeromq.Subscribe + +/** + * A simple publisher for demonstration purposes, repeatedly publishes random Messages + * every one second. + */ +object SimpleZeroMQPublisher { + + def main(args: Array[String]) = { + if (args.length < 2) { + System.err.println("Usage: SimpleZeroMQPublisher <zeroMQUrl> <topic> ") + System.exit(1) + } + + val Seq(url, topic) = args.toSeq + val acs: ActorSystem = ActorSystem() + + val pubSocket = ZeroMQExtension(acs).newSocket(SocketType.Pub, Bind(url)) + val messages: Array[String] = Array("words ", "may ", "count ") + while (true) { + Thread.sleep(1000) + pubSocket ! ZMQMessage(Frame(topic) :: messages.map(x => Frame(x.getBytes)).toList) + } + acs.awaitTermination() + } +} + +/** + * A sample wordcount with ZeroMQStream stream + * + * To work with zeroMQ, some native libraries have to be installed. + * Install zeroMQ (release 2.1) core libraries. [ZeroMQ Install guide](http://www.zeromq.org/intro:get-the-software) + * + * Usage: ZeroMQWordCount <master> <zeroMQurl> <topic> + * In local mode, <master> should be 'local[n]' with n > 1 + * <zeroMQurl> and <topic> describe where zeroMq publisher is running. + * + * To run this example locally, you may run publisher as + * `$ ./run spark.streaming.examples.SimpleZeroMQPublisher tcp://127.0.1.1:1234 foo.bar` + * and run the example as + * `$ ./run spark.streaming.examples.ZeroMQWordCount local[2] tcp://127.0.1.1:1234 foo` + */ +object ZeroMQWordCount { + def main(args: Array[String]) { + if (args.length < 3) { + System.err.println( + "Usage: ZeroMQWordCount <master> <zeroMQurl> <topic>" + + "In local mode, <master> should be 'local[n]' with n > 1") + System.exit(1) + } + val Seq(master, url, topic) = args.toSeq + + // Create the context and set the batch size + val ssc = new StreamingContext(master, "ZeroMQWordCount", Seconds(2), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) + + def bytesToStringIterator(x: Seq[Seq[Byte]]) = (x.map(x => new String(x.toArray))).iterator + + //For this stream, a zeroMQ publisher should be running. + val lines = ssc.zeroMQStream(url, Subscribe(topic), bytesToStringIterator) + val words = lines.flatMap(_.split(" ")) + val wordCounts = words.map(x => (x, 1)).reduceByKey(_ + _) + wordCounts.print() + ssc.start() + } + +} diff --git a/examples/src/main/scala/spark/streaming/examples/clickstream/PageViewStream.scala b/examples/src/main/scala/spark/streaming/examples/clickstream/PageViewStream.scala index a191321d91..e226a4a73a 100644 --- a/examples/src/main/scala/spark/streaming/examples/clickstream/PageViewStream.scala +++ b/examples/src/main/scala/spark/streaming/examples/clickstream/PageViewStream.scala @@ -24,20 +24,20 @@ object PageViewStream { val port = args(2).toInt // Create the context - val ssc = new StreamingContext("local[2]", "PageViewStream", Seconds(1)) + val ssc = new StreamingContext("local[2]", "PageViewStream", Seconds(1), + System.getenv("SPARK_HOME"), Seq(System.getenv("SPARK_EXAMPLES_JAR"))) // Create a NetworkInputDStream on target host:port and convert each line to a PageView - val pageViews = ssc.networkTextStream(host, port) - .flatMap(_.split("\n")) - .map(PageView.fromString(_)) + val pageViews = ssc.socketTextStream(host, port) + .flatMap(_.split("\n")) + .map(PageView.fromString(_)) // Return a count of views per URL seen in each batch - val pageCounts = pageViews.map(view => ((view.url, 1))).countByKey() + val pageCounts = pageViews.map(view => view.url).countByValue() // Return a sliding window of page views per URL in the last ten seconds - val slidingPageCounts = pageViews.map(view => ((view.url, 1))) - .window(Seconds(10), Seconds(2)) - .countByKey() + val slidingPageCounts = pageViews.map(view => view.url) + .countByValueAndWindow(Seconds(10), Seconds(2)) // Return the rate of error pages (a non 200 status) in each zip code over the last 30 seconds @@ -61,7 +61,7 @@ object PageViewStream { .map("Unique active users: " + _) // An external dataset we want to join to this stream - val userList = ssc.sc.parallelize( + val userList = ssc.sparkContext.parallelize( Map(1 -> "Patrick Wendell", 2->"Reynold Xin", 3->"Matei Zaharia").toSeq) metric match { @@ -2,8 +2,8 @@ <project xmlns="http://maven.apache.org/POM/4.0.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd"> <modelVersion>4.0.0</modelVersion> <groupId>org.spark-project</groupId> - <artifactId>parent</artifactId> - <version>0.7.0-SNAPSHOT</version> + <artifactId>spark-parent</artifactId> + <version>0.8.0-SNAPSHOT</version> <packaging>pom</packaging> <name>Spark Project Parent POM</name> <url>http://spark-project.org/</url> @@ -51,13 +51,17 @@ <project.reporting.outputEncoding>UTF-8</project.reporting.outputEncoding> <java.version>1.5</java.version> - <scala.version>2.9.2</scala.version> + <scala.version>2.9.3</scala.version> <mesos.version>0.9.0-incubating</mesos.version> <akka.version>2.0.3</akka.version> <spray.version>1.0-M2.1</spray.version> <spray.json.version>1.1.1</spray.json.version> <slf4j.version>1.6.1</slf4j.version> <cdh.version>4.1.2</cdh.version> + <log4j.version>1.2.17</log4j.version> + + <PermGen>64m</PermGen> + <MaxPermGen>512m</MaxPermGen> </properties> <repositories> @@ -84,9 +88,9 @@ </snapshots> </repository> <repository> - <id>typesafe-repo</id> - <name>Typesafe Repository</name> - <url>http://repo.typesafe.com/typesafe/releases/</url> + <id>akka-repo</id> + <name>Akka Repository</name> + <url>http://repo.akka.io/releases/</url> <releases> <enabled>true</enabled> </releases> @@ -186,9 +190,9 @@ <version>0.8.4</version> </dependency> <dependency> - <groupId>asm</groupId> - <artifactId>asm-all</artifactId> - <version>3.3.1</version> + <groupId>org.ow2.asm</groupId> + <artifactId>asm</artifactId> + <version>4.0</version> </dependency> <dependency> <groupId>com.google.protobuf</groupId> @@ -198,7 +202,7 @@ <dependency> <groupId>de.javakaffee</groupId> <artifactId>kryo-serializers</artifactId> - <version>0.20</version> + <version>0.22</version> </dependency> <dependency> <groupId>com.typesafe.akka</groupId> @@ -237,7 +241,7 @@ </dependency> <dependency> <groupId>cc.spray</groupId> - <artifactId>spray-json_${scala.version}</artifactId> + <artifactId>spray-json_2.9.2</artifactId> <version>${spray.json.version}</version> </dependency> <dependency> @@ -247,7 +251,7 @@ </dependency> <dependency> <groupId>com.github.scala-incubator.io</groupId> - <artifactId>scala-io-file_${scala.version}</artifactId> + <artifactId>scala-io-file_2.9.2</artifactId> <version>0.4.1</version> </dependency> <dependency> @@ -255,6 +259,17 @@ <artifactId>mesos</artifactId> <version>${mesos.version}</version> </dependency> + <dependency> + <groupId>io.netty</groupId> + <artifactId>netty-all</artifactId> + <version>4.0.0.Beta2</version> + </dependency> + <dependency> + <groupId>org.apache.derby</groupId> + <artifactId>derby</artifactId> + <version>10.4.2.0</version> + <scope>test</scope> + </dependency> <dependency> <groupId>org.scala-lang</groupId> @@ -268,9 +283,15 @@ </dependency> <dependency> + <groupId>log4j</groupId> + <artifactId>log4j</artifactId> + <version>${log4j.version}</version> + </dependency> + + <dependency> <groupId>org.scalatest</groupId> <artifactId>scalatest_${scala.version}</artifactId> - <version>1.8</version> + <version>1.9.1</version> <scope>test</scope> </dependency> <dependency> @@ -282,7 +303,7 @@ <dependency> <groupId>org.scalacheck</groupId> <artifactId>scalacheck_${scala.version}</artifactId> - <version>1.9</version> + <version>1.10.0</version> <scope>test</scope> </dependency> <dependency> @@ -374,6 +395,8 @@ <jvmArgs> <jvmArg>-Xms64m</jvmArg> <jvmArg>-Xmx1024m</jvmArg> + <jvmArg>-XX:PermSize=${PermGen}</jvmArg> + <jvmArg>-XX:MaxPermSize=${MaxPermGen}</jvmArg> </jvmArgs> <javacArgs> <javacArg>-source</javacArg> @@ -410,8 +433,9 @@ <configuration> <reportsDirectory>${project.build.directory}/surefire-reports</reportsDirectory> <junitxml>.</junitxml> - <filereports>WDF TestSuite.txt</filereports> + <filereports>${project.build.directory}/SparkTestSuite.txt</filereports> <argLine>-Xms64m -Xmx1024m</argLine> + <stderr/> </configuration> <executions> <execution> @@ -505,7 +529,6 @@ <profiles> <profile> <id>hadoop1</id> - <properties> <hadoop.major.version>1</hadoop.major.version> </properties> @@ -514,7 +537,7 @@ <dependency> <groupId>org.apache.hadoop</groupId> <artifactId>hadoop-core</artifactId> - <version>1.0.3</version> + <version>1.0.4</version> </dependency> </dependencies> </dependencyManagement> @@ -547,6 +570,73 @@ <groupId>org.apache.avro</groupId> <artifactId>avro-ipc</artifactId> <version>1.7.1.cloudera.2</version> + <exclusions> + <exclusion> + <groupId>org.jboss.netty</groupId> + <artifactId>netty</artifactId> + </exclusion> + </exclusions> + </dependency> + </dependencies> + </dependencyManagement> + </profile> + + <profile> + <id>hadoop2-yarn</id> + <properties> + <hadoop.major.version>2</hadoop.major.version> + <!-- 0.23.* is same as 2.0.* - except hardened to run production jobs --> + <!-- <yarn.version>0.23.7</yarn.version> --> + <yarn.version>2.0.2-alpha</yarn.version> + </properties> + + <repositories> + <repository> + <id>maven-root</id> + <name>Maven root repository</name> + <url>http://repo1.maven.org/maven2/</url> + <releases> + <enabled>true</enabled> + </releases> + <snapshots> + <enabled>false</enabled> + </snapshots> + </repository> + </repositories> + + <dependencyManagement> + <dependencies> + <!-- TODO: check versions, bringover from yarn branch ! --> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <version>${yarn.version}</version> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-api</artifactId> + <version>${yarn.version}</version> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> + <version>${yarn.version}</version> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-client</artifactId> + <version>${yarn.version}</version> + </dependency> + <!-- Specify Avro version because Kafka also has it as a dependency --> + <dependency> + <groupId>org.apache.avro</groupId> + <artifactId>avro</artifactId> + <version>1.7.1.cloudera.2</version> + </dependency> + <dependency> + <groupId>org.apache.avro</groupId> + <artifactId>avro-ipc</artifactId> + <version>1.7.1.cloudera.2</version> </dependency> </dependencies> </dependencyManagement> diff --git a/project/SparkBuild.scala b/project/SparkBuild.scala index af8b5ba017..484f97d992 100644 --- a/project/SparkBuild.scala +++ b/project/SparkBuild.scala @@ -1,3 +1,4 @@ + import sbt._ import sbt.Classpaths.publishTask import Keys._ @@ -9,19 +10,26 @@ import twirl.sbt.TwirlPlugin._ object SparkBuild extends Build { // Hadoop version to build against. For example, "0.20.2", "0.20.205.0", or - // "1.0.3" for Apache releases, or "0.20.2-cdh3u5" for Cloudera Hadoop. - val HADOOP_VERSION = "1.0.3" + // "1.0.4" for Apache releases, or "0.20.2-cdh3u5" for Cloudera Hadoop. + val HADOOP_VERSION = "1.0.4" val HADOOP_MAJOR_VERSION = "1" + val HADOOP_YARN = false // For Hadoop 2 versions such as "2.0.0-mr1-cdh4.1.1", set the HADOOP_MAJOR_VERSION to "2" //val HADOOP_VERSION = "2.0.0-mr1-cdh4.1.1" //val HADOOP_MAJOR_VERSION = "2" + //val HADOOP_YARN = false + + // For Hadoop 2 YARN support + //val HADOOP_VERSION = "2.0.2-alpha" + //val HADOOP_MAJOR_VERSION = "2" + //val HADOOP_YARN = true lazy val root = Project("root", file("."), settings = rootSettings) aggregate(core, repl, examples, bagel, streaming) lazy val core = Project("core", file("core"), settings = coreSettings) - lazy val repl = Project("repl", file("repl"), settings = replSettings) dependsOn (core) dependsOn (streaming) + lazy val repl = Project("repl", file("repl"), settings = replSettings) dependsOn (core) lazy val examples = Project("examples", file("examples"), settings = examplesSettings) dependsOn (core) dependsOn (streaming) @@ -35,15 +43,25 @@ object SparkBuild extends Build { def sharedSettings = Defaults.defaultSettings ++ Seq( organization := "org.spark-project", - version := "0.7.0-SNAPSHOT", - scalaVersion := "2.9.2", - scalacOptions := Seq(/*"-deprecation",*/ "-unchecked", "-optimize"), // -deprecation is too noisy due to usage of old Hadoop API, enable it once that's no longer an issue + version := "0.8.0-SNAPSHOT", + scalaVersion := "2.9.3", + scalacOptions := Seq("-unchecked", "-optimize", "-deprecation"), unmanagedJars in Compile <<= baseDirectory map { base => (base / "lib" ** "*.jar").classpath }, retrieveManaged := true, retrievePattern := "[type]s/[artifact](-[revision])(-[classifier]).[ext]", transitiveClassifiers in Scope.GlobalScope := Seq("sources"), testListeners <<= target.map(t => Seq(new eu.henkelmann.sbt.JUnitXmlTestsListener(t.getAbsolutePath))), + // Fork new JVMs for tests and set Java options for those + fork := true, + javaOptions += "-Xmx2500m", + + // Only allow one test at a time, even across projects, since they run in the same JVM + concurrentRestrictions in Global += Tags.limit(Tags.Test, 1), + + // Shared between both core and streaming. + resolvers ++= Seq("Akka Repository" at "http://repo.akka.io/releases/"), + // For Sonatype publishing resolvers ++= Seq("sonatype-snapshots" at "https://oss.sonatype.org/content/repositories/snapshots", "sonatype-staging" at "https://oss.sonatype.org/service/local/staging/deploy/maven2/"), @@ -89,13 +107,13 @@ object SparkBuild extends Build { */ libraryDependencies ++= Seq( - "org.eclipse.jetty" % "jetty-server" % "7.5.3.v20111011", - "org.scalatest" %% "scalatest" % "1.8" % "test", - "org.scalacheck" %% "scalacheck" % "1.9" % "test", - "com.novocode" % "junit-interface" % "0.8" % "test", + "io.netty" % "netty" % "3.5.3.Final", + "org.eclipse.jetty" % "jetty-server" % "7.6.8.v20121106", + "org.scalatest" %% "scalatest" % "1.9.1" % "test", + "org.scalacheck" %% "scalacheck" % "1.10.0" % "test", + "com.novocode" % "junit-interface" % "0.9" % "test", "org.easymock" % "easymock" % "3.1" % "test" ), - parallelExecution := false, /* Workaround for issue #206 (fixed after SBT 0.11.0) */ watchTransitiveSources <<= Defaults.inDependencies[Task[Seq[File]]](watchSources.task, const(std.TaskExtra.constant(Nil)), aggregate = true, includeRoot = true) apply { _.join.map(_.flatten) }, @@ -107,14 +125,17 @@ object SparkBuild extends Build { publishMavenStyle in MavenCompile := true, publishLocal in MavenCompile <<= publishTask(publishLocalConfiguration in MavenCompile, deliverLocal), publishLocalBoth <<= Seq(publishLocal in MavenCompile, publishLocal).dependOn - ) + ) ++ net.virtualvoid.sbt.graph.Plugin.graphSettings + + val slf4jVersion = "1.7.2" - val slf4jVersion = "1.6.1" + val excludeJackson = ExclusionRule(organization = "org.codehaus.jackson") + val excludeNetty = ExclusionRule(organization = "org.jboss.netty") + val excludeAsm = ExclusionRule(organization = "asm") def coreSettings = sharedSettings ++ Seq( name := "spark-core", resolvers ++= Seq( - "Typesafe Repository" at "http://repo.typesafe.com/typesafe/releases/", "JBoss Repository" at "http://repository.jboss.org/nexus/content/repositories/releases/", "Spray Repository" at "http://repo.spray.cc/", "Cloudera Repository" at "https://repository.cloudera.com/artifactory/cloudera-repos/", @@ -126,22 +147,48 @@ object SparkBuild extends Build { "log4j" % "log4j" % "1.2.16", "org.slf4j" % "slf4j-api" % slf4jVersion, "org.slf4j" % "slf4j-log4j12" % slf4jVersion, + "commons-daemon" % "commons-daemon" % "1.0.10", "com.ning" % "compress-lzf" % "0.8.4", - "org.apache.hadoop" % "hadoop-core" % HADOOP_VERSION, - "asm" % "asm-all" % "3.3.1", + "org.ow2.asm" % "asm" % "4.0", "com.google.protobuf" % "protobuf-java" % "2.4.1", - "de.javakaffee" % "kryo-serializers" % "0.20", - "com.typesafe.akka" % "akka-actor" % "2.0.3", - "com.typesafe.akka" % "akka-remote" % "2.0.3", - "com.typesafe.akka" % "akka-slf4j" % "2.0.3", + "de.javakaffee" % "kryo-serializers" % "0.22", + "com.typesafe.akka" % "akka-actor" % "2.0.3" excludeAll(excludeNetty), + "com.typesafe.akka" % "akka-remote" % "2.0.3" excludeAll(excludeNetty), + "com.typesafe.akka" % "akka-slf4j" % "2.0.3" excludeAll(excludeNetty), "it.unimi.dsi" % "fastutil" % "6.4.4", "colt" % "colt" % "1.2.0", - "cc.spray" % "spray-can" % "1.0-M2.1", - "cc.spray" % "spray-server" % "1.0-M2.1", - "cc.spray" %% "spray-json" % "1.1.1", - "org.apache.mesos" % "mesos" % "0.9.0-incubating" - ) ++ (if (HADOOP_MAJOR_VERSION == "2") Some("org.apache.hadoop" % "hadoop-client" % HADOOP_VERSION) else None).toSeq, - unmanagedSourceDirectories in Compile <+= baseDirectory{ _ / ("src/hadoop" + HADOOP_MAJOR_VERSION + "/scala") } + "cc.spray" % "spray-can" % "1.0-M2.1" excludeAll(excludeNetty), + "cc.spray" % "spray-server" % "1.0-M2.1" excludeAll(excludeNetty), + "cc.spray" % "spray-json_2.9.2" % "1.1.1" excludeAll(excludeNetty), + "org.apache.mesos" % "mesos" % "0.9.0-incubating", + "io.netty" % "netty-all" % "4.0.0.Beta2", + "org.apache.derby" % "derby" % "10.4.2.0" % "test" + ) ++ ( + if (HADOOP_MAJOR_VERSION == "2") { + if (HADOOP_YARN) { + Seq( + // Exclude rule required for all ? + "org.apache.hadoop" % "hadoop-client" % HADOOP_VERSION excludeAll(excludeJackson, excludeNetty), + "org.apache.hadoop" % "hadoop-yarn-api" % HADOOP_VERSION excludeAll(excludeJackson, excludeNetty), + "org.apache.hadoop" % "hadoop-yarn-common" % HADOOP_VERSION excludeAll(excludeJackson, excludeNetty), + "org.apache.hadoop" % "hadoop-yarn-client" % HADOOP_VERSION excludeAll(excludeJackson, excludeNetty) + ) + } else { + Seq( + "org.apache.hadoop" % "hadoop-core" % HADOOP_VERSION excludeAll(excludeJackson, excludeNetty), + "org.apache.hadoop" % "hadoop-client" % HADOOP_VERSION excludeAll(excludeJackson, excludeNetty) + ) + } + } else { + Seq("org.apache.hadoop" % "hadoop-core" % HADOOP_VERSION excludeAll(excludeJackson, excludeNetty) ) + }), + unmanagedSourceDirectories in Compile <+= baseDirectory{ _ / + ( if (HADOOP_YARN && HADOOP_MAJOR_VERSION == "2") { + "src/hadoop2-yarn/scala" + } else { + "src/hadoop" + HADOOP_MAJOR_VERSION + "/scala" + } ) + } ) ++ assemblySettings ++ extraAssemblySettings ++ Twirl.settings def rootSettings = sharedSettings ++ Seq( @@ -156,7 +203,18 @@ object SparkBuild extends Build { def examplesSettings = sharedSettings ++ Seq( name := "spark-examples", libraryDependencies ++= Seq( - "org.twitter4j" % "twitter4j-stream" % "3.0.3" + "com.twitter" % "algebird-core_2.9.2" % "0.1.11", + + "org.apache.hbase" % "hbase" % "0.94.6" excludeAll(excludeNetty, excludeAsm), + + "org.apache.cassandra" % "cassandra-all" % "1.2.5" + exclude("com.google.guava", "guava") + exclude("com.googlecode.concurrentlinkedhashmap", "concurrentlinkedhashmap-lru") + exclude("com.ning","compress-lzf") + exclude("io.netty", "netty") + exclude("jline","jline") + exclude("log4j","log4j") + exclude("org.apache.cassandra.deps", "avro") ) ) @@ -164,15 +222,21 @@ object SparkBuild extends Build { def streamingSettings = sharedSettings ++ Seq( name := "spark-streaming", + resolvers ++= Seq( + "Akka Repository" at "http://repo.akka.io/releases/" + ), libraryDependencies ++= Seq( - "org.apache.flume" % "flume-ng-sdk" % "1.2.0" % "compile", - "com.github.sgroschupf" % "zkclient" % "0.1" + "org.apache.flume" % "flume-ng-sdk" % "1.2.0" % "compile" excludeAll(excludeNetty), + "com.github.sgroschupf" % "zkclient" % "0.1" excludeAll(excludeNetty), + "org.twitter4j" % "twitter4j-stream" % "3.0.3" excludeAll(excludeNetty), + "com.typesafe.akka" % "akka-zeromq" % "2.0.3" excludeAll(excludeNetty) ) ) ++ assemblySettings ++ extraAssemblySettings def extraAssemblySettings() = Seq(test in assembly := {}) ++ Seq( mergeStrategy in assembly := { case m if m.toLowerCase.endsWith("manifest.mf") => MergeStrategy.discard + case m if m.toLowerCase.matches("meta-inf/.*\\.sf$") => MergeStrategy.discard case "reference.conf" => MergeStrategy.concat case _ => MergeStrategy.first } diff --git a/project/build.properties b/project/build.properties index d4287112c6..9b860e23c5 100644 --- a/project/build.properties +++ b/project/build.properties @@ -1 +1 @@ -sbt.version=0.11.3 +sbt.version=0.12.3 diff --git a/project/plugins.sbt b/project/plugins.sbt index 4d0e696a11..f806e66481 100644 --- a/project/plugins.sbt +++ b/project/plugins.sbt @@ -4,15 +4,17 @@ resolvers += "Typesafe Repository" at "http://repo.typesafe.com/typesafe/release resolvers += "Spray Repository" at "http://repo.spray.cc/" -addSbtPlugin("com.eed3si9n" % "sbt-assembly" % "0.8.3") +addSbtPlugin("com.eed3si9n" % "sbt-assembly" % "0.8.5") -addSbtPlugin("com.typesafe.sbteclipse" % "sbteclipse-plugin" % "2.1.0-RC1") +addSbtPlugin("com.typesafe.sbteclipse" % "sbteclipse-plugin" % "2.1.1") -addSbtPlugin("com.github.mpeltonen" % "sbt-idea" % "1.0.0") +addSbtPlugin("com.github.mpeltonen" % "sbt-idea" % "1.2.0") -addSbtPlugin("cc.spray" %% "sbt-twirl" % "0.5.2") +addSbtPlugin("io.spray" %% "sbt-twirl" % "0.6.1") // For Sonatype publishing //resolvers += Resolver.url("sbt-plugin-releases", new URL("http://scalasbt.artifactoryonline.com/scalasbt/sbt-plugin-releases/"))(Resolver.ivyStylePatterns) //addSbtPlugin("com.jsuereth" % "xsbt-gpg-plugin" % "0.6") + +addSbtPlugin("net.virtual-void" % "sbt-dependency-graph" % "0.7.3") diff --git a/python/examples/transitive_closure.py b/python/examples/transitive_closure.py index 73f7f8fbaf..7f85a1008e 100644 --- a/python/examples/transitive_closure.py +++ b/python/examples/transitive_closure.py @@ -24,7 +24,7 @@ if __name__ == "__main__": "Usage: PythonTC <master> [<slices>]" exit(-1) sc = SparkContext(sys.argv[1], "PythonTC") - slices = sys.argv[2] if len(sys.argv) > 2 else 2 + slices = int(sys.argv[2]) if len(sys.argv) > 2 else 2 tc = sc.parallelize(generateGraph(), slices).cache() # Linear transitive closure: each round grows paths by one edge, diff --git a/python/pyspark/daemon.py b/python/pyspark/daemon.py new file mode 100644 index 0000000000..78a2da1e18 --- /dev/null +++ b/python/pyspark/daemon.py @@ -0,0 +1,158 @@ +import os +import sys +import multiprocessing +from ctypes import c_bool +from errno import EINTR, ECHILD +from socket import socket, AF_INET, SOCK_STREAM, SOMAXCONN +from signal import signal, SIGHUP, SIGTERM, SIGCHLD, SIG_DFL, SIG_IGN +from pyspark.worker import main as worker_main +from pyspark.serializers import write_int + +try: + POOLSIZE = multiprocessing.cpu_count() +except NotImplementedError: + POOLSIZE = 4 + +exit_flag = multiprocessing.Value(c_bool, False) + + +def should_exit(): + global exit_flag + return exit_flag.value + + +def compute_real_exit_code(exit_code): + # SystemExit's code can be integer or string, but os._exit only accepts integers + import numbers + if isinstance(exit_code, numbers.Integral): + return exit_code + else: + return 1 + + +def worker(listen_sock): + # Redirect stdout to stderr + os.dup2(2, 1) + + # Manager sends SIGHUP to request termination of workers in the pool + def handle_sighup(*args): + assert should_exit() + signal(SIGHUP, handle_sighup) + + # Cleanup zombie children + def handle_sigchld(*args): + pid = status = None + try: + while (pid, status) != (0, 0): + pid, status = os.waitpid(0, os.WNOHANG) + except EnvironmentError as err: + if err.errno == EINTR: + # retry + handle_sigchld() + elif err.errno != ECHILD: + raise + signal(SIGCHLD, handle_sigchld) + + # Handle clients + while not should_exit(): + # Wait until a client arrives or we have to exit + sock = None + while not should_exit() and sock is None: + try: + sock, addr = listen_sock.accept() + except EnvironmentError as err: + if err.errno != EINTR: + raise + + if sock is not None: + # Fork a child to handle the client. + # The client is handled in the child so that the manager + # never receives SIGCHLD unless a worker crashes. + if os.fork() == 0: + # Leave the worker pool + signal(SIGHUP, SIG_DFL) + listen_sock.close() + # Handle the client then exit + sockfile = sock.makefile() + exit_code = 0 + try: + worker_main(sockfile, sockfile) + except SystemExit as exc: + exit_code = exc.code + finally: + sockfile.close() + sock.close() + os._exit(compute_real_exit_code(exit_code)) + else: + sock.close() + + +def launch_worker(listen_sock): + if os.fork() == 0: + try: + worker(listen_sock) + except Exception as err: + import traceback + traceback.print_exc() + os._exit(1) + else: + assert should_exit() + os._exit(0) + + +def manager(): + # Create a new process group to corral our children + os.setpgid(0, 0) + + # Create a listening socket on the AF_INET loopback interface + listen_sock = socket(AF_INET, SOCK_STREAM) + listen_sock.bind(('127.0.0.1', 0)) + listen_sock.listen(max(1024, 2 * POOLSIZE, SOMAXCONN)) + listen_host, listen_port = listen_sock.getsockname() + write_int(listen_port, sys.stdout) + + # Launch initial worker pool + for idx in range(POOLSIZE): + launch_worker(listen_sock) + listen_sock.close() + + def shutdown(): + global exit_flag + exit_flag.value = True + + # Gracefully exit on SIGTERM, don't die on SIGHUP + signal(SIGTERM, lambda signum, frame: shutdown()) + signal(SIGHUP, SIG_IGN) + + # Cleanup zombie children + def handle_sigchld(*args): + try: + pid, status = os.waitpid(0, os.WNOHANG) + if status != 0 and not should_exit(): + raise RuntimeError("worker crashed: %s, %s" % (pid, status)) + except EnvironmentError as err: + if err.errno not in (ECHILD, EINTR): + raise + signal(SIGCHLD, handle_sigchld) + + # Initialization complete + sys.stdout.close() + try: + while not should_exit(): + try: + # Spark tells us to exit by closing stdin + if os.read(0, 512) == '': + shutdown() + except EnvironmentError as err: + if err.errno != EINTR: + shutdown() + raise + finally: + signal(SIGTERM, SIG_DFL) + exit_flag.value = True + # Send SIGHUP to notify workers of shutdown + os.kill(0, SIGHUP) + + +if __name__ == '__main__': + manager() diff --git a/python/pyspark/join.py b/python/pyspark/join.py index 7036c47980..5f4294fb1b 100644 --- a/python/pyspark/join.py +++ b/python/pyspark/join.py @@ -32,13 +32,13 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. """ -def _do_python_join(rdd, other, numSplits, dispatch): +def _do_python_join(rdd, other, numPartitions, dispatch): vs = rdd.map(lambda (k, v): (k, (1, v))) ws = other.map(lambda (k, v): (k, (2, v))) - return vs.union(ws).groupByKey(numSplits).flatMapValues(dispatch) + return vs.union(ws).groupByKey(numPartitions).flatMapValues(dispatch) -def python_join(rdd, other, numSplits): +def python_join(rdd, other, numPartitions): def dispatch(seq): vbuf, wbuf = [], [] for (n, v) in seq: @@ -47,10 +47,10 @@ def python_join(rdd, other, numSplits): elif n == 2: wbuf.append(v) return [(v, w) for v in vbuf for w in wbuf] - return _do_python_join(rdd, other, numSplits, dispatch) + return _do_python_join(rdd, other, numPartitions, dispatch) -def python_right_outer_join(rdd, other, numSplits): +def python_right_outer_join(rdd, other, numPartitions): def dispatch(seq): vbuf, wbuf = [], [] for (n, v) in seq: @@ -61,10 +61,10 @@ def python_right_outer_join(rdd, other, numSplits): if not vbuf: vbuf.append(None) return [(v, w) for v in vbuf for w in wbuf] - return _do_python_join(rdd, other, numSplits, dispatch) + return _do_python_join(rdd, other, numPartitions, dispatch) -def python_left_outer_join(rdd, other, numSplits): +def python_left_outer_join(rdd, other, numPartitions): def dispatch(seq): vbuf, wbuf = [], [] for (n, v) in seq: @@ -75,10 +75,10 @@ def python_left_outer_join(rdd, other, numSplits): if not wbuf: wbuf.append(None) return [(v, w) for v in vbuf for w in wbuf] - return _do_python_join(rdd, other, numSplits, dispatch) + return _do_python_join(rdd, other, numPartitions, dispatch) -def python_cogroup(rdd, other, numSplits): +def python_cogroup(rdd, other, numPartitions): vs = rdd.map(lambda (k, v): (k, (1, v))) ws = other.map(lambda (k, v): (k, (2, v))) def dispatch(seq): @@ -89,4 +89,4 @@ def python_cogroup(rdd, other, numSplits): elif n == 2: wbuf.append(v) return (vbuf, wbuf) - return vs.union(ws).groupByKey(numSplits).mapValues(dispatch) + return vs.union(ws).groupByKey(numPartitions).mapValues(dispatch) diff --git a/python/pyspark/rdd.py b/python/pyspark/rdd.py index 6b6ab6abd9..a9fec17a9d 100644 --- a/python/pyspark/rdd.py +++ b/python/pyspark/rdd.py @@ -215,7 +215,7 @@ class RDD(object): yield pair return java_cartesian.flatMap(unpack_batches) - def groupBy(self, f, numSplits=None): + def groupBy(self, f, numPartitions=None): """ Return an RDD of grouped items. @@ -224,7 +224,7 @@ class RDD(object): >>> sorted([(x, sorted(y)) for (x, y) in result]) [(0, [2, 8]), (1, [1, 1, 3, 5])] """ - return self.map(lambda x: (f(x), x)).groupByKey(numSplits) + return self.map(lambda x: (f(x), x)).groupByKey(numPartitions) def pipe(self, command, env={}): """ @@ -274,7 +274,7 @@ class RDD(object): def reduce(self, f): """ - Reduces the elements of this RDD using the specified commutative and + Reduces the elements of this RDD using the specified commutative and associative binary operator. >>> from operator import add @@ -399,7 +399,7 @@ class RDD(object): >>> sc.parallelize(range(10)).saveAsTextFile(tempFile.name) >>> from fileinput import input >>> from glob import glob - >>> ''.join(input(glob(tempFile.name + "/part-0000*"))) + >>> ''.join(sorted(input(glob(tempFile.name + "/part-0000*")))) '0\\n1\\n2\\n3\\n4\\n5\\n6\\n7\\n8\\n9\\n' """ def func(split, iterator): @@ -422,22 +422,22 @@ class RDD(object): """ return dict(self.collect()) - def reduceByKey(self, func, numSplits=None): + def reduceByKey(self, func, numPartitions=None): """ Merge the values for each key using an associative reduce function. This will also perform the merging locally on each mapper before sending results to a reducer, similarly to a "combiner" in MapReduce. - Output will be hash-partitioned with C{numSplits} splits, or the - default parallelism level if C{numSplits} is not specified. + Output will be hash-partitioned with C{numPartitions} partitions, or + the default parallelism level if C{numPartitions} is not specified. >>> from operator import add >>> rdd = sc.parallelize([("a", 1), ("b", 1), ("a", 1)]) >>> sorted(rdd.reduceByKey(add).collect()) [('a', 2), ('b', 1)] """ - return self.combineByKey(lambda x: x, func, func, numSplits) + return self.combineByKey(lambda x: x, func, func, numPartitions) def reduceByKeyLocally(self, func): """ @@ -474,7 +474,7 @@ class RDD(object): """ return self.map(lambda x: x[0]).countByValue() - def join(self, other, numSplits=None): + def join(self, other, numPartitions=None): """ Return an RDD containing all pairs of elements with matching keys in C{self} and C{other}. @@ -489,9 +489,9 @@ class RDD(object): >>> sorted(x.join(y).collect()) [('a', (1, 2)), ('a', (1, 3))] """ - return python_join(self, other, numSplits) + return python_join(self, other, numPartitions) - def leftOuterJoin(self, other, numSplits=None): + def leftOuterJoin(self, other, numPartitions=None): """ Perform a left outer join of C{self} and C{other}. @@ -506,9 +506,9 @@ class RDD(object): >>> sorted(x.leftOuterJoin(y).collect()) [('a', (1, 2)), ('b', (4, None))] """ - return python_left_outer_join(self, other, numSplits) + return python_left_outer_join(self, other, numPartitions) - def rightOuterJoin(self, other, numSplits=None): + def rightOuterJoin(self, other, numPartitions=None): """ Perform a right outer join of C{self} and C{other}. @@ -523,10 +523,10 @@ class RDD(object): >>> sorted(y.rightOuterJoin(x).collect()) [('a', (2, 1)), ('b', (None, 4))] """ - return python_right_outer_join(self, other, numSplits) + return python_right_outer_join(self, other, numPartitions) # TODO: add option to control map-side combining - def partitionBy(self, numSplits, partitionFunc=hash): + def partitionBy(self, numPartitions, partitionFunc=hash): """ Return a copy of the RDD partitioned using the specified partitioner. @@ -535,22 +535,22 @@ class RDD(object): >>> set(sets[0]).intersection(set(sets[1])) set([]) """ - if numSplits is None: - numSplits = self.ctx.defaultParallelism + if numPartitions is None: + numPartitions = self.ctx.defaultParallelism # Transferring O(n) objects to Java is too expensive. Instead, we'll - # form the hash buckets in Python, transferring O(numSplits) objects + # form the hash buckets in Python, transferring O(numPartitions) objects # to Java. Each object is a (splitNumber, [objects]) pair. def add_shuffle_key(split, iterator): buckets = defaultdict(list) for (k, v) in iterator: - buckets[partitionFunc(k) % numSplits].append((k, v)) + buckets[partitionFunc(k) % numPartitions].append((k, v)) for (split, items) in buckets.iteritems(): yield str(split) yield dump_pickle(Batch(items)) keyed = PipelinedRDD(self, add_shuffle_key) keyed._bypass_serializer = True pairRDD = self.ctx._jvm.PairwiseRDD(keyed._jrdd.rdd()).asJavaPairRDD() - partitioner = self.ctx._jvm.PythonPartitioner(numSplits, + partitioner = self.ctx._jvm.PythonPartitioner(numPartitions, id(partitionFunc)) jrdd = pairRDD.partitionBy(partitioner).values() rdd = RDD(jrdd, self.ctx) @@ -561,7 +561,7 @@ class RDD(object): # TODO: add control over map-side aggregation def combineByKey(self, createCombiner, mergeValue, mergeCombiners, - numSplits=None): + numPartitions=None): """ Generic function to combine the elements for each key using a custom set of aggregation functions. @@ -586,8 +586,8 @@ class RDD(object): >>> sorted(x.combineByKey(str, add, add).collect()) [('a', '11'), ('b', '1')] """ - if numSplits is None: - numSplits = self.ctx.defaultParallelism + if numPartitions is None: + numPartitions = self.ctx.defaultParallelism def combineLocally(iterator): combiners = {} for (k, v) in iterator: @@ -597,7 +597,7 @@ class RDD(object): combiners[k] = mergeValue(combiners[k], v) return combiners.iteritems() locally_combined = self.mapPartitions(combineLocally) - shuffled = locally_combined.partitionBy(numSplits) + shuffled = locally_combined.partitionBy(numPartitions) def _mergeCombiners(iterator): combiners = {} for (k, v) in iterator: @@ -609,10 +609,10 @@ class RDD(object): return shuffled.mapPartitions(_mergeCombiners) # TODO: support variant with custom partitioner - def groupByKey(self, numSplits=None): + def groupByKey(self, numPartitions=None): """ Group the values for each key in the RDD into a single sequence. - Hash-partitions the resulting RDD with into numSplits partitions. + Hash-partitions the resulting RDD with into numPartitions partitions. >>> x = sc.parallelize([("a", 1), ("b", 1), ("a", 1)]) >>> sorted(x.groupByKey().collect()) @@ -630,7 +630,7 @@ class RDD(object): return a + b return self.combineByKey(createCombiner, mergeValue, mergeCombiners, - numSplits) + numPartitions) # TODO: add tests def flatMapValues(self, f): @@ -659,7 +659,7 @@ class RDD(object): return self.cogroup(other) # TODO: add variant with custom parittioner - def cogroup(self, other, numSplits=None): + def cogroup(self, other, numPartitions=None): """ For each key k in C{self} or C{other}, return a resulting RDD that contains a tuple with the list of values for that key in C{self} as well @@ -670,7 +670,7 @@ class RDD(object): >>> sorted(x.cogroup(y).collect()) [('a', ([1], [2])), ('b', ([4], []))] """ - return python_cogroup(self, other, numSplits) + return python_cogroup(self, other, numPartitions) # TODO: `lookup` is disabled because we can't make direct comparisons based # on the key; we need to compare the hash of the key to the hash of the diff --git a/python/pyspark/serializers.py b/python/pyspark/serializers.py index 115cf28cc2..5a95144983 100644 --- a/python/pyspark/serializers.py +++ b/python/pyspark/serializers.py @@ -46,6 +46,10 @@ def read_long(stream): return struct.unpack("!q", length)[0] +def write_long(value, stream): + stream.write(struct.pack("!q", value)) + + def read_int(stream): length = stream.read(4) if length == "": diff --git a/python/pyspark/tests.py b/python/pyspark/tests.py index 6a1962d267..1e34d47365 100644 --- a/python/pyspark/tests.py +++ b/python/pyspark/tests.py @@ -12,6 +12,7 @@ import unittest from pyspark.context import SparkContext from pyspark.files import SparkFiles from pyspark.java_gateway import SPARK_HOME +from pyspark.serializers import read_int class PySparkTestCase(unittest.TestCase): @@ -117,5 +118,47 @@ class TestIO(PySparkTestCase): self.sc.parallelize([1]).foreach(func) +class TestDaemon(unittest.TestCase): + def connect(self, port): + from socket import socket, AF_INET, SOCK_STREAM + sock = socket(AF_INET, SOCK_STREAM) + sock.connect(('127.0.0.1', port)) + # send a split index of -1 to shutdown the worker + sock.send("\xFF\xFF\xFF\xFF") + sock.close() + return True + + def do_termination_test(self, terminator): + from subprocess import Popen, PIPE + from errno import ECONNREFUSED + + # start daemon + daemon_path = os.path.join(os.path.dirname(__file__), "daemon.py") + daemon = Popen([sys.executable, daemon_path], stdin=PIPE, stdout=PIPE) + + # read the port number + port = read_int(daemon.stdout) + + # daemon should accept connections + self.assertTrue(self.connect(port)) + + # request shutdown + terminator(daemon) + time.sleep(1) + + # daemon should no longer accept connections + with self.assertRaises(EnvironmentError) as trap: + self.connect(port) + self.assertEqual(trap.exception.errno, ECONNREFUSED) + + def test_termination_stdin(self): + """Ensure that daemon and workers terminate when stdin is closed.""" + self.do_termination_test(lambda daemon: daemon.stdin.close()) + + def test_termination_sigterm(self): + """Ensure that daemon and workers terminate on SIGTERM.""" + from signal import SIGTERM + self.do_termination_test(lambda daemon: os.kill(daemon.pid, SIGTERM)) + if __name__ == "__main__": unittest.main() diff --git a/python/pyspark/worker.py b/python/pyspark/worker.py index 812e7a9da5..379bbfd4c2 100644 --- a/python/pyspark/worker.py +++ b/python/pyspark/worker.py @@ -3,6 +3,7 @@ Worker that receives input from Piped RDD. """ import os import sys +import time import traceback from base64 import standard_b64decode # CloudPickler needs to be imported so that depicklers are registered using the @@ -12,48 +13,60 @@ from pyspark.broadcast import Broadcast, _broadcastRegistry from pyspark.cloudpickle import CloudPickler from pyspark.files import SparkFiles from pyspark.serializers import write_with_length, read_with_length, write_int, \ - read_long, read_int, dump_pickle, load_pickle, read_from_pickle_file + read_long, write_long, read_int, dump_pickle, load_pickle, read_from_pickle_file -# Redirect stdout to stderr so that users must return values from functions. -old_stdout = os.fdopen(os.dup(1), 'w') -os.dup2(2, 1) +def load_obj(infile): + return load_pickle(standard_b64decode(infile.readline().strip())) -def load_obj(): - return load_pickle(standard_b64decode(sys.stdin.readline().strip())) +def report_times(outfile, boot, init, finish): + write_int(-3, outfile) + write_long(1000 * boot, outfile) + write_long(1000 * init, outfile) + write_long(1000 * finish, outfile) -def main(): - split_index = read_int(sys.stdin) - spark_files_dir = load_pickle(read_with_length(sys.stdin)) +def main(infile, outfile): + boot_time = time.time() + split_index = read_int(infile) + if split_index == -1: # for unit tests + return + spark_files_dir = load_pickle(read_with_length(infile)) SparkFiles._root_directory = spark_files_dir SparkFiles._is_running_on_worker = True sys.path.append(spark_files_dir) - num_broadcast_variables = read_int(sys.stdin) + num_broadcast_variables = read_int(infile) for _ in range(num_broadcast_variables): - bid = read_long(sys.stdin) - value = read_with_length(sys.stdin) + bid = read_long(infile) + value = read_with_length(infile) _broadcastRegistry[bid] = Broadcast(bid, load_pickle(value)) - func = load_obj() - bypassSerializer = load_obj() + func = load_obj(infile) + bypassSerializer = load_obj(infile) if bypassSerializer: dumps = lambda x: x else: dumps = dump_pickle - iterator = read_from_pickle_file(sys.stdin) + init_time = time.time() + iterator = read_from_pickle_file(infile) try: for obj in func(split_index, iterator): - write_with_length(dumps(obj), old_stdout) + write_with_length(dumps(obj), outfile) except Exception as e: - write_int(-2, old_stdout) - write_with_length(traceback.format_exc(), old_stdout) + write_int(-2, outfile) + write_with_length(traceback.format_exc(), outfile) sys.exit(-1) + finish_time = time.time() + report_times(outfile, boot_time, init_time, finish_time) # Mark the beginning of the accumulators section of the output - write_int(-1, old_stdout) + write_int(-1, outfile) for aid, accum in _accumulatorRegistry.items(): - write_with_length(dump_pickle((aid, accum._value)), old_stdout) + write_with_length(dump_pickle((aid, accum._value)), outfile) + write_int(-1, outfile) if __name__ == '__main__': - main() + # Redirect stdout to stderr so that users must return values from functions. + old_stdout = os.fdopen(os.dup(1), 'w') + os.dup2(2, 1) + main(sys.stdin, old_stdout) diff --git a/repl-bin/pom.xml b/repl-bin/pom.xml index 0667b71cc7..7a7280313e 100644 --- a/repl-bin/pom.xml +++ b/repl-bin/pom.xml @@ -3,8 +3,8 @@ <modelVersion>4.0.0</modelVersion> <parent> <groupId>org.spark-project</groupId> - <artifactId>parent</artifactId> - <version>0.7.0-SNAPSHOT</version> + <artifactId>spark-parent</artifactId> + <version>0.8.0-SNAPSHOT</version> <relativePath>../pom.xml</relativePath> </parent> @@ -154,6 +154,61 @@ </dependencies> </profile> <profile> + <id>hadoop2-yarn</id> + <properties> + <classifier>hadoop2-yarn</classifier> + </properties> + <dependencies> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-core</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + </dependency> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-bagel</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + <scope>runtime</scope> + </dependency> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-examples</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + <scope>runtime</scope> + </dependency> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-repl</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + <scope>runtime</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <scope>runtime</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-api</artifactId> + <scope>runtime</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> + <scope>runtime</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-client</artifactId> + <scope>runtime</scope> + </dependency> + </dependencies> + </profile> + <profile> <id>deb</id> <build> <plugins> diff --git a/repl/pom.xml b/repl/pom.xml index 4a296fa630..92a2020b48 100644 --- a/repl/pom.xml +++ b/repl/pom.xml @@ -3,8 +3,8 @@ <modelVersion>4.0.0</modelVersion> <parent> <groupId>org.spark-project</groupId> - <artifactId>parent</artifactId> - <version>0.7.0-SNAPSHOT</version> + <artifactId>spark-parent</artifactId> + <version>0.8.0-SNAPSHOT</version> <relativePath>../pom.xml</relativePath> </parent> @@ -97,13 +97,6 @@ <scope>runtime</scope> </dependency> <dependency> - <groupId>org.spark-project</groupId> - <artifactId>spark-streaming</artifactId> - <version>${project.version}</version> - <classifier>hadoop1</classifier> - <scope>runtime</scope> - </dependency> - <dependency> <groupId>org.apache.hadoop</groupId> <artifactId>hadoop-core</artifactId> <scope>provided</scope> @@ -148,20 +141,84 @@ <scope>runtime</scope> </dependency> <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-core</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.avro</groupId> + <artifactId>avro</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.avro</groupId> + <artifactId>avro-ipc</artifactId> + <scope>provided</scope> + </dependency> + </dependencies> + <build> + <plugins> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-jar-plugin</artifactId> + <configuration> + <classifier>hadoop2</classifier> + </configuration> + </plugin> + </plugins> + </build> + </profile> + <profile> + <id>hadoop2-yarn</id> + <properties> + <classifier>hadoop2-yarn</classifier> + </properties> + <dependencies> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-core</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + </dependency> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-bagel</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + <scope>runtime</scope> + </dependency> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-examples</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + <scope>runtime</scope> + </dependency> + <dependency> <groupId>org.spark-project</groupId> <artifactId>spark-streaming</artifactId> <version>${project.version}</version> - <classifier>hadoop2</classifier> + <classifier>hadoop2-yarn</classifier> <scope>runtime</scope> </dependency> <dependency> <groupId>org.apache.hadoop</groupId> - <artifactId>hadoop-core</artifactId> + <artifactId>hadoop-client</artifactId> <scope>provided</scope> </dependency> <dependency> <groupId>org.apache.hadoop</groupId> - <artifactId>hadoop-client</artifactId> + <artifactId>hadoop-yarn-api</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> <scope>provided</scope> </dependency> <dependency> @@ -181,7 +238,7 @@ <groupId>org.apache.maven.plugins</groupId> <artifactId>maven-jar-plugin</artifactId> <configuration> - <classifier>hadoop2</classifier> + <classifier>hadoop2-yarn</classifier> </configuration> </plugin> </plugins> diff --git a/repl/src/main/scala/spark/repl/ExecutorClassLoader.scala b/repl/src/main/scala/spark/repl/ExecutorClassLoader.scala index 13d81ec1cf..0e9aa863b5 100644 --- a/repl/src/main/scala/spark/repl/ExecutorClassLoader.scala +++ b/repl/src/main/scala/spark/repl/ExecutorClassLoader.scala @@ -8,7 +8,6 @@ import org.apache.hadoop.conf.Configuration import org.apache.hadoop.fs.{FileSystem, Path} import org.objectweb.asm._ -import org.objectweb.asm.commons.EmptyVisitor import org.objectweb.asm.Opcodes._ @@ -83,7 +82,7 @@ extends ClassLoader(parent) { } class ConstructorCleaner(className: String, cv: ClassVisitor) -extends ClassAdapter(cv) { +extends ClassVisitor(ASM4, cv) { override def visitMethod(access: Int, name: String, desc: String, sig: String, exceptions: Array[String]): MethodVisitor = { val mv = cv.visitMethod(access, name, desc, sig, exceptions) diff --git a/repl/src/main/scala/spark/repl/SparkILoop.scala b/repl/src/main/scala/spark/repl/SparkILoop.scala index 22bcb4be8a..59f9d05683 100644 --- a/repl/src/main/scala/spark/repl/SparkILoop.scala +++ b/repl/src/main/scala/spark/repl/SparkILoop.scala @@ -200,7 +200,7 @@ class SparkILoop(in0: Option[BufferedReader], val out: PrintWriter, val master: ____ __ / __/__ ___ _____/ /__ _\ \/ _ \/ _ `/ __/ '_/ - /___/ .__/\_,_/_/ /_/\_\ version 0.7.0 + /___/ .__/\_,_/_/ /_/\_\ version 0.8.0 /_/ """) import Properties._ @@ -822,7 +822,7 @@ class SparkILoop(in0: Option[BufferedReader], val out: PrintWriter, val master: spark.repl.Main.interp.out.println("Spark context available as sc."); spark.repl.Main.interp.out.flush(); """) - command("import spark.SparkContext._"); + command("import spark.SparkContext._") } echo("Type in expressions to have them evaluated.") echo("Type :help for more information.") @@ -838,7 +838,10 @@ class SparkILoop(in0: Option[BufferedReader], val out: PrintWriter, val master: if (prop != null) prop else "local" } } - sparkContext = new SparkContext(master, "Spark shell") + val jars = Option(System.getenv("ADD_JARS")).map(_.split(',')) + .getOrElse(new Array[String](0)) + .map(new java.io.File(_).getAbsolutePath) + sparkContext = new SparkContext(master, "Spark shell", System.getenv("SPARK_HOME"), jars) sparkContext } @@ -850,6 +853,10 @@ class SparkILoop(in0: Option[BufferedReader], val out: PrintWriter, val master: printWelcome() echo("Initializing interpreter...") + // Add JARS specified in Spark's ADD_JARS variable to classpath + val jars = Option(System.getenv("ADD_JARS")).map(_.split(',')).getOrElse(new Array[String](0)) + jars.foreach(settings.classpath.append(_)) + this.settings = settings createInterpreter() diff --git a/repl/src/test/scala/spark/repl/ReplSuite.scala b/repl/src/test/scala/spark/repl/ReplSuite.scala index 43559b96d3..1c64f9b98d 100644 --- a/repl/src/test/scala/spark/repl/ReplSuite.scala +++ b/repl/src/test/scala/spark/repl/ReplSuite.scala @@ -32,6 +32,7 @@ class ReplSuite extends FunSuite { interp.sparkContext.stop() // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") return out.toString } @@ -1,6 +1,6 @@ #!/bin/bash -SCALA_VERSION=2.9.2 +SCALA_VERSION=2.9.3 # Figure out where the Scala framework is installed FWDIR="$(cd `dirname $0`; pwd)" @@ -22,21 +22,43 @@ fi # values for that; it doesn't need a lot if [ "$1" = "spark.deploy.master.Master" -o "$1" = "spark.deploy.worker.Worker" ]; then SPARK_MEM=${SPARK_DAEMON_MEMORY:-512m} + SPARK_DAEMON_JAVA_OPTS="$SPARK_DAEMON_JAVA_OPTS -Dspark.akka.logLifecycleEvents=true" # Do not overwrite SPARK_JAVA_OPTS environment variable in this script - OUR_JAVA_OPTS=$SPARK_DAEMON_JAVA_OPTS # Empty by default + OUR_JAVA_OPTS="$SPARK_DAEMON_JAVA_OPTS" # Empty by default else - OUR_JAVA_OPTS=$SPARK_JAVA_OPTS + OUR_JAVA_OPTS="$SPARK_JAVA_OPTS" fi + +# Add java opts for master, worker, executor. The opts maybe null +case "$1" in + 'spark.deploy.master.Master') + OUR_JAVA_OPTS="$OUR_JAVA_OPTS $SPARK_MASTER_OPTS" + ;; + 'spark.deploy.worker.Worker') + OUR_JAVA_OPTS="$OUR_JAVA_OPTS $SPARK_WORKER_OPTS" + ;; + 'spark.executor.StandaloneExecutorBackend') + OUR_JAVA_OPTS="$OUR_JAVA_OPTS $SPARK_EXECUTOR_OPTS" + ;; + 'spark.executor.MesosExecutorBackend') + OUR_JAVA_OPTS="$OUR_JAVA_OPTS $SPARK_EXECUTOR_OPTS" + ;; + 'spark.repl.Main') + OUR_JAVA_OPTS="$OUR_JAVA_OPTS $SPARK_REPL_OPTS" + ;; +esac + if [ "$SPARK_LAUNCH_WITH_SCALA" == "1" ]; then - if [ `command -v scala` ]; then - RUNNER="scala" + if [ "$SCALA_HOME" ]; then + RUNNER="${SCALA_HOME}/bin/scala" else - if [ -z "$SCALA_HOME" ]; then - echo "SCALA_HOME is not set" >&2 + if [ `command -v scala` ]; then + RUNNER="scala" + else + echo "SCALA_HOME is not set and scala is not in PATH" >&2 exit 1 fi - RUNNER="${SCALA_HOME}/bin/scala" fi else if [ `command -v java` ]; then @@ -66,55 +88,89 @@ export SPARK_MEM # Set JAVA_OPTS to be able to load native libraries and to set heap size JAVA_OPTS="$OUR_JAVA_OPTS" -JAVA_OPTS+=" -Djava.library.path=$SPARK_LIBRARY_PATH" -JAVA_OPTS+=" -Xms$SPARK_MEM -Xmx$SPARK_MEM" +JAVA_OPTS="$JAVA_OPTS -Djava.library.path=$SPARK_LIBRARY_PATH" +JAVA_OPTS="$JAVA_OPTS -Xms$SPARK_MEM -Xmx$SPARK_MEM" # Load extra JAVA_OPTS from conf/java-opts, if it exists if [ -e $FWDIR/conf/java-opts ] ; then - JAVA_OPTS+=" `cat $FWDIR/conf/java-opts`" + JAVA_OPTS="$JAVA_OPTS `cat $FWDIR/conf/java-opts`" fi export JAVA_OPTS # Attention: when changing the way the JAVA_OPTS are assembled, the change must be reflected in ExecutorRunner.scala! CORE_DIR="$FWDIR/core" REPL_DIR="$FWDIR/repl" +REPL_BIN_DIR="$FWDIR/repl-bin" EXAMPLES_DIR="$FWDIR/examples" BAGEL_DIR="$FWDIR/bagel" STREAMING_DIR="$FWDIR/streaming" PYSPARK_DIR="$FWDIR/python" # Exit if the user hasn't compiled Spark -if [ ! -e "$REPL_DIR/target" ]; then - echo "Failed to find Spark classes in $REPL_DIR/target" >&2 +if [ ! -e "$CORE_DIR/target" ]; then + echo "Failed to find Spark classes in $CORE_DIR/target" >&2 echo "You need to compile Spark before running this program" >&2 exit 1 fi +if [[ "$@" = *repl* && ! -e "$REPL_DIR/target" ]]; then + echo "Failed to find Spark classes in $REPL_DIR/target" >&2 + echo "You need to compile Spark repl module before running this program" >&2 + exit 1 +fi + # Build up classpath CLASSPATH="$SPARK_CLASSPATH" -CLASSPATH+=":$FWDIR/conf" -CLASSPATH+=":$CORE_DIR/target/scala-$SCALA_VERSION/classes" +CLASSPATH="$CLASSPATH:$FWDIR/conf" +CLASSPATH="$CLASSPATH:$CORE_DIR/target/scala-$SCALA_VERSION/classes" if [ -n "$SPARK_TESTING" ] ; then - CLASSPATH+=":$CORE_DIR/target/scala-$SCALA_VERSION/test-classes" + CLASSPATH="$CLASSPATH:$CORE_DIR/target/scala-$SCALA_VERSION/test-classes" + CLASSPATH="$CLASSPATH:$STREAMING_DIR/target/scala-$SCALA_VERSION/test-classes" fi -CLASSPATH+=":$CORE_DIR/src/main/resources" -CLASSPATH+=":$REPL_DIR/target/scala-$SCALA_VERSION/classes" -CLASSPATH+=":$EXAMPLES_DIR/target/scala-$SCALA_VERSION/classes" -CLASSPATH+=":$STREAMING_DIR/target/scala-$SCALA_VERSION/classes" +CLASSPATH="$CLASSPATH:$CORE_DIR/src/main/resources" +CLASSPATH="$CLASSPATH:$REPL_DIR/target/scala-$SCALA_VERSION/classes" +CLASSPATH="$CLASSPATH:$EXAMPLES_DIR/target/scala-$SCALA_VERSION/classes" +CLASSPATH="$CLASSPATH:$STREAMING_DIR/target/scala-$SCALA_VERSION/classes" +CLASSPATH="$CLASSPATH:$STREAMING_DIR/lib/org/apache/kafka/kafka/0.7.2-spark/*" # <-- our in-project Kafka Jar if [ -e "$FWDIR/lib_managed" ]; then - CLASSPATH+=":$FWDIR/lib_managed/jars/*" - CLASSPATH+=":$FWDIR/lib_managed/bundles/*" + CLASSPATH="$CLASSPATH:$FWDIR/lib_managed/jars/*" + CLASSPATH="$CLASSPATH:$FWDIR/lib_managed/bundles/*" fi -CLASSPATH+=":$REPL_DIR/lib/*" -if [ -e repl-bin/target ]; then - for jar in `find "repl-bin/target" -name 'spark-repl-*-shaded-hadoop*.jar'`; do - CLASSPATH+=":$jar" +CLASSPATH="$CLASSPATH:$REPL_DIR/lib/*" +# Add the shaded JAR for Maven builds +if [ -e $REPL_BIN_DIR/target ]; then + for jar in `find "$REPL_BIN_DIR/target" -name 'spark-repl-*-shaded-hadoop*.jar'`; do + CLASSPATH="$CLASSPATH:$jar" done + # The shaded JAR doesn't contain examples, so include those separately + EXAMPLES_JAR=`ls "$EXAMPLES_DIR/target/spark-examples"*[0-9T].jar` + CLASSPATH+=":$EXAMPLES_JAR" fi -CLASSPATH+=":$BAGEL_DIR/target/scala-$SCALA_VERSION/classes" +CLASSPATH="$CLASSPATH:$BAGEL_DIR/target/scala-$SCALA_VERSION/classes" for jar in `find $PYSPARK_DIR/lib -name '*jar'`; do - CLASSPATH+=":$jar" + CLASSPATH="$CLASSPATH:$jar" done -export CLASSPATH # Needed for spark-shell + +# Figure out the JAR file that our examples were packaged into. This includes a bit of a hack +# to avoid the -sources and -doc packages that are built by publish-local. +if [ -e "$EXAMPLES_DIR/target/scala-$SCALA_VERSION/spark-examples"*[0-9T].jar ]; then + # Use the JAR from the SBT build + export SPARK_EXAMPLES_JAR=`ls "$EXAMPLES_DIR/target/scala-$SCALA_VERSION/spark-examples"*[0-9T].jar` +fi +if [ -e "$EXAMPLES_DIR/target/spark-examples"*[0-9T].jar ]; then + # Use the JAR from the Maven build + export SPARK_EXAMPLES_JAR=`ls "$EXAMPLES_DIR/target/spark-examples"*[0-9T].jar` +fi + +# Add hadoop conf dir - else FileSystem.*, etc fail ! +# Note, this assumes that there is either a HADOOP_CONF_DIR or YARN_CONF_DIR which hosts +# the configurtion files. +if [ "x" != "x$HADOOP_CONF_DIR" ]; then + CLASSPATH="$CLASSPATH:$HADOOP_CONF_DIR" +fi +if [ "x" != "x$YARN_CONF_DIR" ]; then + CLASSPATH="$CLASSPATH:$YARN_CONF_DIR" +fi + # Figure out whether to run our class with java or with the scala launcher. # In most cases, we'd prefer to execute our process with java because scala @@ -125,11 +181,12 @@ export CLASSPATH # Needed for spark-shell if [ "$SPARK_LAUNCH_WITH_SCALA" == "1" ]; then EXTRA_ARGS="" # Java options will be passed to scala as JAVA_OPTS else - CLASSPATH+=":$SCALA_LIBRARY_PATH/scala-library.jar" - CLASSPATH+=":$SCALA_LIBRARY_PATH/scala-compiler.jar" - CLASSPATH+=":$SCALA_LIBRARY_PATH/jline.jar" + CLASSPATH="$CLASSPATH:$SCALA_LIBRARY_PATH/scala-library.jar" + CLASSPATH="$CLASSPATH:$SCALA_LIBRARY_PATH/scala-compiler.jar" + CLASSPATH="$CLASSPATH:$SCALA_LIBRARY_PATH/jline.jar" # The JVM doesn't read JAVA_OPTS by default so we need to pass it in EXTRA_ARGS="$JAVA_OPTS" fi +export CLASSPATH # Needed for spark-shell exec "$RUNNER" -cp "$CLASSPATH" $EXTRA_ARGS "$@" @@ -1,6 +1,6 @@ @echo off -set SCALA_VERSION=2.9.2 +set SCALA_VERSION=2.9.3 rem Figure out where the Spark framework is installed set FWDIR=%~dp0 @@ -21,6 +21,7 @@ set RUNNING_DAEMON=0 if "%1"=="spark.deploy.master.Master" set RUNNING_DAEMON=1 if "%1"=="spark.deploy.worker.Worker" set RUNNING_DAEMON=1 if "x%SPARK_DAEMON_MEMORY%" == "x" set SPARK_DAEMON_MEMORY=512m +set SPARK_DAEMON_JAVA_OPTS=%SPARK_DAEMON_JAVA_OPTS% -Dspark.akka.logLifecycleEvents=true if "%RUNNING_DAEMON%"=="1" set SPARK_MEM=%SPARK_DAEMON_MEMORY% rem Do not overwrite SPARK_JAVA_OPTS environment variable in this script if "%RUNNING_DAEMON%"=="0" set OUR_JAVA_OPTS=%SPARK_JAVA_OPTS% @@ -50,11 +51,14 @@ set CORE_DIR=%FWDIR%core set REPL_DIR=%FWDIR%repl set EXAMPLES_DIR=%FWDIR%examples set BAGEL_DIR=%FWDIR%bagel +set STREAMING_DIR=%FWDIR%streaming set PYSPARK_DIR=%FWDIR%python rem Build up classpath set CLASSPATH=%SPARK_CLASSPATH%;%MESOS_CLASSPATH%;%FWDIR%conf;%CORE_DIR%\target\scala-%SCALA_VERSION%\classes set CLASSPATH=%CLASSPATH%;%CORE_DIR%\target\scala-%SCALA_VERSION%\test-classes;%CORE_DIR%\src\main\resources +set CLASSPATH=%CLASSPATH%;%STREAMING_DIR%\target\scala-%SCALA_VERSION%\classes;%STREAMING_DIR%\target\scala-%SCALA_VERSION%\test-classes +set CLASSPATH=%CLASSPATH%;%STREAMING_DIR%\lib\org\apache\kafka\kafka\0.7.2-spark\* set CLASSPATH=%CLASSPATH%;%REPL_DIR%\target\scala-%SCALA_VERSION%\classes;%EXAMPLES_DIR%\target\scala-%SCALA_VERSION%\classes set CLASSPATH=%CLASSPATH%;%FWDIR%lib_managed\jars\* set CLASSPATH=%CLASSPATH%;%FWDIR%lib_managed\bundles\* @@ -62,6 +66,29 @@ set CLASSPATH=%CLASSPATH%;%FWDIR%repl\lib\* set CLASSPATH=%CLASSPATH%;%FWDIR%python\lib\* set CLASSPATH=%CLASSPATH%;%BAGEL_DIR%\target\scala-%SCALA_VERSION%\classes +rem Add hadoop conf dir - else FileSystem.*, etc fail +rem Note, this assumes that there is either a HADOOP_CONF_DIR or YARN_CONF_DIR which hosts +rem the configurtion files. +if "x%HADOOP_CONF_DIR%"=="x" goto no_hadoop_conf_dir + set CLASSPATH=%CLASSPATH%;%HADOOP_CONF_DIR% +:no_hadoop_conf_dir + +if "x%YARN_CONF_DIR%"=="x" goto no_yarn_conf_dir + set CLASSPATH=%CLASSPATH%;%YARN_CONF_DIR% +:no_yarn_conf_dir + + + +rem Figure out the JAR file that our examples were packaged into. +rem First search in the build path from SBT: +for %%d in ("examples/target/scala-%SCALA_VERSION%/spark-examples*.jar") do ( + set SPARK_EXAMPLES_JAR=examples/target/scala-%SCALA_VERSION%/%%d +) +rem Then search in the build path from Maven: +for %%d in ("examples/target/spark-examples*hadoop*.jar") do ( + set SPARK_EXAMPLES_JAR=examples/target/%%d +) + rem Figure out whether to run our class with java or with the scala launcher. rem In most cases, we'd prefer to execute our process with java because scala rem creates a shell script as the parent of its Java process, which makes it @@ -5,4 +5,4 @@ if [ "$MESOS_HOME" != "" ]; then fi export SPARK_HOME=$(cd "$(dirname $0)/.."; pwd) export SPARK_TESTING=1 # To put test classes on classpath -java -Xmx1200M -XX:MaxPermSize=250m $EXTRA_ARGS -jar $SPARK_HOME/sbt/sbt-launch-*.jar "$@" +java -Xmx1200m -XX:MaxPermSize=250m -XX:ReservedCodeCacheSize=128m $EXTRA_ARGS -jar $SPARK_HOME/sbt/sbt-launch-*.jar "$@" diff --git a/streaming/pom.xml b/streaming/pom.xml index 6ee7e59df3..4dc9a19d51 100644 --- a/streaming/pom.xml +++ b/streaming/pom.xml @@ -3,8 +3,8 @@ <modelVersion>4.0.0</modelVersion> <parent> <groupId>org.spark-project</groupId> - <artifactId>parent</artifactId> - <version>0.7.0-SNAPSHOT</version> + <artifactId>spark-parent</artifactId> + <version>0.8.0-SNAPSHOT</version> <relativePath>../pom.xml</relativePath> </parent> @@ -41,13 +41,28 @@ <groupId>org.apache.flume</groupId> <artifactId>flume-ng-sdk</artifactId> <version>1.2.0</version> + <exclusions> + <exclusion> + <groupId>org.jboss.netty</groupId> + <artifactId>netty</artifactId> + </exclusion> + </exclusions> </dependency> <dependency> <groupId>com.github.sgroschupf</groupId> <artifactId>zkclient</artifactId> <version>0.1</version> </dependency> - + <dependency> + <groupId>org.twitter4j</groupId> + <artifactId>twitter4j-stream</artifactId> + <version>3.0.3</version> + </dependency> + <dependency> + <groupId>com.typesafe.akka</groupId> + <artifactId>akka-zeromq</artifactId> + <version>2.0.3</version> + </dependency> <dependency> <groupId>org.scalatest</groupId> <artifactId>scalatest_${scala.version}</artifactId> @@ -140,5 +155,42 @@ </plugins> </build> </profile> + <profile> + <id>hadoop2-yarn</id> + <dependencies> + <dependency> + <groupId>org.spark-project</groupId> + <artifactId>spark-core</artifactId> + <version>${project.version}</version> + <classifier>hadoop2-yarn</classifier> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-api</artifactId> + <scope>provided</scope> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> + <scope>provided</scope> + </dependency> + </dependencies> + <build> + <plugins> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-jar-plugin</artifactId> + <configuration> + <classifier>hadoop2-yarn</classifier> + </configuration> + </plugin> + </plugins> + </build> + </profile> </profiles> </project> diff --git a/streaming/src/main/scala/spark/streaming/Checkpoint.scala b/streaming/src/main/scala/spark/streaming/Checkpoint.scala index 80244520a3..66e67cbfa1 100644 --- a/streaming/src/main/scala/spark/streaming/Checkpoint.scala +++ b/streaming/src/main/scala/spark/streaming/Checkpoint.scala @@ -6,6 +6,8 @@ import org.apache.hadoop.fs.{FileUtil, Path} import org.apache.hadoop.conf.Configuration import java.io._ +import com.ning.compress.lzf.{LZFInputStream, LZFOutputStream} +import java.util.concurrent.Executors private[streaming] @@ -15,9 +17,11 @@ class Checkpoint(@transient ssc: StreamingContext, val checkpointTime: Time) val framework = ssc.sc.appName val sparkHome = ssc.sc.sparkHome val jars = ssc.sc.jars + val environment = ssc.sc.environment val graph = ssc.graph val checkpointDir = ssc.checkpointDir - val checkpointDuration: Duration = ssc.checkpointDuration + val checkpointDuration = ssc.checkpointDuration + val pendingTimes = ssc.scheduler.jobManager.getPendingTimes() def validate() { assert(master != null, "Checkpoint.master is null") @@ -34,35 +38,72 @@ class Checkpoint(@transient ssc: StreamingContext, val checkpointTime: Time) private[streaming] class CheckpointWriter(checkpointDir: String) extends Logging { val file = new Path(checkpointDir, "graph") + // The file to which we actually write - and then "move" to file. + private val writeFile = new Path(file.getParent, file.getName + ".next") + private val bakFile = new Path(file.getParent, file.getName + ".bk") + + private var stopped = false + val conf = new Configuration() var fs = file.getFileSystem(conf) val maxAttempts = 3 + val executor = Executors.newFixedThreadPool(1) - def write(checkpoint: Checkpoint) { - // TODO: maybe do this in a different thread from the main stream execution thread - var attempts = 0 - while (attempts < maxAttempts) { - attempts += 1 - try { - logDebug("Saving checkpoint for time " + checkpoint.checkpointTime + " to file '" + file + "'") - if (fs.exists(file)) { - val bkFile = new Path(file.getParent, file.getName + ".bk") - FileUtil.copy(fs, file, fs, bkFile, true, true, conf) - logDebug("Moved existing checkpoint file to " + bkFile) + // Removed code which validates whether there is only one CheckpointWriter per path 'file' since + // I did not notice any errors - reintroduce it ? + + class CheckpointWriteHandler(checkpointTime: Time, bytes: Array[Byte]) extends Runnable { + def run() { + var attempts = 0 + val startTime = System.currentTimeMillis() + while (attempts < maxAttempts) { + attempts += 1 + try { + logDebug("Saving checkpoint for time " + checkpointTime + " to file '" + file + "'") + // This is inherently thread unsafe .. so alleviating it by writing to '.new' and then doing moves : which should be pretty fast. + val fos = fs.create(writeFile) + fos.write(bytes) + fos.close() + if (fs.exists(file) && fs.rename(file, bakFile)) { + logDebug("Moved existing checkpoint file to " + bakFile) + } + // paranoia + fs.delete(file, false) + fs.rename(writeFile, file) + + val finishTime = System.currentTimeMillis(); + logInfo("Checkpoint for time " + checkpointTime + " saved to file '" + file + + "', took " + bytes.length + " bytes and " + (finishTime - startTime) + " milliseconds") + return + } catch { + case ioe: IOException => + logWarning("Error writing checkpoint to file in " + attempts + " attempts", ioe) } - val fos = fs.create(file) - val oos = new ObjectOutputStream(fos) - oos.writeObject(checkpoint) - oos.close() - logInfo("Checkpoint for time " + checkpoint.checkpointTime + " saved to file '" + file + "'") - fos.close() - return - } catch { - case ioe: IOException => - logWarning("Error writing checkpoint to file in " + attempts + " attempts", ioe) } + logError("Could not write checkpoint for time " + checkpointTime + " to file '" + file + "'") + } + } + + def write(checkpoint: Checkpoint) { + val bos = new ByteArrayOutputStream() + val zos = new LZFOutputStream(bos) + val oos = new ObjectOutputStream(zos) + oos.writeObject(checkpoint) + oos.close() + bos.close() + executor.execute(new CheckpointWriteHandler(checkpoint.checkpointTime, bos.toByteArray)) + } + + def stop() { + synchronized { + if (stopped) return ; + stopped = true } - logError("Could not write checkpoint for time " + checkpoint.checkpointTime + " to file '" + file + "'") + executor.shutdown() + val startTime = System.currentTimeMillis() + val terminated = executor.awaitTermination(10, java.util.concurrent.TimeUnit.SECONDS) + val endTime = System.currentTimeMillis() + logInfo("CheckpointWriter executor terminated ? " + terminated + ", waited for " + (endTime - startTime) + " ms.") } } @@ -84,7 +125,8 @@ object CheckpointReader extends Logging { // of ObjectInputStream is used to explicitly use the current thread's default class // loader to find and load classes. This is a well know Java issue and has popped up // in other places (e.g., http://jira.codehaus.org/browse/GROOVY-1627) - val ois = new ObjectInputStreamWithLoader(fis, Thread.currentThread().getContextClassLoader) + val zis = new LZFInputStream(fis) + val ois = new ObjectInputStreamWithLoader(zis, Thread.currentThread().getContextClassLoader) val cp = ois.readObject.asInstanceOf[Checkpoint] ois.close() fs.close() diff --git a/streaming/src/main/scala/spark/streaming/DStream.scala b/streaming/src/main/scala/spark/streaming/DStream.scala index 352f83fe0c..9be7926a4a 100644 --- a/streaming/src/main/scala/spark/streaming/DStream.scala +++ b/streaming/src/main/scala/spark/streaming/DStream.scala @@ -12,7 +12,7 @@ import scala.collection.mutable.HashMap import java.io.{ObjectInputStream, IOException, ObjectOutputStream} -import org.apache.hadoop.fs.Path +import org.apache.hadoop.fs.{FileSystem, Path} import org.apache.hadoop.conf.Configuration /** @@ -75,7 +75,7 @@ abstract class DStream[T: ClassManifest] ( // Checkpoint details protected[streaming] val mustCheckpoint = false protected[streaming] var checkpointDuration: Duration = null - protected[streaming] var checkpointData = new DStreamCheckpointData(HashMap[Time, Any]()) + protected[streaming] val checkpointData = new DStreamCheckpointData(this) // Reference to whole DStream graph protected[streaming] var graph: DStreamGraph = null @@ -85,10 +85,10 @@ abstract class DStream[T: ClassManifest] ( // Duration for which the DStream requires its parent DStream to remember each RDD created protected[streaming] def parentRememberDuration = rememberDuration - /** Returns the StreamingContext associated with this DStream */ - def context() = ssc + /** Return the StreamingContext associated with this DStream */ + def context = ssc - /** Persists the RDDs of this DStream with the given storage level */ + /** Persist the RDDs of this DStream with the given storage level */ def persist(level: StorageLevel): DStream[T] = { if (this.isInitialized) { throw new UnsupportedOperationException( @@ -132,7 +132,7 @@ abstract class DStream[T: ClassManifest] ( // Set the checkpoint interval to be slideDuration or 10 seconds, which ever is larger if (mustCheckpoint && checkpointDuration == null) { - checkpointDuration = slideDuration.max(Seconds(10)) + checkpointDuration = slideDuration * math.ceil(Seconds(10) / slideDuration).toInt logInfo("Checkpoint interval automatically set to " + checkpointDuration) } @@ -159,7 +159,7 @@ abstract class DStream[T: ClassManifest] ( ) assert( - checkpointDuration == null || ssc.sc.checkpointDir.isDefined, + checkpointDuration == null || context.sparkContext.checkpointDir.isDefined, "The checkpoint directory has not been set. Please use StreamingContext.checkpoint()" + " or SparkContext.checkpoint() to set the checkpoint directory." ) @@ -238,13 +238,15 @@ abstract class DStream[T: ClassManifest] ( dependencies.foreach(_.remember(parentRememberDuration)) } - /** This method checks whether the 'time' is valid wrt slideDuration for generating RDD */ + /** Checks whether the 'time' is valid wrt slideDuration for generating RDD */ protected def isTimeValid(time: Time): Boolean = { if (!isInitialized) { throw new Exception (this + " has not been initialized") } else if (time <= zeroTime || ! (time - zeroTime).isMultipleOf(slideDuration)) { + logInfo("Time " + time + " is invalid as zeroTime is " + zeroTime + " and slideDuration is " + slideDuration + " and difference is " + (time - zeroTime)) false } else { + logInfo("Time " + time + " is valid") true } } @@ -292,14 +294,14 @@ abstract class DStream[T: ClassManifest] ( * Generate a SparkStreaming job for the given time. This is an internal method that * should not be called directly. This default implementation creates a job * that materializes the corresponding RDD. Subclasses of DStream may override this - * (eg. ForEachDStream). + * to generate their own jobs. */ protected[streaming] def generateJob(time: Time): Option[Job] = { getOrCompute(time) match { case Some(rdd) => { val jobFunc = () => { - val emptyFunc = { (iterator: Iterator[T]) => {} } - ssc.sc.runJob(rdd, emptyFunc) + val emptyFunc = { (iterator: Iterator[T]) => {} } + context.sparkContext.runJob(rdd, emptyFunc) } Some(new Job(time, jobFunc)) } @@ -308,20 +310,18 @@ abstract class DStream[T: ClassManifest] ( } /** - * Dereference RDDs that are older than rememberDuration. + * Clear metadata that are older than `rememberDuration` of this DStream. + * This is an internal method that should not be called directly. This default + * implementation clears the old generated RDDs. Subclasses of DStream may override + * this to clear their own metadata along with the generated RDDs. */ - protected[streaming] def forgetOldRDDs(time: Time) { - val keys = generatedRDDs.keys + protected[streaming] def clearOldMetadata(time: Time) { var numForgotten = 0 - keys.foreach(t => { - if (t <= (time - rememberDuration)) { - generatedRDDs.remove(t) - numForgotten += 1 - logInfo("Forgot RDD of time " + t + " from " + this) - } - }) - logInfo("Forgot " + numForgotten + " RDDs from " + this) - dependencies.foreach(_.forgetOldRDDs(time)) + val oldRDDs = generatedRDDs.filter(_._1 <= (time - rememberDuration)) + generatedRDDs --= oldRDDs.keys + logInfo("Cleared " + oldRDDs.size + " RDDs that were older than " + + (time - rememberDuration) + ": " + oldRDDs.keys.mkString(", ")) + dependencies.foreach(_.clearOldMetadata(time)) } /* Adds metadata to the Stream while it is running. @@ -342,40 +342,10 @@ abstract class DStream[T: ClassManifest] ( */ protected[streaming] def updateCheckpointData(currentTime: Time) { logInfo("Updating checkpoint data for time " + currentTime) - - // Get the checkpointed RDDs from the generated RDDs - val newRdds = generatedRDDs.filter(_._2.getCheckpointFile.isDefined) - .map(x => (x._1, x._2.getCheckpointFile.get)) - - // Make a copy of the existing checkpoint data (checkpointed RDDs) - val oldRdds = checkpointData.rdds.clone() - - // If the new checkpoint data has checkpoints then replace existing with the new one - if (newRdds.size > 0) { - checkpointData.rdds.clear() - checkpointData.rdds ++= newRdds - } - - // Make parent DStreams update their checkpoint data + checkpointData.update() dependencies.foreach(_.updateCheckpointData(currentTime)) - - // TODO: remove this, this is just for debugging - newRdds.foreach { - case (time, data) => { logInfo("Added checkpointed RDD for time " + time + " to stream checkpoint") } - } - - if (newRdds.size > 0) { - (oldRdds -- newRdds.keySet).foreach { - case (time, data) => { - val path = new Path(data.toString) - val fs = path.getFileSystem(new Configuration()) - fs.delete(path, true) - logInfo("Deleted checkpoint file '" + path + "' for time " + time) - } - } - } - logInfo("Updated checkpoint data for time " + currentTime + ", " + checkpointData.rdds.size + " checkpoints, " - + "[" + checkpointData.rdds.mkString(",") + "]") + checkpointData.cleanup() + logDebug("Updated checkpoint data for time " + currentTime + ": " + checkpointData) } /** @@ -386,14 +356,8 @@ abstract class DStream[T: ClassManifest] ( */ protected[streaming] def restoreCheckpointData() { // Create RDDs from the checkpoint data - logInfo("Restoring checkpoint data from " + checkpointData.rdds.size + " checkpointed RDDs") - checkpointData.rdds.foreach { - case(time, data) => { - logInfo("Restoring checkpointed RDD for time " + time + " from file '" + data.toString + "'") - val rdd = ssc.sc.checkpointFile[T](data.toString) - generatedRDDs += ((time, rdd)) - } - } + logInfo("Restoring checkpoint data") + checkpointData.restore() dependencies.foreach(_.restoreCheckpointData()) logInfo("Restored checkpoint data") } @@ -433,7 +397,7 @@ abstract class DStream[T: ClassManifest] ( /** Return a new DStream by applying a function to all elements of this DStream. */ def map[U: ClassManifest](mapFunc: T => U): DStream[U] = { - new MappedDStream(this, ssc.sc.clean(mapFunc)) + new MappedDStream(this, context.sparkContext.clean(mapFunc)) } /** @@ -441,7 +405,7 @@ abstract class DStream[T: ClassManifest] ( * and then flattening the results */ def flatMap[U: ClassManifest](flatMapFunc: T => Traversable[U]): DStream[U] = { - new FlatMappedDStream(this, ssc.sc.clean(flatMapFunc)) + new FlatMappedDStream(this, context.sparkContext.clean(flatMapFunc)) } /** Return a new DStream containing only the elements that satisfy a predicate. */ @@ -463,7 +427,7 @@ abstract class DStream[T: ClassManifest] ( mapPartFunc: Iterator[T] => Iterator[U], preservePartitioning: Boolean = false ): DStream[U] = { - new MapPartitionedDStream(this, ssc.sc.clean(mapPartFunc), preservePartitioning) + new MapPartitionedDStream(this, context.sparkContext.clean(mapPartFunc), preservePartitioning) } /** @@ -477,14 +441,28 @@ abstract class DStream[T: ClassManifest] ( * Return a new DStream in which each RDD has a single element generated by counting each RDD * of this DStream. */ - def count(): DStream[Long] = this.map(_ => 1L).reduce(_ + _) + def count(): DStream[Long] = { + this.map(_ => (null, 1L)) + .transform(_.union(context.sparkContext.makeRDD(Seq((null, 0L)), 1))) + .reduceByKey(_ + _) + .map(_._2) + } + + /** + * Return a new DStream in which each RDD contains the counts of each distinct value in + * each RDD of this DStream. Hash partitioning is used to generate + * the RDDs with `numPartitions` partitions (Spark's default number of partitions if + * `numPartitions` not specified). + */ + def countByValue(numPartitions: Int = ssc.sc.defaultParallelism): DStream[(T, Long)] = + this.map(x => (x, 1L)).reduceByKey((x: Long, y: Long) => x + y, numPartitions) /** * Apply a function to each RDD in this DStream. This is an output operator, so * this DStream will be registered as an output stream and therefore materialized. */ def foreach(foreachFunc: RDD[T] => Unit) { - foreach((r: RDD[T], t: Time) => foreachFunc(r)) + this.foreach((r: RDD[T], t: Time) => foreachFunc(r)) } /** @@ -492,7 +470,7 @@ abstract class DStream[T: ClassManifest] ( * this DStream will be registered as an output stream and therefore materialized. */ def foreach(foreachFunc: (RDD[T], Time) => Unit) { - val newStream = new ForEachDStream(this, ssc.sc.clean(foreachFunc)) + val newStream = new ForEachDStream(this, context.sparkContext.clean(foreachFunc)) ssc.registerOutputStream(newStream) newStream } @@ -510,7 +488,7 @@ abstract class DStream[T: ClassManifest] ( * on each RDD of this DStream. */ def transform[U: ClassManifest](transformFunc: (RDD[T], Time) => RDD[U]): DStream[U] = { - new TransformedDStream(this, ssc.sc.clean(transformFunc)) + new TransformedDStream(this, context.sparkContext.clean(transformFunc)) } /** @@ -527,19 +505,21 @@ abstract class DStream[T: ClassManifest] ( if (first11.size > 10) println("...") println() } - val newStream = new ForEachDStream(this, ssc.sc.clean(foreachFunc)) + val newStream = new ForEachDStream(this, context.sparkContext.clean(foreachFunc)) ssc.registerOutputStream(newStream) } /** - * Return a new DStream which is computed based on windowed batches of this DStream. - * The new DStream generates RDDs with the same interval as this DStream. + * Return a new DStream in which each RDD contains all the elements in seen in a + * sliding window of time over this DStream. The new DStream generates RDDs with + * the same interval as this DStream. * @param windowDuration width of the window; must be a multiple of this DStream's interval. */ def window(windowDuration: Duration): DStream[T] = window(windowDuration, this.slideDuration) /** - * Return a new DStream which is computed based on windowed batches of this DStream. + * Return a new DStream in which each RDD contains all the elements in seen in a + * sliding window of time over this DStream. * @param windowDuration width of the window; must be a multiple of this DStream's * batching interval * @param slideDuration sliding interval of the window (i.e., the interval after which @@ -551,27 +531,39 @@ abstract class DStream[T: ClassManifest] ( } /** - * Return a new DStream which computed based on tumbling window on this DStream. - * This is equivalent to window(batchTime, batchTime). - * @param batchDuration tumbling window duration; must be a multiple of this DStream's - * batching interval - */ - def tumble(batchDuration: Duration): DStream[T] = window(batchDuration, batchDuration) - - /** * Return a new DStream in which each RDD has a single element generated by reducing all - * elements in a window over this DStream. windowDuration and slideDuration are as defined - * in the window() operation. This is equivalent to - * window(windowDuration, slideDuration).reduce(reduceFunc) + * elements in a sliding window over this DStream. + * @param reduceFunc associative reduce function + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval */ def reduceByWindow( reduceFunc: (T, T) => T, windowDuration: Duration, slideDuration: Duration ): DStream[T] = { - this.window(windowDuration, slideDuration).reduce(reduceFunc) + this.reduce(reduceFunc).window(windowDuration, slideDuration).reduce(reduceFunc) } + /** + * Return a new DStream in which each RDD has a single element generated by reducing all + * elements in a sliding window over this DStream. However, the reduction is done incrementally + * using the old window's reduced value : + * 1. reduce the new values that entered the window (e.g., adding new counts) + * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) + * This is more efficient than reduceByWindow without "inverse reduce" function. + * However, it is applicable to only "invertible reduce functions". + * @param reduceFunc associative reduce function + * @param invReduceFunc inverse reduce function + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval + */ def reduceByWindow( reduceFunc: (T, T) => T, invReduceFunc: (T, T) => T, @@ -585,14 +577,47 @@ abstract class DStream[T: ClassManifest] ( /** * Return a new DStream in which each RDD has a single element generated by counting the number - * of elements in a window over this DStream. windowDuration and slideDuration are as defined in the - * window() operation. This is equivalent to window(windowDuration, slideDuration).count() + * of elements in a sliding window over this DStream. Hash partitioning is used to generate the RDDs with + * Spark's default number of partitions. + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval */ def countByWindow(windowDuration: Duration, slideDuration: Duration): DStream[Long] = { this.map(_ => 1L).reduceByWindow(_ + _, _ - _, windowDuration, slideDuration) } /** + * Return a new DStream in which each RDD contains the count of distinct elements in + * RDDs in a sliding window over this DStream. Hash partitioning is used to generate + * the RDDs with `numPartitions` partitions (Spark's default number of partitions if + * `numPartitions` not specified). + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval + * @param numPartitions number of partitions of each RDD in the new DStream. + */ + def countByValueAndWindow( + windowDuration: Duration, + slideDuration: Duration, + numPartitions: Int = ssc.sc.defaultParallelism + ): DStream[(T, Long)] = { + + this.map(x => (x, 1L)).reduceByKeyAndWindow( + (x: Long, y: Long) => x + y, + (x: Long, y: Long) => x - y, + windowDuration, + slideDuration, + numPartitions, + (x: (T, Long)) => x._2 != 0L + ) + } + + /** * Return a new DStream by unifying data of another DStream with this DStream. * @param that Another DStream having the same slideDuration as this DStream. */ @@ -609,16 +634,21 @@ abstract class DStream[T: ClassManifest] ( * Return all the RDDs between 'fromTime' to 'toTime' (both included) */ def slice(fromTime: Time, toTime: Time): Seq[RDD[T]] = { - val rdds = new ArrayBuffer[RDD[T]]() - var time = toTime.floor(slideDuration) - while (time >= zeroTime && time >= fromTime) { - getOrCompute(time) match { - case Some(rdd) => rdds += rdd - case None => //throw new Exception("Could not get RDD for time " + time) - } - time -= slideDuration + if (!(fromTime - zeroTime).isMultipleOf(slideDuration)) { + logWarning("fromTime (" + fromTime + ") is not a multiple of slideDuration (" + slideDuration + ")") + } + if (!(toTime - zeroTime).isMultipleOf(slideDuration)) { + logWarning("toTime (" + fromTime + ") is not a multiple of slideDuration (" + slideDuration + ")") } - rdds.toSeq + val alignedToTime = toTime.floor(slideDuration) + val alignedFromTime = fromTime.floor(slideDuration) + + logInfo("Slicing from " + fromTime + " to " + toTime + + " (aligned to " + alignedFromTime + " and " + alignedToTime + ")") + + alignedFromTime.to(alignedToTime, slideDuration).flatMap(time => { + if (time >= zeroTime) getOrCompute(time) else None + }) } /** @@ -651,7 +681,3 @@ abstract class DStream[T: ClassManifest] ( ssc.registerOutputStream(this) } } - -private[streaming] -case class DStreamCheckpointData(rdds: HashMap[Time, Any]) - diff --git a/streaming/src/main/scala/spark/streaming/DStreamCheckpointData.scala b/streaming/src/main/scala/spark/streaming/DStreamCheckpointData.scala new file mode 100644 index 0000000000..6b0fade7c6 --- /dev/null +++ b/streaming/src/main/scala/spark/streaming/DStreamCheckpointData.scala @@ -0,0 +1,93 @@ +package spark.streaming + +import org.apache.hadoop.fs.Path +import org.apache.hadoop.fs.FileSystem +import org.apache.hadoop.conf.Configuration +import collection.mutable.HashMap +import spark.Logging + + + +private[streaming] +class DStreamCheckpointData[T: ClassManifest] (dstream: DStream[T]) + extends Serializable with Logging { + protected val data = new HashMap[Time, AnyRef]() + + @transient private var fileSystem : FileSystem = null + @transient private var lastCheckpointFiles: HashMap[Time, String] = null + + protected[streaming] def checkpointFiles = data.asInstanceOf[HashMap[Time, String]] + + /** + * Updates the checkpoint data of the DStream. This gets called every time + * the graph checkpoint is initiated. Default implementation records the + * checkpoint files to which the generate RDDs of the DStream has been saved. + */ + def update() { + + // Get the checkpointed RDDs from the generated RDDs + val newCheckpointFiles = dstream.generatedRDDs.filter(_._2.getCheckpointFile.isDefined) + .map(x => (x._1, x._2.getCheckpointFile.get)) + + // Make a copy of the existing checkpoint data (checkpointed RDDs) + lastCheckpointFiles = checkpointFiles.clone() + + // If the new checkpoint data has checkpoints then replace existing with the new one + if (newCheckpointFiles.size > 0) { + checkpointFiles.clear() + checkpointFiles ++= newCheckpointFiles + } + + // TODO: remove this, this is just for debugging + newCheckpointFiles.foreach { + case (time, data) => { logInfo("Added checkpointed RDD for time " + time + " to stream checkpoint") } + } + } + + /** + * Cleanup old checkpoint data. This gets called every time the graph + * checkpoint is initiated, but after `update` is called. Default + * implementation, cleans up old checkpoint files. + */ + def cleanup() { + // If there is at least on checkpoint file in the current checkpoint files, + // then delete the old checkpoint files. + if (checkpointFiles.size > 0 && lastCheckpointFiles != null) { + (lastCheckpointFiles -- checkpointFiles.keySet).foreach { + case (time, file) => { + try { + val path = new Path(file) + if (fileSystem == null) { + fileSystem = path.getFileSystem(new Configuration()) + } + fileSystem.delete(path, true) + logInfo("Deleted checkpoint file '" + file + "' for time " + time) + } catch { + case e: Exception => + logWarning("Error deleting old checkpoint file '" + file + "' for time " + time, e) + } + } + } + } + } + + /** + * Restore the checkpoint data. This gets called once when the DStream graph + * (along with its DStreams) are being restored from a graph checkpoint file. + * Default implementation restores the RDDs from their checkpoint files. + */ + def restore() { + // Create RDDs from the checkpoint data + checkpointFiles.foreach { + case(time, file) => { + logInfo("Restoring checkpointed RDD for time " + time + " from file '" + file + "'") + dstream.generatedRDDs += ((time, dstream.context.sparkContext.checkpointFile[T](file))) + } + } + } + + override def toString() = { + "[\n" + checkpointFiles.size + " checkpoint files \n" + checkpointFiles.mkString("\n") + "\n]" + } +} + diff --git a/streaming/src/main/scala/spark/streaming/DStreamGraph.scala b/streaming/src/main/scala/spark/streaming/DStreamGraph.scala index bc4a40d7bc..3b331956f5 100644 --- a/streaming/src/main/scala/spark/streaming/DStreamGraph.scala +++ b/streaming/src/main/scala/spark/streaming/DStreamGraph.scala @@ -11,17 +11,20 @@ final private[streaming] class DStreamGraph extends Serializable with Logging { private val inputStreams = new ArrayBuffer[InputDStream[_]]() private val outputStreams = new ArrayBuffer[DStream[_]]() - private[streaming] var zeroTime: Time = null - private[streaming] var batchDuration: Duration = null - private[streaming] var rememberDuration: Duration = null - private[streaming] var checkpointInProgress = false + var rememberDuration: Duration = null + var checkpointInProgress = false - private[streaming] def start(time: Time) { + var zeroTime: Time = null + var startTime: Time = null + var batchDuration: Duration = null + + def start(time: Time) { this.synchronized { if (zeroTime != null) { throw new Exception("DStream graph computation already started") } zeroTime = time + startTime = time outputStreams.foreach(_.initialize(zeroTime)) outputStreams.foreach(_.remember(rememberDuration)) outputStreams.foreach(_.validate) @@ -29,81 +32,94 @@ final private[streaming] class DStreamGraph extends Serializable with Logging { } } - private[streaming] def stop() { + def restart(time: Time) { + this.synchronized { startTime = time } + } + + def stop() { this.synchronized { inputStreams.par.foreach(_.stop()) } } - private[streaming] def setContext(ssc: StreamingContext) { + def setContext(ssc: StreamingContext) { this.synchronized { outputStreams.foreach(_.setContext(ssc)) } } - private[streaming] def setBatchDuration(duration: Duration) { + def setBatchDuration(duration: Duration) { this.synchronized { if (batchDuration != null) { throw new Exception("Batch duration already set as " + batchDuration + ". cannot set it again.") } + batchDuration = duration } - batchDuration = duration } - private[streaming] def remember(duration: Duration) { + def remember(duration: Duration) { this.synchronized { if (rememberDuration != null) { throw new Exception("Batch duration already set as " + batchDuration + ". cannot set it again.") } + rememberDuration = duration } - rememberDuration = duration } - private[streaming] def addInputStream(inputStream: InputDStream[_]) { + def addInputStream(inputStream: InputDStream[_]) { this.synchronized { inputStream.setGraph(this) inputStreams += inputStream } } - private[streaming] def addOutputStream(outputStream: DStream[_]) { + def addOutputStream(outputStream: DStream[_]) { this.synchronized { outputStream.setGraph(this) outputStreams += outputStream } } - private[streaming] def getInputStreams() = this.synchronized { inputStreams.toArray } + def getInputStreams() = this.synchronized { inputStreams.toArray } - private[streaming] def getOutputStreams() = this.synchronized { outputStreams.toArray } + def getOutputStreams() = this.synchronized { outputStreams.toArray } - private[streaming] def generateRDDs(time: Time): Seq[Job] = { + def generateJobs(time: Time): Seq[Job] = { this.synchronized { - outputStreams.flatMap(outputStream => outputStream.generateJob(time)) + logInfo("Generating jobs for time " + time) + val jobs = outputStreams.flatMap(outputStream => outputStream.generateJob(time)) + logInfo("Generated " + jobs.length + " jobs for time " + time) + jobs } } - private[streaming] def forgetOldRDDs(time: Time) { + def clearOldMetadata(time: Time) { this.synchronized { - outputStreams.foreach(_.forgetOldRDDs(time)) + logInfo("Clearing old metadata for time " + time) + outputStreams.foreach(_.clearOldMetadata(time)) + logInfo("Cleared old metadata for time " + time) } } - private[streaming] def updateCheckpointData(time: Time) { + def updateCheckpointData(time: Time) { this.synchronized { + logInfo("Updating checkpoint data for time " + time) outputStreams.foreach(_.updateCheckpointData(time)) + logInfo("Updated checkpoint data for time " + time) } } - private[streaming] def restoreCheckpointData() { + def restoreCheckpointData() { this.synchronized { + logInfo("Restoring checkpoint data") outputStreams.foreach(_.restoreCheckpointData()) + logInfo("Restored checkpoint data") } } - private[streaming] def validate() { + def validate() { this.synchronized { assert(batchDuration != null, "Batch duration has not been set") //assert(batchDuration >= Milliseconds(100), "Batch duration of " + batchDuration + " is very low") diff --git a/streaming/src/main/scala/spark/streaming/Duration.scala b/streaming/src/main/scala/spark/streaming/Duration.scala index e4dc579a17..ee26206e24 100644 --- a/streaming/src/main/scala/spark/streaming/Duration.scala +++ b/streaming/src/main/scala/spark/streaming/Duration.scala @@ -16,7 +16,7 @@ case class Duration (private val millis: Long) { def * (times: Int): Duration = new Duration(millis * times) - def / (that: Duration): Long = millis / that.millis + def / (that: Duration): Double = millis.toDouble / that.millis.toDouble def isMultipleOf(that: Duration): Boolean = (this.millis % that.millis == 0) diff --git a/streaming/src/main/scala/spark/streaming/Interval.scala b/streaming/src/main/scala/spark/streaming/Interval.scala index dc21dfb722..6a8b81760e 100644 --- a/streaming/src/main/scala/spark/streaming/Interval.scala +++ b/streaming/src/main/scala/spark/streaming/Interval.scala @@ -30,6 +30,7 @@ class Interval(val beginTime: Time, val endTime: Time) { override def toString = "[" + beginTime + ", " + endTime + "]" } +private[streaming] object Interval { def currentInterval(duration: Duration): Interval = { val time = new Time(System.currentTimeMillis) diff --git a/streaming/src/main/scala/spark/streaming/JobManager.scala b/streaming/src/main/scala/spark/streaming/JobManager.scala index 3b910538e0..7696c4a592 100644 --- a/streaming/src/main/scala/spark/streaming/JobManager.scala +++ b/streaming/src/main/scala/spark/streaming/JobManager.scala @@ -3,6 +3,8 @@ package spark.streaming import spark.Logging import spark.SparkEnv import java.util.concurrent.Executors +import collection.mutable.HashMap +import collection.mutable.ArrayBuffer private[streaming] @@ -13,21 +15,57 @@ class JobManager(ssc: StreamingContext, numThreads: Int = 1) extends Logging { SparkEnv.set(ssc.env) try { val timeTaken = job.run() - logInfo("Total delay: %.5f s for job %s (execution: %.5f s)".format( - (System.currentTimeMillis() - job.time.milliseconds) / 1000.0, job.id, timeTaken / 1000.0)) + logInfo("Total delay: %.5f s for job %s of time %s (execution: %.5f s)".format( + (System.currentTimeMillis() - job.time.milliseconds) / 1000.0, job.id, job.time.milliseconds, timeTaken / 1000.0)) } catch { case e: Exception => logError("Running " + job + " failed", e) } + clearJob(job) } } initLogging() val jobExecutor = Executors.newFixedThreadPool(numThreads) - + val jobs = new HashMap[Time, ArrayBuffer[Job]] + def runJob(job: Job) { + jobs.synchronized { + jobs.getOrElseUpdate(job.time, new ArrayBuffer[Job]) += job + } jobExecutor.execute(new JobHandler(ssc, job)) logInfo("Added " + job + " to queue") } + + def stop() { + jobExecutor.shutdown() + } + + private def clearJob(job: Job) { + var timeCleared = false + val time = job.time + jobs.synchronized { + val jobsOfTime = jobs.get(time) + if (jobsOfTime.isDefined) { + jobsOfTime.get -= job + if (jobsOfTime.get.isEmpty) { + jobs -= time + timeCleared = true + } + } else { + throw new Exception("Job finished for time " + job.time + + " but time does not exist in jobs") + } + } + if (timeCleared) { + ssc.scheduler.clearOldMetadata(time) + } + } + + def getPendingTimes(): Array[Time] = { + jobs.synchronized { + jobs.keySet.toArray + } + } } diff --git a/streaming/src/main/scala/spark/streaming/NetworkInputTracker.scala b/streaming/src/main/scala/spark/streaming/NetworkInputTracker.scala index e4152f3a61..b159d26c02 100644 --- a/streaming/src/main/scala/spark/streaming/NetworkInputTracker.scala +++ b/streaming/src/main/scala/spark/streaming/NetworkInputTracker.scala @@ -4,6 +4,7 @@ import spark.streaming.dstream.{NetworkInputDStream, NetworkReceiver} import spark.streaming.dstream.{StopReceiver, ReportBlock, ReportError} import spark.Logging import spark.SparkEnv +import spark.SparkContext._ import scala.collection.mutable.HashMap import scala.collection.mutable.Queue @@ -23,7 +24,7 @@ private[streaming] case class DeregisterReceiver(streamId: Int, msg: String) ext */ private[streaming] class NetworkInputTracker( - @transient ssc: StreamingContext, + @transient ssc: StreamingContext, @transient networkInputStreams: Array[NetworkInputDStream[_]]) extends Logging { @@ -65,12 +66,12 @@ class NetworkInputTracker( def receive = { case RegisterReceiver(streamId, receiverActor) => { if (!networkInputStreamMap.contains(streamId)) { - throw new Exception("Register received for unexpected id " + streamId) + throw new Exception("Register received for unexpected id " + streamId) } receiverInfo += ((streamId, receiverActor)) logInfo("Registered receiver for network stream " + streamId + " from " + sender.path.address) sender ! true - } + } case AddBlocks(streamId, blockIds, metadata) => { val tmp = receivedBlockIds.synchronized { if (!receivedBlockIds.contains(streamId)) { @@ -85,7 +86,7 @@ class NetworkInputTracker( } case DeregisterReceiver(streamId, msg) => { receiverInfo -= streamId - logInfo("De-registered receiver for network stream " + streamId + logError("De-registered receiver for network stream " + streamId + " with message " + msg) //TODO: Do something about the corresponding NetworkInputDStream } @@ -95,8 +96,8 @@ class NetworkInputTracker( /** This thread class runs all the receivers on the cluster. */ class ReceiverExecutor extends Thread { val env = ssc.env - - override def run() { + + override def run() { try { SparkEnv.set(env) startReceivers() @@ -113,7 +114,7 @@ class NetworkInputTracker( */ def startReceivers() { val receivers = networkInputStreams.map(nis => { - val rcvr = nis.createReceiver() + val rcvr = nis.getReceiver() rcvr.setStreamId(nis.id) rcvr }) @@ -138,10 +139,14 @@ class NetworkInputTracker( } iterator.next().start() } + // Run the dummy Spark job to ensure that all slaves have registered. + // This avoids all the receivers to be scheduled on the same node. + ssc.sparkContext.makeRDD(1 to 50, 50).map(x => (x, 1)).reduceByKey(_ + _, 20).collect() + // Distribute the receivers and start them - ssc.sc.runJob(tempRDD, startReceiver) + ssc.sparkContext.runJob(tempRDD, startReceiver) } - + /** Stops the receivers. */ def stopReceivers() { // Signal the receivers to stop diff --git a/streaming/src/main/scala/spark/streaming/PairDStreamFunctions.scala b/streaming/src/main/scala/spark/streaming/PairDStreamFunctions.scala index 5db3844f1d..3ec922957d 100644 --- a/streaming/src/main/scala/spark/streaming/PairDStreamFunctions.scala +++ b/streaming/src/main/scala/spark/streaming/PairDStreamFunctions.scala @@ -18,15 +18,15 @@ import org.apache.hadoop.conf.Configuration class PairDStreamFunctions[K: ClassManifest, V: ClassManifest](self: DStream[(K,V)]) extends Serializable { - - def ssc = self.ssc + + private[streaming] def ssc = self.ssc private[streaming] def defaultPartitioner(numPartitions: Int = self.ssc.sc.defaultParallelism) = { new HashPartitioner(numPartitions) } /** - * Create a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to + * Return a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to * generate the RDDs with Spark's default number of partitions. */ def groupByKey(): DStream[(K, Seq[V])] = { @@ -34,7 +34,7 @@ extends Serializable { } /** - * Create a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to + * Return a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to * generate the RDDs with `numPartitions` partitions. */ def groupByKey(numPartitions: Int): DStream[(K, Seq[V])] = { @@ -42,7 +42,7 @@ extends Serializable { } /** - * Create a new DStream by applying `groupByKey` on each RDD. The supplied [[spark.Partitioner]] + * Return a new DStream by applying `groupByKey` on each RDD. The supplied [[spark.Partitioner]] * is used to control the partitioning of each RDD. */ def groupByKey(partitioner: Partitioner): DStream[(K, Seq[V])] = { @@ -54,7 +54,7 @@ extends Serializable { } /** - * Create a new DStream by applying `reduceByKey` to each RDD. The values for each key are + * Return a new DStream by applying `reduceByKey` to each RDD. The values for each key are * merged using the associative reduce function. Hash partitioning is used to generate the RDDs * with Spark's default number of partitions. */ @@ -63,7 +63,7 @@ extends Serializable { } /** - * Create a new DStream by applying `reduceByKey` to each RDD. The values for each key are + * Return a new DStream by applying `reduceByKey` to each RDD. The values for each key are * merged using the supplied reduce function. Hash partitioning is used to generate the RDDs * with `numPartitions` partitions. */ @@ -72,7 +72,7 @@ extends Serializable { } /** - * Create a new DStream by applying `reduceByKey` to each RDD. The values for each key are + * Return a new DStream by applying `reduceByKey` to each RDD. The values for each key are * merged using the supplied reduce function. [[spark.Partitioner]] is used to control the * partitioning of each RDD. */ @@ -82,7 +82,7 @@ extends Serializable { } /** - * Combine elements of each key in DStream's RDDs using custom function. This is similar to the + * Combine elements of each key in DStream's RDDs using custom functions. This is similar to the * combineByKey for RDDs. Please refer to combineByKey in [[spark.PairRDDFunctions]] for more * information. */ @@ -95,15 +95,7 @@ extends Serializable { } /** - * Create a new DStream by counting the number of values of each key in each RDD. Hash - * partitioning is used to generate the RDDs with Spark's `numPartitions` partitions. - */ - def countByKey(numPartitions: Int = self.ssc.sc.defaultParallelism): DStream[(K, Long)] = { - self.map(x => (x._1, 1L)).reduceByKey((x: Long, y: Long) => x + y, numPartitions) - } - - /** - * Creates a new DStream by applying `groupByKey` over a sliding window. This is similar to + * Return a new DStream by applying `groupByKey` over a sliding window. This is similar to * `DStream.groupByKey()` but applies it over a sliding window. The new DStream generates RDDs * with the same interval as this DStream. Hash partitioning is used to generate the RDDs with * Spark's default number of partitions. @@ -115,7 +107,7 @@ extends Serializable { } /** - * Create a new DStream by applying `groupByKey` over a sliding window. Similar to + * Return a new DStream by applying `groupByKey` over a sliding window. Similar to * `DStream.groupByKey()`, but applies it over a sliding window. Hash partitioning is used to * generate the RDDs with Spark's default number of partitions. * @param windowDuration width of the window; must be a multiple of this DStream's @@ -129,7 +121,7 @@ extends Serializable { } /** - * Create a new DStream by applying `groupByKey` over a sliding window on `this` DStream. + * Return a new DStream by applying `groupByKey` over a sliding window on `this` DStream. * Similar to `DStream.groupByKey()`, but applies it over a sliding window. * Hash partitioning is used to generate the RDDs with `numPartitions` partitions. * @param windowDuration width of the window; must be a multiple of this DStream's @@ -137,7 +129,8 @@ extends Serializable { * @param slideDuration sliding interval of the window (i.e., the interval after which * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval - * @param numPartitions Number of partitions of each RDD in the new DStream. + * @param numPartitions number of partitions of each RDD in the new DStream; if not specified + * then Spark's default number of partitions will be used */ def groupByKeyAndWindow( windowDuration: Duration, @@ -155,7 +148,7 @@ extends Serializable { * @param slideDuration sliding interval of the window (i.e., the interval after which * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval - * @param partitioner Partitioner for controlling the partitioning of each RDD in the new DStream. + * @param partitioner partitioner for controlling the partitioning of each RDD in the new DStream. */ def groupByKeyAndWindow( windowDuration: Duration, @@ -166,7 +159,7 @@ extends Serializable { } /** - * Create a new DStream by applying `reduceByKey` over a sliding window on `this` DStream. + * Return a new DStream by applying `reduceByKey` over a sliding window on `this` DStream. * Similar to `DStream.reduceByKey()`, but applies it over a sliding window. The new DStream * generates RDDs with the same interval as this DStream. Hash partitioning is used to generate * the RDDs with Spark's default number of partitions. @@ -182,7 +175,7 @@ extends Serializable { } /** - * Create a new DStream by applying `reduceByKey` over a sliding window. This is similar to + * Return a new DStream by applying `reduceByKey` over a sliding window. This is similar to * `DStream.reduceByKey()` but applies it over a sliding window. Hash partitioning is used to * generate the RDDs with Spark's default number of partitions. * @param reduceFunc associative reduce function @@ -201,7 +194,7 @@ extends Serializable { } /** - * Create a new DStream by applying `reduceByKey` over a sliding window. This is similar to + * Return a new DStream by applying `reduceByKey` over a sliding window. This is similar to * `DStream.reduceByKey()` but applies it over a sliding window. Hash partitioning is used to * generate the RDDs with `numPartitions` partitions. * @param reduceFunc associative reduce function @@ -210,10 +203,10 @@ extends Serializable { * @param slideDuration sliding interval of the window (i.e., the interval after which * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval - * @param numPartitions Number of partitions of each RDD in the new DStream. + * @param numPartitions number of partitions of each RDD in the new DStream. */ def reduceByKeyAndWindow( - reduceFunc: (V, V) => V, + reduceFunc: (V, V) => V, windowDuration: Duration, slideDuration: Duration, numPartitions: Int @@ -222,7 +215,7 @@ extends Serializable { } /** - * Create a new DStream by applying `reduceByKey` over a sliding window. Similar to + * Return a new DStream by applying `reduceByKey` over a sliding window. Similar to * `DStream.reduceByKey()`, but applies it over a sliding window. * @param reduceFunc associative reduce function * @param windowDuration width of the window; must be a multiple of this DStream's @@ -230,7 +223,8 @@ extends Serializable { * @param slideDuration sliding interval of the window (i.e., the interval after which * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval - * @param partitioner Partitioner for controlling the partitioning of each RDD in the new DStream. + * @param partitioner partitioner for controlling the partitioning of each RDD + * in the new DStream. */ def reduceByKeyAndWindow( reduceFunc: (V, V) => V, @@ -245,118 +239,78 @@ extends Serializable { } /** - * Create a new DStream by reducing over a using incremental computation. - * The reduced value of over a new window is calculated using the old window's reduce value : + * Return a new DStream by applying incremental `reduceByKey` over a sliding window. + * The reduced value of over a new window is calculated using the old window's reduced value : * 1. reduce the new values that entered the window (e.g., adding new counts) + * * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) - * This is more efficient that reduceByKeyAndWindow without "inverse reduce" function. + * + * This is more efficient than reduceByKeyAndWindow without "inverse reduce" function. * However, it is applicable to only "invertible reduce functions". * Hash partitioning is used to generate the RDDs with Spark's default number of partitions. * @param reduceFunc associative reduce function - * @param invReduceFunc inverse function + * @param invReduceFunc inverse reduce function * @param windowDuration width of the window; must be a multiple of this DStream's * batching interval * @param slideDuration sliding interval of the window (i.e., the interval after which * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval + * @param filterFunc Optional function to filter expired key-value pairs; + * only pairs that satisfy the function are retained */ def reduceByKeyAndWindow( reduceFunc: (V, V) => V, invReduceFunc: (V, V) => V, windowDuration: Duration, - slideDuration: Duration + slideDuration: Duration = self.slideDuration, + numPartitions: Int = ssc.sc.defaultParallelism, + filterFunc: ((K, V)) => Boolean = null ): DStream[(K, V)] = { reduceByKeyAndWindow( - reduceFunc, invReduceFunc, windowDuration, slideDuration, defaultPartitioner()) - } - - /** - * Create a new DStream by reducing over a using incremental computation. - * The reduced value of over a new window is calculated using the old window's reduce value : - * 1. reduce the new values that entered the window (e.g., adding new counts) - * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) - * This is more efficient that reduceByKeyAndWindow without "inverse reduce" function. - * However, it is applicable to only "invertible reduce functions". - * Hash partitioning is used to generate the RDDs with `numPartitions` partitions. - * @param reduceFunc associative reduce function - * @param invReduceFunc inverse function - * @param windowDuration width of the window; must be a multiple of this DStream's - * batching interval - * @param slideDuration sliding interval of the window (i.e., the interval after which - * the new DStream will generate RDDs); must be a multiple of this - * DStream's batching interval - * @param numPartitions Number of partitions of each RDD in the new DStream. - */ - def reduceByKeyAndWindow( - reduceFunc: (V, V) => V, - invReduceFunc: (V, V) => V, - windowDuration: Duration, - slideDuration: Duration, - numPartitions: Int - ): DStream[(K, V)] = { - - reduceByKeyAndWindow( - reduceFunc, invReduceFunc, windowDuration, slideDuration, defaultPartitioner(numPartitions)) + reduceFunc, invReduceFunc, windowDuration, + slideDuration, defaultPartitioner(numPartitions), filterFunc + ) } /** - * Create a new DStream by reducing over a using incremental computation. - * The reduced value of over a new window is calculated using the old window's reduce value : + * Return a new DStream by applying incremental `reduceByKey` over a sliding window. + * The reduced value of over a new window is calculated using the old window's reduced value : * 1. reduce the new values that entered the window (e.g., adding new counts) * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) - * This is more efficient that reduceByKeyAndWindow without "inverse reduce" function. + * This is more efficient than reduceByKeyAndWindow without "inverse reduce" function. * However, it is applicable to only "invertible reduce functions". - * @param reduceFunc associative reduce function - * @param invReduceFunc inverse function + * @param reduceFunc associative reduce function + * @param invReduceFunc inverse reduce function * @param windowDuration width of the window; must be a multiple of this DStream's * batching interval * @param slideDuration sliding interval of the window (i.e., the interval after which * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval - * @param partitioner Partitioner for controlling the partitioning of each RDD in the new DStream. + * @param partitioner partitioner for controlling the partitioning of each RDD in the new DStream. + * @param filterFunc Optional function to filter expired key-value pairs; + * only pairs that satisfy the function are retained */ def reduceByKeyAndWindow( reduceFunc: (V, V) => V, invReduceFunc: (V, V) => V, windowDuration: Duration, slideDuration: Duration, - partitioner: Partitioner + partitioner: Partitioner, + filterFunc: ((K, V)) => Boolean ): DStream[(K, V)] = { val cleanedReduceFunc = ssc.sc.clean(reduceFunc) val cleanedInvReduceFunc = ssc.sc.clean(invReduceFunc) + val cleanedFilterFunc = if (filterFunc != null) Some(ssc.sc.clean(filterFunc)) else None new ReducedWindowedDStream[K, V]( - self, cleanedReduceFunc, cleanedInvReduceFunc, windowDuration, slideDuration, partitioner) - } - - /** - * Create a new DStream by counting the number of values for each key over a window. - * Hash partitioning is used to generate the RDDs with `numPartitions` partitions. - * @param windowDuration width of the window; must be a multiple of this DStream's - * batching interval - * @param slideDuration sliding interval of the window (i.e., the interval after which - * the new DStream will generate RDDs); must be a multiple of this - * DStream's batching interval - * @param numPartitions Number of partitions of each RDD in the new DStream. - */ - def countByKeyAndWindow( - windowDuration: Duration, - slideDuration: Duration, - numPartitions: Int = self.ssc.sc.defaultParallelism - ): DStream[(K, Long)] = { - - self.map(x => (x._1, 1L)).reduceByKeyAndWindow( - (x: Long, y: Long) => x + y, - (x: Long, y: Long) => x - y, - windowDuration, - slideDuration, - numPartitions + self, cleanedReduceFunc, cleanedInvReduceFunc, cleanedFilterFunc, + windowDuration, slideDuration, partitioner ) } /** - * Create a new "state" DStream where the state for each key is updated by applying + * Return a new "state" DStream where the state for each key is updated by applying * the given function on the previous state of the key and the new values of each key. * Hash partitioning is used to generate the RDDs with Spark's default number of partitions. * @param updateFunc State update function. If `this` function returns None, then @@ -370,7 +324,7 @@ extends Serializable { } /** - * Create a new "state" DStream where the state for each key is updated by applying + * Return a new "state" DStream where the state for each key is updated by applying * the given function on the previous state of the key and the new values of each key. * Hash partitioning is used to generate the RDDs with `numPartitions` partitions. * @param updateFunc State update function. If `this` function returns None, then @@ -405,7 +359,7 @@ extends Serializable { } /** - * Create a new "state" DStream where the state for each key is updated by applying + * Return a new "state" DStream where the state for each key is updated by applying * the given function on the previous state of the key and the new values of each key. * [[spark.Paxrtitioner]] is used to control the partitioning of each RDD. * @param updateFunc State update function. If `this` function returns None, then @@ -447,7 +401,7 @@ extends Serializable { } /** - * Cogroup `this` DStream with `other` DStream. For each key k in corresponding RDDs of `this` + * Cogroup `this` DStream with `other` DStream using a partitioner. For each key k in corresponding RDDs of `this` * or `other` DStreams, the generated RDD will contains a tuple with the list of values for that * key in both RDDs. Partitioner is used to partition each generated RDD. */ diff --git a/streaming/src/main/scala/spark/streaming/Scheduler.scala b/streaming/src/main/scala/spark/streaming/Scheduler.scala index c04ed37de8..1c4b22a898 100644 --- a/streaming/src/main/scala/spark/streaming/Scheduler.scala +++ b/streaming/src/main/scala/spark/streaming/Scheduler.scala @@ -9,11 +9,8 @@ class Scheduler(ssc: StreamingContext) extends Logging { initLogging() - val graph = ssc.graph - val concurrentJobs = System.getProperty("spark.streaming.concurrentJobs", "1").toInt val jobManager = new JobManager(ssc, concurrentJobs) - val checkpointWriter = if (ssc.checkpointDuration != null && ssc.checkpointDir != null) { new CheckpointWriter(ssc.checkpointDir) } else { @@ -23,54 +20,93 @@ class Scheduler(ssc: StreamingContext) extends Logging { val clockClass = System.getProperty("spark.streaming.clock", "spark.streaming.util.SystemClock") val clock = Class.forName(clockClass).newInstance().asInstanceOf[Clock] val timer = new RecurringTimer(clock, ssc.graph.batchDuration.milliseconds, - longTime => generateRDDs(new Time(longTime))) + longTime => generateJobs(new Time(longTime))) + val graph = ssc.graph + var latestTime: Time = null - def start() { - // If context was started from checkpoint, then restart timer such that - // this timer's triggers occur at the same time as the original timer. - // Otherwise just start the timer from scratch, and initialize graph based - // on this first trigger time of the timer. + def start() = synchronized { if (ssc.isCheckpointPresent) { - // If manual clock is being used for testing, then - // either set the manual clock to the last checkpointed time, - // or if the property is defined set it to that time - if (clock.isInstanceOf[ManualClock]) { - val lastTime = ssc.getInitialCheckpoint.checkpointTime.milliseconds - val jumpTime = System.getProperty("spark.streaming.manualClock.jump", "0").toLong - clock.asInstanceOf[ManualClock].setTime(lastTime + jumpTime) - } - timer.restart(graph.zeroTime.milliseconds) - logInfo("Scheduler's timer restarted") + restart() } else { - val firstTime = new Time(timer.start()) - graph.start(firstTime - ssc.graph.batchDuration) - logInfo("Scheduler's timer started") + startFirstTime() } logInfo("Scheduler started") } - def stop() { + def stop() = synchronized { timer.stop() - graph.stop() + jobManager.stop() + if (checkpointWriter != null) checkpointWriter.stop() + ssc.graph.stop() logInfo("Scheduler stopped") } - - private def generateRDDs(time: Time) { + + private def startFirstTime() { + val startTime = new Time(timer.getStartTime()) + graph.start(startTime - graph.batchDuration) + timer.start(startTime.milliseconds) + logInfo("Scheduler's timer started at " + startTime) + } + + private def restart() { + + // If manual clock is being used for testing, then + // either set the manual clock to the last checkpointed time, + // or if the property is defined set it to that time + if (clock.isInstanceOf[ManualClock]) { + val lastTime = ssc.initialCheckpoint.checkpointTime.milliseconds + val jumpTime = System.getProperty("spark.streaming.manualClock.jump", "0").toLong + clock.asInstanceOf[ManualClock].setTime(lastTime + jumpTime) + } + + val batchDuration = ssc.graph.batchDuration + + // Batches when the master was down, that is, + // between the checkpoint and current restart time + val checkpointTime = ssc.initialCheckpoint.checkpointTime + val restartTime = new Time(timer.getRestartTime(graph.zeroTime.milliseconds)) + val downTimes = checkpointTime.until(restartTime, batchDuration) + logInfo("Batches during down time: " + downTimes.mkString(", ")) + + // Batches that were unprocessed before failure + val pendingTimes = ssc.initialCheckpoint.pendingTimes + logInfo("Batches pending processing: " + pendingTimes.mkString(", ")) + // Reschedule jobs for these times + val timesToReschedule = (pendingTimes ++ downTimes).distinct.sorted(Time.ordering) + logInfo("Batches to reschedule: " + timesToReschedule.mkString(", ")) + timesToReschedule.foreach(time => + graph.generateJobs(time).foreach(jobManager.runJob) + ) + + // Restart the timer + timer.start(restartTime.milliseconds) + logInfo("Scheduler's timer restarted at " + restartTime) + } + + /** Generate jobs and perform checkpoint for the given `time`. */ + def generateJobs(time: Time) { SparkEnv.set(ssc.env) logInfo("\n-----------------------------------------------------\n") - graph.generateRDDs(time).foreach(jobManager.runJob) - graph.forgetOldRDDs(time) + graph.generateJobs(time).foreach(jobManager.runJob) + latestTime = time + doCheckpoint(time) + } + + /** + * Clear old metadata assuming jobs of `time` have finished processing. + * And also perform checkpoint. + */ + def clearOldMetadata(time: Time) { + ssc.graph.clearOldMetadata(time) doCheckpoint(time) - logInfo("Generated RDDs for time " + time) } - private def doCheckpoint(time: Time) { + /** Perform checkpoint for the give `time`. */ + def doCheckpoint(time: Time) = synchronized { if (ssc.checkpointDuration != null && (time - graph.zeroTime).isMultipleOf(ssc.checkpointDuration)) { - val startTime = System.currentTimeMillis() + logInfo("Checkpointing graph for time " + time) ssc.graph.updateCheckpointData(time) checkpointWriter.write(new Checkpoint(ssc, time)) - val stopTime = System.currentTimeMillis() - logInfo("Checkpointing the graph took " + (stopTime - startTime) + " ms") } } } diff --git a/streaming/src/main/scala/spark/streaming/StreamingContext.scala b/streaming/src/main/scala/spark/streaming/StreamingContext.scala index 0cce2b13cf..f2c4073f22 100644 --- a/streaming/src/main/scala/spark/streaming/StreamingContext.scala +++ b/streaming/src/main/scala/spark/streaming/StreamingContext.scala @@ -1,22 +1,33 @@ package spark.streaming +import akka.actor.Props +import akka.actor.SupervisorStrategy +import akka.zeromq.Subscribe + import spark.streaming.dstream._ -import spark.{RDD, Logging, SparkEnv, SparkContext} +import spark._ +import spark.streaming.receivers.ActorReceiver +import spark.streaming.receivers.ReceiverSupervisorStrategy +import spark.streaming.receivers.ZeroMQReceiver import spark.storage.StorageLevel import spark.util.MetadataCleaner +import spark.streaming.receivers.ActorReceiver import scala.collection.mutable.Queue +import scala.collection.Map import java.io.InputStream import java.util.concurrent.atomic.AtomicInteger +import java.util.UUID import org.apache.hadoop.io.LongWritable import org.apache.hadoop.io.Text import org.apache.hadoop.mapreduce.{InputFormat => NewInputFormat} import org.apache.hadoop.mapreduce.lib.input.TextInputFormat import org.apache.hadoop.fs.Path -import java.util.UUID +import twitter4j.Status + /** * A StreamingContext is the main entry point for Spark Streaming functionality. Besides the basic @@ -30,23 +41,34 @@ class StreamingContext private ( ) extends Logging { /** - * Creates a StreamingContext using an existing SparkContext. + * Create a StreamingContext using an existing SparkContext. * @param sparkContext Existing SparkContext * @param batchDuration The time interval at which streaming data will be divided into batches */ - def this(sparkContext: SparkContext, batchDuration: Duration) = this(sparkContext, null, batchDuration) + def this(sparkContext: SparkContext, batchDuration: Duration) = { + this(sparkContext, null, batchDuration) + } /** - * Creates a StreamingContext by providing the details necessary for creating a new SparkContext. + * Create a StreamingContext by providing the details necessary for creating a new SparkContext. * @param master Cluster URL to connect to (e.g. mesos://host:port, spark://host:port, local[4]). * @param appName A name for your job, to display on the cluster web UI * @param batchDuration The time interval at which streaming data will be divided into batches */ - def this(master: String, appName: String, batchDuration: Duration) = - this(StreamingContext.createNewSparkContext(master, appName), null, batchDuration) + def this( + master: String, + appName: String, + batchDuration: Duration, + sparkHome: String = null, + jars: Seq[String] = Nil, + environment: Map[String, String] = Map()) = { + this(StreamingContext.createNewSparkContext(master, appName, sparkHome, jars, environment), + null, batchDuration) + } + /** - * Re-creates a StreamingContext from a checkpoint file. + * Re-create a StreamingContext from a checkpoint file. * @param path Path either to the directory that was specified as the checkpoint directory, or * to the checkpoint file 'graph' or 'graph.bk'. */ @@ -55,15 +77,20 @@ class StreamingContext private ( initLogging() if (sc_ == null && cp_ == null) { - throw new Exception("Streaming Context cannot be initilalized with " + + throw new Exception("Spark Streaming cannot be initialized with " + "both SparkContext and checkpoint as null") } + if (MetadataCleaner.getDelaySeconds < 0) { + throw new SparkException("Spark Streaming cannot be used without setting spark.cleaner.ttl; " + + "set this property before creating a SparkContext (use SPARK_JAVA_OPTS for the shell)") + } + protected[streaming] val isCheckpointPresent = (cp_ != null) - val sc: SparkContext = { + protected[streaming] val sc: SparkContext = { if (isCheckpointPresent) { - new SparkContext(cp_.master, cp_.framework, cp_.sparkHome, cp_.jars) + new SparkContext(cp_.master, cp_.framework, cp_.sparkHome, cp_.jars, cp_.environment) } else { sc_ } @@ -101,7 +128,12 @@ class StreamingContext private ( protected[streaming] var scheduler: Scheduler = null /** - * Sets each DStreams in this context to remember RDDs it generated in the last given duration. + * Return the associated Spark context + */ + def sparkContext = sc + + /** + * Set each DStreams in this context to remember RDDs it generated in the last given duration. * DStreams remember RDDs only for a limited duration of time and releases them for garbage * collection. This method allows the developer to specify how to long to remember the RDDs ( * if the developer wishes to query old data outside the DStream computation). @@ -112,71 +144,136 @@ class StreamingContext private ( } /** - * Sets the context to periodically checkpoint the DStream operations for master - * fault-tolerance. By default, the graph will be checkpointed every batch interval. + * Set the context to periodically checkpoint the DStream operations for master + * fault-tolerance. The graph will be checkpointed every batch interval. * @param directory HDFS-compatible directory where the checkpoint data will be reliably stored - * @param interval checkpoint interval */ - def checkpoint(directory: String, interval: Duration = null) { + def checkpoint(directory: String) { if (directory != null) { sc.setCheckpointDir(StreamingContext.getSparkCheckpointDir(directory)) checkpointDir = directory - checkpointDuration = interval } else { checkpointDir = null - checkpointDuration = null } } - protected[streaming] def getInitialCheckpoint(): Checkpoint = { + protected[streaming] def initialCheckpoint: Checkpoint = { if (isCheckpointPresent) cp_ else null } protected[streaming] def getNewNetworkStreamId() = nextNetworkInputStreamId.getAndIncrement() /** - * Create an input stream that pulls messages form a Kafka Broker. - * @param hostname Zookeper hostname. - * @param port Zookeper port. + * Create an input stream with any arbitrary user implemented network receiver. + * @param receiver Custom implementation of NetworkReceiver + */ + def networkStream[T: ClassManifest]( + receiver: NetworkReceiver[T]): DStream[T] = { + val inputStream = new PluggableInputDStream[T](this, + receiver) + graph.addInputStream(inputStream) + inputStream + } + + /** + * Create an input stream with any arbitrary user implemented actor receiver. + * @param props Props object defining creation of the actor + * @param name Name of the actor + * @param storageLevel RDD storage level. Defaults to memory-only. + * + * @note An important point to note: + * Since Actor may exist outside the spark framework, It is thus user's responsibility + * to ensure the type safety, i.e parametrized type of data received and actorStream + * should be same. + */ + def actorStream[T: ClassManifest]( + props: Props, + name: String, + storageLevel: StorageLevel = StorageLevel.MEMORY_ONLY_SER_2, + supervisorStrategy: SupervisorStrategy = ReceiverSupervisorStrategy.defaultStrategy + ): DStream[T] = { + networkStream(new ActorReceiver[T](props, name, storageLevel, supervisorStrategy)) + } + + /** + * Create an input stream that receives messages pushed by a zeromq publisher. + * @param publisherUrl Url of remote zeromq publisher + * @param subscribe topic to subscribe to + * @param bytesToObjects A zeroMQ stream publishes sequence of frames for each topic + * and each frame has sequence of byte thus it needs the converter + * (which might be deserializer of bytes) to translate from sequence + * of sequence of bytes, where sequence refer to a frame + * and sub sequence refer to its payload. + * @param storageLevel RDD storage level. Defaults to memory-only. + */ + def zeroMQStream[T: ClassManifest]( + publisherUrl:String, + subscribe: Subscribe, + bytesToObjects: Seq[Seq[Byte]] ⇒ Iterator[T], + storageLevel: StorageLevel = StorageLevel.MEMORY_ONLY_SER_2, + supervisorStrategy: SupervisorStrategy = ReceiverSupervisorStrategy.defaultStrategy + ): DStream[T] = { + actorStream(Props(new ZeroMQReceiver(publisherUrl,subscribe,bytesToObjects)), + "ZeroMQReceiver", storageLevel, supervisorStrategy) + } + + /** + * Create an input stream that pulls messages from a Kafka Broker. + * @param zkQuorum Zookeper quorum (hostname:port,hostname:port,..). * @param groupId The group id for this consumer. * @param topics Map of (topic_name -> numPartitions) to consume. Each partition is consumed - * in its own thread. - * @param initialOffsets Optional initial offsets for each of the partitions to consume. - * By default the value is pulled from zookeper. - * @param storageLevel RDD storage level. Defaults to memory-only. + * in its own thread. + * @param storageLevel Storage level to use for storing the received objects + * (default: StorageLevel.MEMORY_AND_DISK_SER_2) */ - def kafkaStream[T: ClassManifest]( - hostname: String, - port: Int, + def kafkaStream( + zkQuorum: String, groupId: String, topics: Map[String, Int], - initialOffsets: Map[KafkaPartitionKey, Long] = Map[KafkaPartitionKey, Long](), storageLevel: StorageLevel = StorageLevel.MEMORY_ONLY_SER_2 + ): DStream[String] = { + val kafkaParams = Map[String, String]( + "zk.connect" -> zkQuorum, "groupid" -> groupId, "zk.connectiontimeout.ms" -> "10000") + kafkaStream[String, kafka.serializer.StringDecoder](kafkaParams, topics, storageLevel) + } + + /** + * Create an input stream that pulls messages from a Kafka Broker. + * @param kafkaParams Map of kafka configuration paramaters. + * See: http://kafka.apache.org/configuration.html + * @param topics Map of (topic_name -> numPartitions) to consume. Each partition is consumed + * in its own thread. + * @param storageLevel Storage level to use for storing the received objects + */ + def kafkaStream[T: ClassManifest, D <: kafka.serializer.Decoder[_]: Manifest]( + kafkaParams: Map[String, String], + topics: Map[String, Int], + storageLevel: StorageLevel ): DStream[T] = { - val inputStream = new KafkaInputDStream[T](this, hostname, port, groupId, topics, initialOffsets, storageLevel) + val inputStream = new KafkaInputDStream[T, D](this, kafkaParams, topics, storageLevel) registerInputStream(inputStream) inputStream } /** - * Create a input stream from network source hostname:port. Data is received using - * a TCP socket and the receive bytes is interpreted as UTF8 encoded \n delimited + * Create a input stream from TCP source hostname:port. Data is received using + * a TCP socket and the receive bytes is interpreted as UTF8 encoded `\n` delimited * lines. * @param hostname Hostname to connect to for receiving data * @param port Port to connect to for receiving data * @param storageLevel Storage level to use for storing the received objects * (default: StorageLevel.MEMORY_AND_DISK_SER_2) */ - def networkTextStream( + def socketTextStream( hostname: String, port: Int, storageLevel: StorageLevel = StorageLevel.MEMORY_AND_DISK_SER_2 ): DStream[String] = { - networkStream[String](hostname, port, SocketReceiver.bytesToLines, storageLevel) + socketStream[String](hostname, port, SocketReceiver.bytesToLines, storageLevel) } /** - * Create a input stream from network source hostname:port. Data is received using + * Create a input stream from TCP source hostname:port. Data is received using * a TCP socket and the receive bytes it interepreted as object using the given * converter. * @param hostname Hostname to connect to for receiving data @@ -185,7 +282,7 @@ class StreamingContext private ( * @param storageLevel Storage level to use for storing the received objects * @tparam T Type of the objects received (after converting bytes to objects) */ - def networkStream[T: ClassManifest]( + def socketStream[T: ClassManifest]( hostname: String, port: Int, converter: (InputStream) => Iterator[T], @@ -197,7 +294,7 @@ class StreamingContext private ( } /** - * Creates a input stream from a Flume source. + * Create a input stream from a Flume source. * @param hostname Hostname of the slave machine to which the flume data will be sent * @param port Port of the slave machine to which the flume data will be sent * @param storageLevel Storage level to use for storing the received objects @@ -222,7 +319,7 @@ class StreamingContext private ( * @param storageLevel Storage level to use for storing the received objects * @tparam T Type of the objects in the received blocks */ - def rawNetworkStream[T: ClassManifest]( + def rawSocketStream[T: ClassManifest]( hostname: String, port: Int, storageLevel: StorageLevel = StorageLevel.MEMORY_AND_DISK_SER_2 @@ -233,7 +330,7 @@ class StreamingContext private ( } /** - * Creates a input stream that monitors a Hadoop-compatible filesystem + * Create a input stream that monitors a Hadoop-compatible filesystem * for new files and reads them using the given key-value types and input format. * File names starting with . are ignored. * @param directory HDFS directory to monitor for new file @@ -252,7 +349,7 @@ class StreamingContext private ( } /** - * Creates a input stream that monitors a Hadoop-compatible filesystem + * Create a input stream that monitors a Hadoop-compatible filesystem * for new files and reads them using the given key-value types and input format. * @param directory HDFS directory to monitor for new file * @param filter Function to filter paths to process @@ -271,9 +368,8 @@ class StreamingContext private ( inputStream } - /** - * Creates a input stream that monitors a Hadoop-compatible filesystem + * Create a input stream that monitors a Hadoop-compatible filesystem * for new files and reads them as text files (using key as LongWritable, value * as Text and input format as TextInputFormat). File names starting with . are ignored. * @param directory HDFS directory to monitor for new file @@ -283,17 +379,50 @@ class StreamingContext private ( } /** - * Creates a input stream from an queue of RDDs. In each batch, + * Create a input stream that returns tweets received from Twitter. + * @param username Twitter username + * @param password Twitter password + * @param filters Set of filter strings to get only those tweets that match them + * @param storageLevel Storage level to use for storing the received objects + */ + def twitterStream( + username: String, + password: String, + filters: Seq[String] = Nil, + storageLevel: StorageLevel = StorageLevel.MEMORY_AND_DISK_SER_2 + ): DStream[Status] = { + val inputStream = new TwitterInputDStream(this, username, password, filters, storageLevel) + registerInputStream(inputStream) + inputStream + } + + /** + * Create an input stream from a queue of RDDs. In each batch, + * it will process either one or all of the RDDs returned by the queue. + * @param queue Queue of RDDs + * @param oneAtATime Whether only one RDD should be consumed from the queue in every interval + * @tparam T Type of objects in the RDD + */ + def queueStream[T: ClassManifest]( + queue: Queue[RDD[T]], + oneAtATime: Boolean = true + ): DStream[T] = { + queueStream(queue, oneAtATime, sc.makeRDD(Seq[T](), 1)) + } + + /** + * Create an input stream from a queue of RDDs. In each batch, * it will process either one or all of the RDDs returned by the queue. * @param queue Queue of RDDs * @param oneAtATime Whether only one RDD should be consumed from the queue in every interval - * @param defaultRDD Default RDD is returned by the DStream when the queue is empty + * @param defaultRDD Default RDD is returned by the DStream when the queue is empty. + * Set as null if no RDD should be returned when empty * @tparam T Type of objects in the RDD */ def queueStream[T: ClassManifest]( queue: Queue[RDD[T]], - oneAtATime: Boolean = true, - defaultRDD: RDD[T] = null + oneAtATime: Boolean, + defaultRDD: RDD[T] ): DStream[T] = { val inputStream = new QueueInputDStream(this, queue, oneAtATime, defaultRDD) registerInputStream(inputStream) @@ -308,7 +437,7 @@ class StreamingContext private ( } /** - * Registers an input stream that will be started (InputDStream.start() called) to get the + * Register an input stream that will be started (InputDStream.start() called) to get the * input data. */ def registerInputStream(inputStream: InputDStream[_]) { @@ -316,7 +445,7 @@ class StreamingContext private ( } /** - * Registers an output stream that will be computed every interval + * Register an output stream that will be computed every interval */ def registerOutputStream(outputStream: DStream[_]) { graph.addOutputStream(outputStream) @@ -334,7 +463,7 @@ class StreamingContext private ( } /** - * Starts the execution of the streams. + * Start the execution of the streams. */ def start() { if (checkpointDir != null && checkpointDuration == null && graph != null) { @@ -362,7 +491,7 @@ class StreamingContext private ( } /** - * Sstops the execution of the streams. + * Stop the execution of the streams. */ def stop() { try { @@ -384,14 +513,18 @@ object StreamingContext { new PairDStreamFunctions[K, V](stream) } - protected[streaming] def createNewSparkContext(master: String, appName: String): SparkContext = { - + protected[streaming] def createNewSparkContext( + master: String, + appName: String, + sparkHome: String, + jars: Seq[String], + environment: Map[String, String]): SparkContext = { // Set the default cleaner delay to an hour if not already set. // This should be sufficient for even 1 second interval. if (MetadataCleaner.getDelaySeconds < 0) { MetadataCleaner.setDelaySeconds(3600) } - new SparkContext(master, appName) + new SparkContext(master, appName, sparkHome, jars, environment) } protected[streaming] def rddToFileName[T](prefix: String, suffix: String, time: Time): String = { @@ -408,4 +541,3 @@ object StreamingContext { new Path(sscCheckpointDir, UUID.randomUUID.toString).toString } } - diff --git a/streaming/src/main/scala/spark/streaming/Time.scala b/streaming/src/main/scala/spark/streaming/Time.scala index 5daeb761dd..f14decf08b 100644 --- a/streaming/src/main/scala/spark/streaming/Time.scala +++ b/streaming/src/main/scala/spark/streaming/Time.scala @@ -37,6 +37,19 @@ case class Time(private val millis: Long) { def max(that: Time): Time = if (this > that) this else that + def until(that: Time, interval: Duration): Seq[Time] = { + (this.milliseconds) until (that.milliseconds) by (interval.milliseconds) map (new Time(_)) + } + + def to(that: Time, interval: Duration): Seq[Time] = { + (this.milliseconds) to (that.milliseconds) by (interval.milliseconds) map (new Time(_)) + } + + override def toString: String = (millis.toString + " ms") +} + +object Time { + val ordering = Ordering.by((time: Time) => time.millis) }
\ No newline at end of file diff --git a/streaming/src/main/scala/spark/streaming/api/java/JavaDStream.scala b/streaming/src/main/scala/spark/streaming/api/java/JavaDStream.scala index 2e7466b16c..4d93f0a5f7 100644 --- a/streaming/src/main/scala/spark/streaming/api/java/JavaDStream.scala +++ b/streaming/src/main/scala/spark/streaming/api/java/JavaDStream.scala @@ -4,6 +4,7 @@ import spark.streaming.{Duration, Time, DStream} import spark.api.java.function.{Function => JFunction} import spark.api.java.JavaRDD import spark.storage.StorageLevel +import spark.RDD /** * A Discretized Stream (DStream), the basic abstraction in Spark Streaming, is a continuous @@ -16,9 +17,7 @@ import spark.storage.StorageLevel * * This class contains the basic operations available on all DStreams, such as `map`, `filter` and * `window`. In addition, [[spark.streaming.api.java.JavaPairDStream]] contains operations available - * only on DStreams of key-value pairs, such as `groupByKeyAndWindow` and `join`. These operations - * are automatically available on any DStream of the right type (e.g., DStream[(Int, Int)] through - * implicit conversions when `spark.streaming.StreamingContext._` is imported. + * only on DStreams of key-value pairs, such as `groupByKeyAndWindow` and `join`. * * DStreams internally is characterized by a few basic properties: * - A list of other DStreams that the DStream depends on @@ -26,7 +25,9 @@ import spark.storage.StorageLevel * - A function that is used to generate an RDD after each time interval */ class JavaDStream[T](val dstream: DStream[T])(implicit val classManifest: ClassManifest[T]) - extends JavaDStreamLike[T, JavaDStream[T]] { + extends JavaDStreamLike[T, JavaDStream[T], JavaRDD[T]] { + + override def wrapRDD(rdd: RDD[T]): JavaRDD[T] = JavaRDD.fromRDD(rdd) /** Return a new DStream containing only the elements that satisfy a predicate. */ def filter(f: JFunction[T, java.lang.Boolean]): JavaDStream[T] = @@ -36,7 +37,7 @@ class JavaDStream[T](val dstream: DStream[T])(implicit val classManifest: ClassM def cache(): JavaDStream[T] = dstream.cache() /** Persist RDDs of this DStream with the default storage level (MEMORY_ONLY_SER) */ - def persist(): JavaDStream[T] = dstream.cache() + def persist(): JavaDStream[T] = dstream.persist() /** Persist the RDDs of this DStream with the given storage level */ def persist(storageLevel: StorageLevel): JavaDStream[T] = dstream.persist(storageLevel) @@ -50,34 +51,27 @@ class JavaDStream[T](val dstream: DStream[T])(implicit val classManifest: ClassM } /** - * Return a new DStream which is computed based on windowed batches of this DStream. - * The new DStream generates RDDs with the same interval as this DStream. + * Return a new DStream in which each RDD contains all the elements in seen in a + * sliding window of time over this DStream. The new DStream generates RDDs with + * the same interval as this DStream. * @param windowDuration width of the window; must be a multiple of this DStream's interval. - * @return */ def window(windowDuration: Duration): JavaDStream[T] = dstream.window(windowDuration) /** - * Return a new DStream which is computed based on windowed batches of this DStream. - * @param windowDuration duration (i.e., width) of the window; - * must be a multiple of this DStream's interval + * Return a new DStream in which each RDD contains all the elements in seen in a + * sliding window of time over this DStream. + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval * @param slideDuration sliding interval of the window (i.e., the interval after which - * the new DStream will generate RDDs); must be a multiple of this - * DStream's interval + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval */ def window(windowDuration: Duration, slideDuration: Duration): JavaDStream[T] = dstream.window(windowDuration, slideDuration) /** - * Return a new DStream which computed based on tumbling window on this DStream. - * This is equivalent to window(batchDuration, batchDuration). - * @param batchDuration tumbling window duration; must be a multiple of this DStream's interval - */ - def tumble(batchDuration: Duration): JavaDStream[T] = - dstream.tumble(batchDuration) - - /** * Return a new DStream by unifying data of another DStream with this DStream. * @param that Another DStream having the same interval (i.e., slideDuration) as this DStream. */ diff --git a/streaming/src/main/scala/spark/streaming/api/java/JavaDStreamLike.scala b/streaming/src/main/scala/spark/streaming/api/java/JavaDStreamLike.scala index ec546c8190..548809a359 100644 --- a/streaming/src/main/scala/spark/streaming/api/java/JavaDStreamLike.scala +++ b/streaming/src/main/scala/spark/streaming/api/java/JavaDStreamLike.scala @@ -6,17 +6,20 @@ import java.lang.{Long => JLong} import scala.collection.JavaConversions._ import spark.streaming._ -import spark.api.java.JavaRDD +import spark.api.java.{JavaPairRDD, JavaRDDLike, JavaRDD} import spark.api.java.function.{Function2 => JFunction2, Function => JFunction, _} import java.util import spark.RDD import JavaDStream._ -trait JavaDStreamLike[T, This <: JavaDStreamLike[T, This]] extends Serializable { +trait JavaDStreamLike[T, This <: JavaDStreamLike[T, This, R], R <: JavaRDDLike[T, R]] + extends Serializable { implicit val classManifest: ClassManifest[T] def dstream: DStream[T] + def wrapRDD(in: RDD[T]): R + implicit def scalaIntToJavaLong(in: DStream[Long]): JavaDStream[JLong] = { in.map(new JLong(_)) } @@ -34,6 +37,26 @@ trait JavaDStreamLike[T, This <: JavaDStreamLike[T, This]] extends Serializable def count(): JavaDStream[JLong] = dstream.count() /** + * Return a new DStream in which each RDD contains the counts of each distinct value in + * each RDD of this DStream. Hash partitioning is used to generate the RDDs with + * Spark's default number of partitions. + */ + def countByValue(): JavaPairDStream[T, JLong] = { + JavaPairDStream.scalaToJavaLong(dstream.countByValue()) + } + + /** + * Return a new DStream in which each RDD contains the counts of each distinct value in + * each RDD of this DStream. Hash partitioning is used to generate the RDDs with `numPartitions` + * partitions. + * @param numPartitions number of partitions of each RDD in the new DStream. + */ + def countByValue(numPartitions: Int): JavaPairDStream[T, JLong] = { + JavaPairDStream.scalaToJavaLong(dstream.countByValue(numPartitions)) + } + + + /** * Return a new DStream in which each RDD has a single element generated by counting the number * of elements in a window over this DStream. windowDuration and slideDuration are as defined in the * window() operation. This is equivalent to window(windowDuration, slideDuration).count() @@ -43,6 +66,39 @@ trait JavaDStreamLike[T, This <: JavaDStreamLike[T, This]] extends Serializable } /** + * Return a new DStream in which each RDD contains the count of distinct elements in + * RDDs in a sliding window over this DStream. Hash partitioning is used to generate the RDDs with + * Spark's default number of partitions. + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval + */ + def countByValueAndWindow(windowDuration: Duration, slideDuration: Duration) + : JavaPairDStream[T, JLong] = { + JavaPairDStream.scalaToJavaLong( + dstream.countByValueAndWindow(windowDuration, slideDuration)) + } + + /** + * Return a new DStream in which each RDD contains the count of distinct elements in + * RDDs in a sliding window over this DStream. Hash partitioning is used to generate the RDDs with `numPartitions` + * partitions. + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval + * @param numPartitions number of partitions of each RDD in the new DStream. + */ + def countByValueAndWindow(windowDuration: Duration, slideDuration: Duration, numPartitions: Int) + : JavaPairDStream[T, JLong] = { + JavaPairDStream.scalaToJavaLong( + dstream.countByValueAndWindow(windowDuration, slideDuration, numPartitions)) + } + + /** * Return a new DStream in which each RDD is generated by applying glom() to each RDD of * this DStream. Applying glom() to an RDD coalesces all elements within each partition into * an array. @@ -114,8 +170,38 @@ trait JavaDStreamLike[T, This <: JavaDStreamLike[T, This]] extends Serializable /** * Return a new DStream in which each RDD has a single element generated by reducing all - * elements in a window over this DStream. windowDuration and slideDuration are as defined in the - * window() operation. This is equivalent to window(windowDuration, slideDuration).reduce(reduceFunc) + * elements in a sliding window over this DStream. + * @param reduceFunc associative reduce function + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval + */ + def reduceByWindow( + reduceFunc: (T, T) => T, + windowDuration: Duration, + slideDuration: Duration + ): DStream[T] = { + dstream.reduceByWindow(reduceFunc, windowDuration, slideDuration) + } + + + /** + * Return a new DStream in which each RDD has a single element generated by reducing all + * elements in a sliding window over this DStream. However, the reduction is done incrementally + * using the old window's reduced value : + * 1. reduce the new values that entered the window (e.g., adding new counts) + * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) + * This is more efficient than reduceByWindow without "inverse reduce" function. + * However, it is applicable to only "invertible reduce functions". + * @param reduceFunc associative reduce function + * @param invReduceFunc inverse reduce function + * @param windowDuration width of the window; must be a multiple of this DStream's + * batching interval + * @param slideDuration sliding interval of the window (i.e., the interval after which + * the new DStream will generate RDDs); must be a multiple of this + * DStream's batching interval */ def reduceByWindow( reduceFunc: JFunction2[T, T, T], @@ -129,35 +215,35 @@ trait JavaDStreamLike[T, This <: JavaDStreamLike[T, This]] extends Serializable /** * Return all the RDDs between 'fromDuration' to 'toDuration' (both included) */ - def slice(fromDuration: Duration, toDuration: Duration): JList[JavaRDD[T]] = { - new util.ArrayList(dstream.slice(fromDuration, toDuration).map(new JavaRDD(_)).toSeq) + def slice(fromTime: Time, toTime: Time): JList[R] = { + new util.ArrayList(dstream.slice(fromTime, toTime).map(wrapRDD(_)).toSeq) } /** * Apply a function to each RDD in this DStream. This is an output operator, so * this DStream will be registered as an output stream and therefore materialized. */ - def foreach(foreachFunc: JFunction[JavaRDD[T], Void]) { - dstream.foreach(rdd => foreachFunc.call(new JavaRDD(rdd))) + def foreach(foreachFunc: JFunction[R, Void]) { + dstream.foreach(rdd => foreachFunc.call(wrapRDD(rdd))) } /** * Apply a function to each RDD in this DStream. This is an output operator, so * this DStream will be registered as an output stream and therefore materialized. */ - def foreach(foreachFunc: JFunction2[JavaRDD[T], Time, Void]) { - dstream.foreach((rdd, time) => foreachFunc.call(new JavaRDD(rdd), time)) + def foreach(foreachFunc: JFunction2[R, Time, Void]) { + dstream.foreach((rdd, time) => foreachFunc.call(wrapRDD(rdd), time)) } /** * Return a new DStream in which each RDD is generated by applying a function * on each RDD of this DStream. */ - def transform[U](transformFunc: JFunction[JavaRDD[T], JavaRDD[U]]): JavaDStream[U] = { + def transform[U](transformFunc: JFunction[R, JavaRDD[U]]): JavaDStream[U] = { implicit val cm: ClassManifest[U] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[U]] def scalaTransform (in: RDD[T]): RDD[U] = - transformFunc.call(new JavaRDD[T](in)).rdd + transformFunc.call(wrapRDD(in)).rdd dstream.transform(scalaTransform(_)) } @@ -165,11 +251,41 @@ trait JavaDStreamLike[T, This <: JavaDStreamLike[T, This]] extends Serializable * Return a new DStream in which each RDD is generated by applying a function * on each RDD of this DStream. */ - def transform[U](transformFunc: JFunction2[JavaRDD[T], Time, JavaRDD[U]]): JavaDStream[U] = { + def transform[U](transformFunc: JFunction2[R, Time, JavaRDD[U]]): JavaDStream[U] = { implicit val cm: ClassManifest[U] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[U]] def scalaTransform (in: RDD[T], time: Time): RDD[U] = - transformFunc.call(new JavaRDD[T](in), time).rdd + transformFunc.call(wrapRDD(in), time).rdd + dstream.transform(scalaTransform(_, _)) + } + + /** + * Return a new DStream in which each RDD is generated by applying a function + * on each RDD of this DStream. + */ + def transform[K2, V2](transformFunc: JFunction[R, JavaPairRDD[K2, V2]]): + JavaPairDStream[K2, V2] = { + implicit val cmk: ClassManifest[K2] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[K2]] + implicit val cmv: ClassManifest[V2] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[V2]] + def scalaTransform (in: RDD[T]): RDD[(K2, V2)] = + transformFunc.call(wrapRDD(in)).rdd + dstream.transform(scalaTransform(_)) + } + + /** + * Return a new DStream in which each RDD is generated by applying a function + * on each RDD of this DStream. + */ + def transform[K2, V2](transformFunc: JFunction2[R, Time, JavaPairRDD[K2, V2]]): + JavaPairDStream[K2, V2] = { + implicit val cmk: ClassManifest[K2] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[K2]] + implicit val cmv: ClassManifest[V2] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[V2]] + def scalaTransform (in: RDD[T], time: Time): RDD[(K2, V2)] = + transformFunc.call(wrapRDD(in), time).rdd dstream.transform(scalaTransform(_, _)) } diff --git a/streaming/src/main/scala/spark/streaming/api/java/JavaPairDStream.scala b/streaming/src/main/scala/spark/streaming/api/java/JavaPairDStream.scala index eb2495e3ac..30240cad98 100644 --- a/streaming/src/main/scala/spark/streaming/api/java/JavaPairDStream.scala +++ b/streaming/src/main/scala/spark/streaming/api/java/JavaPairDStream.scala @@ -15,27 +15,30 @@ import org.apache.hadoop.conf.Configuration import spark.api.java.{JavaRDD, JavaPairRDD} import spark.storage.StorageLevel import com.google.common.base.Optional +import spark.RDD class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( implicit val kManifiest: ClassManifest[K], implicit val vManifest: ClassManifest[V]) - extends JavaDStreamLike[(K, V), JavaPairDStream[K, V]] { + extends JavaDStreamLike[(K, V), JavaPairDStream[K, V], JavaPairRDD[K, V]] { + + override def wrapRDD(rdd: RDD[(K, V)]): JavaPairRDD[K, V] = JavaPairRDD.fromRDD(rdd) // ======================================================================= // Methods common to all DStream's // ======================================================================= - /** Returns a new DStream containing only the elements that satisfy a predicate. */ + /** Return a new DStream containing only the elements that satisfy a predicate. */ def filter(f: JFunction[(K, V), java.lang.Boolean]): JavaPairDStream[K, V] = dstream.filter((x => f(x).booleanValue())) - /** Persists RDDs of this DStream with the default storage level (MEMORY_ONLY_SER) */ + /** Persist RDDs of this DStream with the default storage level (MEMORY_ONLY_SER) */ def cache(): JavaPairDStream[K, V] = dstream.cache() - /** Persists RDDs of this DStream with the default storage level (MEMORY_ONLY_SER) */ - def persist(): JavaPairDStream[K, V] = dstream.cache() + /** Persist RDDs of this DStream with the default storage level (MEMORY_ONLY_SER) */ + def persist(): JavaPairDStream[K, V] = dstream.persist() - /** Persists the RDDs of this DStream with the given storage level */ + /** Persist the RDDs of this DStream with the given storage level */ def persist(storageLevel: StorageLevel): JavaPairDStream[K, V] = dstream.persist(storageLevel) /** Method that generates a RDD for the given Duration */ @@ -67,70 +70,32 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( dstream.window(windowDuration, slideDuration) /** - * Returns a new DStream which computed based on tumbling window on this DStream. - * This is equivalent to window(batchDuration, batchDuration). - * @param batchDuration tumbling window duration; must be a multiple of this DStream's interval - */ - def tumble(batchDuration: Duration): JavaPairDStream[K, V] = - dstream.tumble(batchDuration) - - /** - * Returns a new DStream by unifying data of another DStream with this DStream. + * Return a new DStream by unifying data of another DStream with this DStream. * @param that Another DStream having the same interval (i.e., slideDuration) as this DStream. */ def union(that: JavaPairDStream[K, V]): JavaPairDStream[K, V] = dstream.union(that.dstream) - /** - * Return a new DStream in which each RDD is generated by applying a function - * on each RDD of this DStream. - */ - def transform[K2, V2](transformFunc: JFunction[JavaPairRDD[K, V], JavaPairRDD[K2, V2]]): - JavaPairDStream[K2, V2] = { - implicit val cmk: ClassManifest[K2] = - implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[K2]] - implicit val cmv: ClassManifest[V2] = - implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[V2]] - def scalaTransform (in: RDD[(K, V)]): RDD[(K2, V2)] = - transformFunc.call(new JavaPairRDD[K, V](in)).rdd - dstream.transform(scalaTransform(_)) - } - - /** - * Return a new DStream in which each RDD is generated by applying a function - * on each RDD of this DStream. - */ - def transform[K2, V2](transformFunc: JFunction2[JavaPairRDD[K, V], Time, JavaPairRDD[K2, V2]]): - JavaPairDStream[K2, V2] = { - implicit val cmk: ClassManifest[K2] = - implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[K2]] - implicit val cmv: ClassManifest[V2] = - implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[V2]] - def scalaTransform (in: RDD[(K, V)], time: Time): RDD[(K2, V2)] = - transformFunc.call(new JavaPairRDD[K, V](in), time).rdd - dstream.transform(scalaTransform(_, _)) - } - // ======================================================================= // Methods only for PairDStream's // ======================================================================= /** - * Create a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to + * Return a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to * generate the RDDs with Spark's default number of partitions. */ def groupByKey(): JavaPairDStream[K, JList[V]] = dstream.groupByKey().mapValues(seqAsJavaList _) /** - * Create a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to + * Return a new DStream by applying `groupByKey` to each RDD. Hash partitioning is used to * generate the RDDs with `numPartitions` partitions. */ def groupByKey(numPartitions: Int): JavaPairDStream[K, JList[V]] = dstream.groupByKey(numPartitions).mapValues(seqAsJavaList _) /** - * Creates a new DStream by applying `groupByKey` on each RDD of `this` DStream. + * Return a new DStream by applying `groupByKey` on each RDD of `this` DStream. * Therefore, the values for each key in `this` DStream's RDDs are grouped into a * single sequence to generate the RDDs of the new DStream. [[spark.Partitioner]] * is used to control the partitioning of each RDD. @@ -139,7 +104,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( dstream.groupByKey(partitioner).mapValues(seqAsJavaList _) /** - * Create a new DStream by applying `reduceByKey` to each RDD. The values for each key are + * Return a new DStream by applying `reduceByKey` to each RDD. The values for each key are * merged using the associative reduce function. Hash partitioning is used to generate the RDDs * with Spark's default number of partitions. */ @@ -147,7 +112,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( dstream.reduceByKey(func) /** - * Create a new DStream by applying `reduceByKey` to each RDD. The values for each key are + * Return a new DStream by applying `reduceByKey` to each RDD. The values for each key are * merged using the supplied reduce function. Hash partitioning is used to generate the RDDs * with `numPartitions` partitions. */ @@ -155,7 +120,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( dstream.reduceByKey(func, numPartitions) /** - * Create a new DStream by applying `reduceByKey` to each RDD. The values for each key are + * Return a new DStream by applying `reduceByKey` to each RDD. The values for each key are * merged using the supplied reduce function. [[spark.Partitioner]] is used to control the * partitioning of each RDD. */ @@ -179,24 +144,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by counting the number of values of each key in each RDD. Hash - * partitioning is used to generate the RDDs with Spark's `numPartitions` partitions. - */ - def countByKey(numPartitions: Int): JavaPairDStream[K, JLong] = { - JavaPairDStream.scalaToJavaLong(dstream.countByKey(numPartitions)); - } - - - /** - * Create a new DStream by counting the number of values of each key in each RDD. Hash - * partitioning is used to generate the RDDs with the default number of partitions. - */ - def countByKey(): JavaPairDStream[K, JLong] = { - JavaPairDStream.scalaToJavaLong(dstream.countByKey()); - } - - /** - * Creates a new DStream by applying `groupByKey` over a sliding window. This is similar to + * Return a new DStream by applying `groupByKey` over a sliding window. This is similar to * `DStream.groupByKey()` but applies it over a sliding window. The new DStream generates RDDs * with the same interval as this DStream. Hash partitioning is used to generate the RDDs with * Spark's default number of partitions. @@ -208,7 +156,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by applying `groupByKey` over a sliding window. Similar to + * Return a new DStream by applying `groupByKey` over a sliding window. Similar to * `DStream.groupByKey()`, but applies it over a sliding window. Hash partitioning is used to * generate the RDDs with Spark's default number of partitions. * @param windowDuration width of the window; must be a multiple of this DStream's @@ -223,7 +171,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by applying `groupByKey` over a sliding window on `this` DStream. + * Return a new DStream by applying `groupByKey` over a sliding window on `this` DStream. * Similar to `DStream.groupByKey()`, but applies it over a sliding window. * Hash partitioning is used to generate the RDDs with `numPartitions` partitions. * @param windowDuration width of the window; must be a multiple of this DStream's @@ -240,7 +188,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by applying `groupByKey` over a sliding window on `this` DStream. + * Return a new DStream by applying `groupByKey` over a sliding window on `this` DStream. * Similar to `DStream.groupByKey()`, but applies it over a sliding window. * @param windowDuration width of the window; must be a multiple of this DStream's * batching interval @@ -273,7 +221,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by applying `reduceByKey` over a sliding window. This is similar to + * Return a new DStream by applying `reduceByKey` over a sliding window. This is similar to * `DStream.reduceByKey()` but applies it over a sliding window. Hash partitioning is used to * generate the RDDs with Spark's default number of partitions. * @param reduceFunc associative reduce function @@ -292,7 +240,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by applying `reduceByKey` over a sliding window. This is similar to + * Return a new DStream by applying `reduceByKey` over a sliding window. This is similar to * `DStream.reduceByKey()` but applies it over a sliding window. Hash partitioning is used to * generate the RDDs with `numPartitions` partitions. * @param reduceFunc associative reduce function @@ -313,7 +261,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by applying `reduceByKey` over a sliding window. Similar to + * Return a new DStream by applying `reduceByKey` over a sliding window. Similar to * `DStream.reduceByKey()`, but applies it over a sliding window. * @param reduceFunc associative reduce function * @param windowDuration width of the window; must be a multiple of this DStream's @@ -333,7 +281,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by reducing over a using incremental computation. + * Return a new DStream by reducing over a using incremental computation. * The reduced value of over a new window is calculated using the old window's reduce value : * 1. reduce the new values that entered the window (e.g., adding new counts) * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) @@ -358,7 +306,7 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( } /** - * Create a new DStream by reducing over a using incremental computation. + * Return a new DStream by applying incremental `reduceByKey` over a sliding window. * The reduced value of over a new window is calculated using the old window's reduce value : * 1. reduce the new values that entered the window (e.g., adding new counts) * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) @@ -372,25 +320,31 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( * @param slideDuration sliding interval of the window (i.e., the interval after which * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval - * @param numPartitions Number of partitions of each RDD in the new DStream. + * @param numPartitions number of partitions of each RDD in the new DStream. + * @param filterFunc function to filter expired key-value pairs; + * only pairs that satisfy the function are retained + * set this to null if you do not want to filter */ def reduceByKeyAndWindow( reduceFunc: Function2[V, V, V], invReduceFunc: Function2[V, V, V], windowDuration: Duration, slideDuration: Duration, - numPartitions: Int + numPartitions: Int, + filterFunc: JFunction[(K, V), java.lang.Boolean] ): JavaPairDStream[K, V] = { dstream.reduceByKeyAndWindow( reduceFunc, invReduceFunc, windowDuration, slideDuration, - numPartitions) + numPartitions, + (p: (K, V)) => filterFunc(p).booleanValue() + ) } /** - * Create a new DStream by reducing over a using incremental computation. + * Return a new DStream by applying incremental `reduceByKey` over a sliding window. * The reduced value of over a new window is calculated using the old window's reduce value : * 1. reduce the new values that entered the window (e.g., adding new counts) * 2. "inverse reduce" the old values that left the window (e.g., subtracting old counts) @@ -404,49 +358,26 @@ class JavaPairDStream[K, V](val dstream: DStream[(K, V)])( * the new DStream will generate RDDs); must be a multiple of this * DStream's batching interval * @param partitioner Partitioner for controlling the partitioning of each RDD in the new DStream. + * @param filterFunc function to filter expired key-value pairs; + * only pairs that satisfy the function are retained + * set this to null if you do not want to filter */ def reduceByKeyAndWindow( reduceFunc: Function2[V, V, V], invReduceFunc: Function2[V, V, V], windowDuration: Duration, slideDuration: Duration, - partitioner: Partitioner - ): JavaPairDStream[K, V] = { + partitioner: Partitioner, + filterFunc: JFunction[(K, V), java.lang.Boolean] + ): JavaPairDStream[K, V] = { dstream.reduceByKeyAndWindow( reduceFunc, invReduceFunc, windowDuration, slideDuration, - partitioner) - } - - /** - * Create a new DStream by counting the number of values for each key over a window. - * Hash partitioning is used to generate the RDDs with `numPartitions` partitions. - * @param windowDuration width of the window; must be a multiple of this DStream's - * batching interval - * @param slideDuration sliding interval of the window (i.e., the interval after which - * the new DStream will generate RDDs); must be a multiple of this - * DStream's batching interval - */ - def countByKeyAndWindow(windowDuration: Duration, slideDuration: Duration) - : JavaPairDStream[K, JLong] = { - JavaPairDStream.scalaToJavaLong(dstream.countByKeyAndWindow(windowDuration, slideDuration)) - } - - /** - * Create a new DStream by counting the number of values for each key over a window. - * Hash partitioning is used to generate the RDDs with `numPartitions` partitions. - * @param windowDuration width of the window; must be a multiple of this DStream's - * batching interval - * @param slideDuration sliding interval of the window (i.e., the interval after which - * the new DStream will generate RDDs); must be a multiple of this - * DStream's batching interval - * @param numPartitions Number of partitions of each RDD in the new DStream. - */ - def countByKeyAndWindow(windowDuration: Duration, slideDuration: Duration, numPartitions: Int) - : JavaPairDStream[K, Long] = { - dstream.countByKeyAndWindow(windowDuration, slideDuration, numPartitions) + partitioner, + (p: (K, V)) => filterFunc(p).booleanValue() + ) } private def convertUpdateStateFunction[S](in: JFunction2[JList[V], Optional[S], Optional[S]]): diff --git a/streaming/src/main/scala/spark/streaming/api/java/JavaStreamingContext.scala b/streaming/src/main/scala/spark/streaming/api/java/JavaStreamingContext.scala index e5b5e9ac23..4259d4891c 100644 --- a/streaming/src/main/scala/spark/streaming/api/java/JavaStreamingContext.scala +++ b/streaming/src/main/scala/spark/streaming/api/java/JavaStreamingContext.scala @@ -1,16 +1,26 @@ package spark.streaming.api.java -import scala.collection.JavaConversions._ -import java.lang.{Long => JLong, Integer => JInt} - import spark.streaming._ -import dstream._ +import receivers.{ActorReceiver, ReceiverSupervisorStrategy} +import spark.streaming.dstream._ import spark.storage.StorageLevel + import spark.api.java.function.{Function => JFunction, Function2 => JFunction2} +import spark.api.java.{JavaSparkContext, JavaRDD} + import org.apache.hadoop.mapreduce.{InputFormat => NewInputFormat} + +import twitter4j.Status + +import akka.actor.Props +import akka.actor.SupervisorStrategy +import akka.zeromq.Subscribe + +import scala.collection.JavaConversions._ + +import java.lang.{Long => JLong, Integer => JInt} import java.io.InputStream import java.util.{Map => JMap} -import spark.api.java.{JavaSparkContext, JavaRDD} /** * A StreamingContext is the main entry point for Spark Streaming functionality. Besides the basic @@ -31,10 +41,63 @@ class JavaStreamingContext(val ssc: StreamingContext) { * @param batchDuration The time interval at which streaming data will be divided into batches */ def this(master: String, appName: String, batchDuration: Duration) = - this(new StreamingContext(master, appName, batchDuration)) + this(new StreamingContext(master, appName, batchDuration, null, Nil, Map())) /** * Creates a StreamingContext. + * @param master Name of the Spark Master + * @param appName Name to be used when registering with the scheduler + * @param batchDuration The time interval at which streaming data will be divided into batches + * @param sparkHome The SPARK_HOME directory on the slave nodes + * @param jarFile JAR file containing job code, to ship to cluster. This can be a path on the local + * file system or an HDFS, HTTP, HTTPS, or FTP URL. + */ + def this( + master: String, + appName: String, + batchDuration: Duration, + sparkHome: String, + jarFile: String) = + this(new StreamingContext(master, appName, batchDuration, sparkHome, Seq(jarFile), Map())) + + /** + * Creates a StreamingContext. + * @param master Name of the Spark Master + * @param appName Name to be used when registering with the scheduler + * @param batchDuration The time interval at which streaming data will be divided into batches + * @param sparkHome The SPARK_HOME directory on the slave nodes + * @param jars Collection of JARs to send to the cluster. These can be paths on the local file + * system or HDFS, HTTP, HTTPS, or FTP URLs. + */ + def this( + master: String, + appName: String, + batchDuration: Duration, + sparkHome: String, + jars: Array[String]) = + this(new StreamingContext(master, appName, batchDuration, sparkHome, jars, Map())) + + /** + * Creates a StreamingContext. + * @param master Name of the Spark Master + * @param appName Name to be used when registering with the scheduler + * @param batchDuration The time interval at which streaming data will be divided into batches + * @param sparkHome The SPARK_HOME directory on the slave nodes + * @param jars Collection of JARs to send to the cluster. These can be paths on the local file + * system or HDFS, HTTP, HTTPS, or FTP URLs. + * @param environment Environment variables to set on worker nodes + */ + def this( + master: String, + appName: String, + batchDuration: Duration, + sparkHome: String, + jars: Array[String], + environment: JMap[String, String]) = + this(new StreamingContext(master, appName, batchDuration, sparkHome, jars, environment)) + + /** + * Creates a StreamingContext using an existing SparkContext. * @param sparkContext The underlying JavaSparkContext to use * @param batchDuration The time interval at which streaming data will be divided into batches */ @@ -53,77 +116,66 @@ class JavaStreamingContext(val ssc: StreamingContext) { /** * Create an input stream that pulls messages form a Kafka Broker. - * @param hostname Zookeper hostname. - * @param port Zookeper port. + * @param zkQuorum Zookeper quorum (hostname:port,hostname:port,..). * @param groupId The group id for this consumer. * @param topics Map of (topic_name -> numPartitions) to consume. Each partition is consumed * in its own thread. */ - def kafkaStream[T]( - hostname: String, - port: Int, + def kafkaStream( + zkQuorum: String, groupId: String, topics: JMap[String, JInt]) - : JavaDStream[T] = { - implicit val cmt: ClassManifest[T] = - implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] - ssc.kafkaStream[T](hostname, port, groupId, Map(topics.mapValues(_.intValue()).toSeq: _*)) + : JavaDStream[String] = { + implicit val cmt: ClassManifest[String] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[String]] + ssc.kafkaStream(zkQuorum, groupId, Map(topics.mapValues(_.intValue()).toSeq: _*), + StorageLevel.MEMORY_ONLY_SER_2) } /** * Create an input stream that pulls messages form a Kafka Broker. - * @param hostname Zookeper hostname. - * @param port Zookeper port. + * @param zkQuorum Zookeper quorum (hostname:port,hostname:port,..). * @param groupId The group id for this consumer. * @param topics Map of (topic_name -> numPartitions) to consume. Each partition is consumed - * in its own thread. - * @param initialOffsets Optional initial offsets for each of the partitions to consume. - * By default the value is pulled from zookeper. + * in its own thread. + * @param storageLevel RDD storage level. Defaults to memory-only + * */ - def kafkaStream[T]( - hostname: String, - port: Int, + def kafkaStream( + zkQuorum: String, groupId: String, topics: JMap[String, JInt], - initialOffsets: JMap[KafkaPartitionKey, JLong]) - : JavaDStream[T] = { - implicit val cmt: ClassManifest[T] = - implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] - ssc.kafkaStream[T]( - hostname, - port, - groupId, - Map(topics.mapValues(_.intValue()).toSeq: _*), - Map(initialOffsets.mapValues(_.longValue()).toSeq: _*)) + storageLevel: StorageLevel) + : JavaDStream[String] = { + implicit val cmt: ClassManifest[String] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[String]] + ssc.kafkaStream(zkQuorum, groupId, Map(topics.mapValues(_.intValue()).toSeq: _*), + storageLevel) } /** * Create an input stream that pulls messages form a Kafka Broker. - * @param hostname Zookeper hostname. - * @param port Zookeper port. - * @param groupId The group id for this consumer. + * @param typeClass Type of RDD + * @param decoderClass Type of kafka decoder + * @param kafkaParams Map of kafka configuration paramaters. + * See: http://kafka.apache.org/configuration.html * @param topics Map of (topic_name -> numPartitions) to consume. Each partition is consumed * in its own thread. - * @param initialOffsets Optional initial offsets for each of the partitions to consume. - * By default the value is pulled from zookeper. * @param storageLevel RDD storage level. Defaults to memory-only */ - def kafkaStream[T]( - hostname: String, - port: Int, - groupId: String, + def kafkaStream[T, D <: kafka.serializer.Decoder[_]]( + typeClass: Class[T], + decoderClass: Class[D], + kafkaParams: JMap[String, String], topics: JMap[String, JInt], - initialOffsets: JMap[KafkaPartitionKey, JLong], storageLevel: StorageLevel) : JavaDStream[T] = { implicit val cmt: ClassManifest[T] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] - ssc.kafkaStream[T]( - hostname, - port, - groupId, + implicit val cmd: Manifest[D] = implicitly[Manifest[AnyRef]].asInstanceOf[Manifest[D]] + ssc.kafkaStream[T, D]( + kafkaParams.toMap, Map(topics.mapValues(_.intValue()).toSeq: _*), - Map(initialOffsets.mapValues(_.longValue()).toSeq: _*), storageLevel) } @@ -136,9 +188,9 @@ class JavaStreamingContext(val ssc: StreamingContext) { * @param storageLevel Storage level to use for storing the received objects * (default: StorageLevel.MEMORY_AND_DISK_SER_2) */ - def networkTextStream(hostname: String, port: Int, storageLevel: StorageLevel) + def socketTextStream(hostname: String, port: Int, storageLevel: StorageLevel) : JavaDStream[String] = { - ssc.networkTextStream(hostname, port, storageLevel) + ssc.socketTextStream(hostname, port, storageLevel) } /** @@ -148,8 +200,8 @@ class JavaStreamingContext(val ssc: StreamingContext) { * @param hostname Hostname to connect to for receiving data * @param port Port to connect to for receiving data */ - def networkTextStream(hostname: String, port: Int): JavaDStream[String] = { - ssc.networkTextStream(hostname, port) + def socketTextStream(hostname: String, port: Int): JavaDStream[String] = { + ssc.socketTextStream(hostname, port) } /** @@ -162,7 +214,7 @@ class JavaStreamingContext(val ssc: StreamingContext) { * @param storageLevel Storage level to use for storing the received objects * @tparam T Type of the objects received (after converting bytes to objects) */ - def networkStream[T]( + def socketStream[T]( hostname: String, port: Int, converter: JFunction[InputStream, java.lang.Iterable[T]], @@ -171,7 +223,7 @@ class JavaStreamingContext(val ssc: StreamingContext) { def fn = (x: InputStream) => converter.apply(x).toIterator implicit val cmt: ClassManifest[T] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] - ssc.networkStream(hostname, port, fn, storageLevel) + ssc.socketStream(hostname, port, fn, storageLevel) } /** @@ -194,13 +246,13 @@ class JavaStreamingContext(val ssc: StreamingContext) { * @param storageLevel Storage level to use for storing the received objects * @tparam T Type of the objects in the received blocks */ - def rawNetworkStream[T]( + def rawSocketStream[T]( hostname: String, port: Int, storageLevel: StorageLevel): JavaDStream[T] = { implicit val cmt: ClassManifest[T] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] - JavaDStream.fromDStream(ssc.rawNetworkStream(hostname, port, storageLevel)) + JavaDStream.fromDStream(ssc.rawSocketStream(hostname, port, storageLevel)) } /** @@ -212,10 +264,10 @@ class JavaStreamingContext(val ssc: StreamingContext) { * @param port Port to connect to for receiving data * @tparam T Type of the objects in the received blocks */ - def rawNetworkStream[T](hostname: String, port: Int): JavaDStream[T] = { + def rawSocketStream[T](hostname: String, port: Int): JavaDStream[T] = { implicit val cmt: ClassManifest[T] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] - JavaDStream.fromDStream(ssc.rawNetworkStream(hostname, port)) + JavaDStream.fromDStream(ssc.rawSocketStream(hostname, port)) } /** @@ -254,15 +306,182 @@ class JavaStreamingContext(val ssc: StreamingContext) { * @param hostname Hostname of the slave machine to which the flume data will be sent * @param port Port of the slave machine to which the flume data will be sent */ - def flumeStream(hostname: String, port: Int): - JavaDStream[SparkFlumeEvent] = { + def flumeStream(hostname: String, port: Int): JavaDStream[SparkFlumeEvent] = { ssc.flumeStream(hostname, port) } /** + * Create a input stream that returns tweets received from Twitter. + * @param username Twitter username + * @param password Twitter password + * @param filters Set of filter strings to get only those tweets that match them + * @param storageLevel Storage level to use for storing the received objects + */ + def twitterStream( + username: String, + password: String, + filters: Array[String], + storageLevel: StorageLevel + ): JavaDStream[Status] = { + ssc.twitterStream(username, password, filters, storageLevel) + } + + /** + * Create a input stream that returns tweets received from Twitter. + * @param username Twitter username + * @param password Twitter password + * @param filters Set of filter strings to get only those tweets that match them + */ + def twitterStream( + username: String, + password: String, + filters: Array[String] + ): JavaDStream[Status] = { + ssc.twitterStream(username, password, filters) + } + + /** + * Create a input stream that returns tweets received from Twitter. + * @param username Twitter username + * @param password Twitter password + */ + def twitterStream( + username: String, + password: String + ): JavaDStream[Status] = { + ssc.twitterStream(username, password) + } + + /** + * Create an input stream with any arbitrary user implemented actor receiver. + * @param props Props object defining creation of the actor + * @param name Name of the actor + * @param storageLevel Storage level to use for storing the received objects + * + * @note An important point to note: + * Since Actor may exist outside the spark framework, It is thus user's responsibility + * to ensure the type safety, i.e parametrized type of data received and actorStream + * should be same. + */ + def actorStream[T]( + props: Props, + name: String, + storageLevel: StorageLevel, + supervisorStrategy: SupervisorStrategy + ): JavaDStream[T] = { + implicit val cm: ClassManifest[T] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] + ssc.actorStream[T](props, name, storageLevel, supervisorStrategy) + } + + /** + * Create an input stream with any arbitrary user implemented actor receiver. + * @param props Props object defining creation of the actor + * @param name Name of the actor + * @param storageLevel Storage level to use for storing the received objects + * + * @note An important point to note: + * Since Actor may exist outside the spark framework, It is thus user's responsibility + * to ensure the type safety, i.e parametrized type of data received and actorStream + * should be same. + */ + def actorStream[T]( + props: Props, + name: String, + storageLevel: StorageLevel + ): JavaDStream[T] = { + implicit val cm: ClassManifest[T] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] + ssc.actorStream[T](props, name, storageLevel) + } + + /** + * Create an input stream with any arbitrary user implemented actor receiver. + * @param props Props object defining creation of the actor + * @param name Name of the actor + * + * @note An important point to note: + * Since Actor may exist outside the spark framework, It is thus user's responsibility + * to ensure the type safety, i.e parametrized type of data received and actorStream + * should be same. + */ + def actorStream[T]( + props: Props, + name: String + ): JavaDStream[T] = { + implicit val cm: ClassManifest[T] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] + ssc.actorStream[T](props, name) + } + + /** + * Create an input stream that receives messages pushed by a zeromq publisher. + * @param publisherUrl Url of remote zeromq publisher + * @param subscribe topic to subscribe to + * @param bytesToObjects A zeroMQ stream publishes sequence of frames for each topic and each frame has sequence + * of byte thus it needs the converter(which might be deserializer of bytes) + * to translate from sequence of sequence of bytes, where sequence refer to a frame + * and sub sequence refer to its payload. + * @param storageLevel Storage level to use for storing the received objects + */ + def zeroMQStream[T]( + publisherUrl:String, + subscribe: Subscribe, + bytesToObjects: Seq[Seq[Byte]] ⇒ Iterator[T], + storageLevel: StorageLevel, + supervisorStrategy: SupervisorStrategy + ): JavaDStream[T] = { + implicit val cm: ClassManifest[T] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] + ssc.zeroMQStream[T](publisherUrl, subscribe, bytesToObjects, storageLevel, supervisorStrategy) + } + + /** + * Create an input stream that receives messages pushed by a zeromq publisher. + * @param publisherUrl Url of remote zeromq publisher + * @param subscribe topic to subscribe to + * @param bytesToObjects A zeroMQ stream publishes sequence of frames for each topic and each frame has sequence + * of byte thus it needs the converter(which might be deserializer of bytes) + * to translate from sequence of sequence of bytes, where sequence refer to a frame + * and sub sequence refer to its payload. + * @param storageLevel RDD storage level. Defaults to memory-only. + */ + def zeroMQStream[T]( + publisherUrl:String, + subscribe: Subscribe, + bytesToObjects: JFunction[Array[Array[Byte]], java.lang.Iterable[T]], + storageLevel: StorageLevel + ): JavaDStream[T] = { + implicit val cm: ClassManifest[T] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] + def fn(x: Seq[Seq[Byte]]) = bytesToObjects.apply(x.map(_.toArray).toArray).toIterator + ssc.zeroMQStream[T](publisherUrl, subscribe, fn, storageLevel) + } + + /** + * Create an input stream that receives messages pushed by a zeromq publisher. + * @param publisherUrl Url of remote zeromq publisher + * @param subscribe topic to subscribe to + * @param bytesToObjects A zeroMQ stream publishes sequence of frames for each topic and each frame has sequence + * of byte thus it needs the converter(which might be deserializer of bytes) + * to translate from sequence of sequence of bytes, where sequence refer to a frame + * and sub sequence refer to its payload. + */ + def zeroMQStream[T]( + publisherUrl:String, + subscribe: Subscribe, + bytesToObjects: JFunction[Array[Array[Byte]], java.lang.Iterable[T]] + ): JavaDStream[T] = { + implicit val cm: ClassManifest[T] = + implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] + def fn(x: Seq[Seq[Byte]]) = bytesToObjects.apply(x.map(_.toArray).toArray).toIterator + ssc.zeroMQStream[T](publisherUrl, subscribe, fn) + } + + /** * Registers an output stream that will be computed every interval */ - def registerOutputStream(outputStream: JavaDStreamLike[_, _]) { + def registerOutputStream(outputStream: JavaDStreamLike[_, _, _]) { ssc.registerOutputStream(outputStream.dstream) } @@ -322,12 +541,11 @@ class JavaStreamingContext(val ssc: StreamingContext) { /** * Sets the context to periodically checkpoint the DStream operations for master - * fault-tolerance. By default, the graph will be checkpointed every batch interval. + * fault-tolerance. The graph will be checkpointed every batch interval. * @param directory HDFS-compatible directory where the checkpoint data will be reliably stored - * @param interval checkpoint interval */ - def checkpoint(directory: String, interval: Duration = null) { - ssc.checkpoint(directory, interval) + def checkpoint(directory: String) { + ssc.checkpoint(directory) } /** diff --git a/streaming/src/main/scala/spark/streaming/dstream/FileInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/FileInputDStream.scala index 1e6ad84b44..41b9bd9461 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/FileInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/FileInputDStream.scala @@ -2,13 +2,14 @@ package spark.streaming.dstream import spark.RDD import spark.rdd.UnionRDD -import spark.streaming.{StreamingContext, Time} +import spark.streaming.{DStreamCheckpointData, StreamingContext, Time} import org.apache.hadoop.fs.{FileSystem, Path, PathFilter} import org.apache.hadoop.conf.Configuration import org.apache.hadoop.mapreduce.{InputFormat => NewInputFormat} -import scala.collection.mutable.HashSet +import scala.collection.mutable.{HashSet, HashMap} +import java.io.{ObjectInputStream, IOException} private[streaming] class FileInputDStream[K: ClassManifest, V: ClassManifest, F <: NewInputFormat[K,V] : ClassManifest]( @@ -18,28 +19,23 @@ class FileInputDStream[K: ClassManifest, V: ClassManifest, F <: NewInputFormat[K newFilesOnly: Boolean = true) extends InputDStream[(K, V)](ssc_) { - @transient private var path_ : Path = null - @transient private var fs_ : FileSystem = null - - var lastModTime = 0L - val lastModTimeFiles = new HashSet[String]() + protected[streaming] override val checkpointData = new FileInputDStreamCheckpointData - def path(): Path = { - if (path_ == null) path_ = new Path(directory) - path_ - } + // Latest file mod time seen till any point of time + private val lastModTimeFiles = new HashSet[String]() + private var lastModTime = 0L - def fs(): FileSystem = { - if (fs_ == null) fs_ = path.getFileSystem(new Configuration()) - fs_ - } + @transient private var path_ : Path = null + @transient private var fs_ : FileSystem = null + @transient private[streaming] var files = new HashMap[Time, Array[String]] override def start() { if (newFilesOnly) { - lastModTime = System.currentTimeMillis() + lastModTime = graph.zeroTime.milliseconds } else { lastModTime = 0 } + logDebug("LastModTime initialized to " + lastModTime + ", new files only = " + newFilesOnly) } override def stop() { } @@ -49,38 +45,50 @@ class FileInputDStream[K: ClassManifest, V: ClassManifest, F <: NewInputFormat[K * a union RDD out of them. Note that this maintains the list of files that were processed * in the latest modification time in the previous call to this method. This is because the * modification time returned by the FileStatus API seems to return times only at the - * granularity of seconds. Hence, new files may have the same modification time as the - * latest modification time in the previous call to this method and the list of files - * maintained is used to filter the one that have been processed. + * granularity of seconds. And new files may have the same modification time as the + * latest modification time in the previous call to this method yet was not reported in + * the previous call. */ override def compute(validTime: Time): Option[RDD[(K, V)]] = { + assert(validTime.milliseconds >= lastModTime, "Trying to get new files for really old time [" + validTime + " < " + lastModTime) + // Create the filter for selecting new files val newFilter = new PathFilter() { + // Latest file mod time seen in this round of fetching files and its corresponding files var latestModTime = 0L val latestModTimeFiles = new HashSet[String]() def accept(path: Path): Boolean = { - if (!filter(path)) { + if (!filter(path)) { // Reject file if it does not satisfy filter + logDebug("Rejected by filter " + path) return false - } else { + } else { // Accept file only if val modTime = fs.getFileStatus(path).getModificationTime() - if (modTime < lastModTime){ - return false + logDebug("Mod time for " + path + " is " + modTime) + if (modTime < lastModTime) { + logDebug("Mod time less than last mod time") + return false // If the file was created before the last time it was called } else if (modTime == lastModTime && lastModTimeFiles.contains(path.toString)) { - return false + logDebug("Mod time equal to last mod time, but file considered already") + return false // If the file was created exactly as lastModTime but not reported yet + } else if (modTime > validTime.milliseconds) { + logDebug("Mod time more than valid time") + return false // If the file was created after the time this function call requires } if (modTime > latestModTime) { latestModTime = modTime latestModTimeFiles.clear() + logDebug("Latest mod time updated to " + latestModTime) } latestModTimeFiles += path.toString + logDebug("Accepted " + path) return true } } } - - val newFiles = fs.listStatus(path, newFilter) - logInfo("New files: " + newFiles.map(_.getPath).mkString(", ")) + logDebug("Finding new files at time " + validTime + " for last mod time = " + lastModTime) + val newFiles = fs.listStatus(path, newFilter).map(_.getPath.toString) + logInfo("New files at time " + validTime + ":\n" + newFiles.mkString("\n")) if (newFiles.length > 0) { // Update the modification time and the files processed for that modification time if (lastModTime != newFilter.latestModTime) { @@ -88,10 +96,81 @@ class FileInputDStream[K: ClassManifest, V: ClassManifest, F <: NewInputFormat[K lastModTimeFiles.clear() } lastModTimeFiles ++= newFilter.latestModTimeFiles + logDebug("Last mod time updated to " + lastModTime) + } + files += ((validTime, newFiles)) + Some(filesToRDD(newFiles)) + } + + /** Clear the old time-to-files mappings along with old RDDs */ + protected[streaming] override def clearOldMetadata(time: Time) { + super.clearOldMetadata(time) + val oldFiles = files.filter(_._1 <= (time - rememberDuration)) + files --= oldFiles.keys + logInfo("Cleared " + oldFiles.size + " old files that were older than " + + (time - rememberDuration) + ": " + oldFiles.keys.mkString(", ")) + logDebug("Cleared files are:\n" + + oldFiles.map(p => (p._1, p._2.mkString(", "))).mkString("\n")) + } + + /** Generate one RDD from an array of files */ + protected[streaming] def filesToRDD(files: Seq[String]): RDD[(K, V)] = { + new UnionRDD( + context.sparkContext, + files.map(file => context.sparkContext.newAPIHadoopFile[K, V, F](file)) + ) + } + + private def path: Path = { + if (path_ == null) path_ = new Path(directory) + path_ + } + + private def fs: FileSystem = { + if (fs_ == null) fs_ = path.getFileSystem(new Configuration()) + fs_ + } + + @throws(classOf[IOException]) + private def readObject(ois: ObjectInputStream) { + logDebug(this.getClass().getSimpleName + ".readObject used") + ois.defaultReadObject() + generatedRDDs = new HashMap[Time, RDD[(K,V)]] () + files = new HashMap[Time, Array[String]] + } + + /** + * A custom version of the DStreamCheckpointData that stores names of + * Hadoop files as checkpoint data. + */ + private[streaming] + class FileInputDStreamCheckpointData extends DStreamCheckpointData(this) { + + def hadoopFiles = data.asInstanceOf[HashMap[Time, Array[String]]] + + override def update() { + hadoopFiles.clear() + hadoopFiles ++= files + } + + override def cleanup() { } + + override def restore() { + hadoopFiles.foreach { + case (t, f) => { + // Restore the metadata in both files and generatedRDDs + logInfo("Restoring files for time " + t + " - " + + f.mkString("[", ", ", "]") ) + files += ((t, f)) + generatedRDDs += ((t, filesToRDD(f))) + } + } + } + + override def toString() = { + "[\n" + hadoopFiles.size + " file sets\n" + + hadoopFiles.map(p => (p._1, p._2.mkString(", "))).mkString("\n") + "\n]" } - val newRDD = new UnionRDD(ssc.sc, newFiles.map( - file => ssc.sc.newAPIHadoopFile[K, V, F](file.getPath.toString))) - Some(newRDD) } } @@ -100,3 +179,4 @@ object FileInputDStream { def defaultFilter(path: Path): Boolean = !path.getName().startsWith(".") } + diff --git a/streaming/src/main/scala/spark/streaming/dstream/FlumeInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/FlumeInputDStream.scala index efc7058480..c9644b3a83 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/FlumeInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/FlumeInputDStream.scala @@ -25,7 +25,7 @@ class FlumeInputDStream[T: ClassManifest]( storageLevel: StorageLevel ) extends NetworkInputDStream[SparkFlumeEvent](ssc_) { - override def createReceiver(): NetworkReceiver[SparkFlumeEvent] = { + override def getReceiver(): NetworkReceiver[SparkFlumeEvent] = { new FlumeReceiver(host, port, storageLevel) } } @@ -134,4 +134,4 @@ class FlumeReceiver( } override def getLocationPreference = Some(host) -}
\ No newline at end of file +} diff --git a/streaming/src/main/scala/spark/streaming/dstream/InputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/InputDStream.scala index 980ca5177e..3c5d43a609 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/InputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/InputDStream.scala @@ -1,10 +1,42 @@ package spark.streaming.dstream -import spark.streaming.{Duration, StreamingContext, DStream} +import spark.streaming.{Time, Duration, StreamingContext, DStream} +/** + * This is the abstract base class for all input streams. This class provides to methods + * start() and stop() which called by the scheduler to start and stop receiving data/ + * Input streams that can generated RDDs from new data just by running a service on + * the driver node (that is, without running a receiver onworker nodes) can be + * implemented by directly subclassing this InputDStream. For example, + * FileInputDStream, a subclass of InputDStream, monitors a HDFS directory for + * new files and generates RDDs on the new files. For implementing input streams + * that requires running a receiver on the worker nodes, use NetworkInputDStream + * as the parent class. + */ abstract class InputDStream[T: ClassManifest] (@transient ssc_ : StreamingContext) extends DStream[T](ssc_) { + var lastValidTime: Time = null + + /** + * Checks whether the 'time' is valid wrt slideDuration for generating RDD. + * Additionally it also ensures valid times are in strictly increasing order. + * This ensures that InputDStream.compute() is called strictly on increasing + * times. + */ + override protected def isTimeValid(time: Time): Boolean = { + if (!super.isTimeValid(time)) { + false // Time not valid + } else { + // Time is valid, but check it it is more than lastValidTime + if (lastValidTime != null && time < lastValidTime) { + logWarning("isTimeValid called with " + time + " where as last valid time is " + lastValidTime) + } + lastValidTime = time + true + } + } + override def dependencies = List() override def slideDuration: Duration = { @@ -13,7 +45,9 @@ abstract class InputDStream[T: ClassManifest] (@transient ssc_ : StreamingContex ssc.graph.batchDuration } + /** Method called to start receiving data. Subclasses must implement this method. */ def start() + /** Method called to stop receiving data. Subclasses must implement this method. */ def stop() } diff --git a/streaming/src/main/scala/spark/streaming/dstream/KafkaInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/KafkaInputDStream.scala index 2b4740bdf7..55d2957be4 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/KafkaInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/KafkaInputDStream.scala @@ -9,112 +9,51 @@ import java.util.concurrent.Executors import kafka.consumer._ import kafka.message.{Message, MessageSet, MessageAndMetadata} -import kafka.serializer.StringDecoder +import kafka.serializer.Decoder import kafka.utils.{Utils, ZKGroupTopicDirs} import kafka.utils.ZkUtils._ +import kafka.utils.ZKStringSerializer +import org.I0Itec.zkclient._ +import scala.collection.Map import scala.collection.mutable.HashMap import scala.collection.JavaConversions._ -// Key for a specific Kafka Partition: (broker, topic, group, part) -case class KafkaPartitionKey(brokerId: Int, topic: String, groupId: String, partId: Int) -// NOT USED - Originally intended for fault-tolerance -// Metadata for a Kafka Stream that it sent to the Master -private[streaming] -case class KafkaInputDStreamMetadata(timestamp: Long, data: Map[KafkaPartitionKey, Long]) -// NOT USED - Originally intended for fault-tolerance -// Checkpoint data specific to a KafkaInputDstream -private[streaming] -case class KafkaDStreamCheckpointData(kafkaRdds: HashMap[Time, Any], - savedOffsets: Map[KafkaPartitionKey, Long]) extends DStreamCheckpointData(kafkaRdds) - /** * Input stream that pulls messages from a Kafka Broker. * - * @param host Zookeper hostname. - * @param port Zookeper port. - * @param groupId The group id for this consumer. + * @param kafkaParams Map of kafka configuration paramaters. See: http://kafka.apache.org/configuration.html * @param topics Map of (topic_name -> numPartitions) to consume. Each partition is consumed * in its own thread. - * @param initialOffsets Optional initial offsets for each of the partitions to consume. - * By default the value is pulled from zookeper. * @param storageLevel RDD storage level. */ private[streaming] -class KafkaInputDStream[T: ClassManifest]( +class KafkaInputDStream[T: ClassManifest, D <: Decoder[_]: Manifest]( @transient ssc_ : StreamingContext, - host: String, - port: Int, - groupId: String, + kafkaParams: Map[String, String], topics: Map[String, Int], - initialOffsets: Map[KafkaPartitionKey, Long], storageLevel: StorageLevel ) extends NetworkInputDStream[T](ssc_ ) with Logging { - // Metadata that keeps track of which messages have already been consumed. - var savedOffsets = HashMap[Long, Map[KafkaPartitionKey, Long]]() - - /* NOT USED - Originally intended for fault-tolerance - - // In case of a failure, the offets for a particular timestamp will be restored. - @transient var restoredOffsets : Map[KafkaPartitionKey, Long] = null - - - override protected[streaming] def addMetadata(metadata: Any) { - metadata match { - case x : KafkaInputDStreamMetadata => - savedOffsets(x.timestamp) = x.data - // TOOD: Remove logging - logInfo("New saved Offsets: " + savedOffsets) - case _ => logInfo("Received unknown metadata: " + metadata.toString) - } - } - - override protected[streaming] def updateCheckpointData(currentTime: Time) { - super.updateCheckpointData(currentTime) - if(savedOffsets.size > 0) { - // Find the offets that were stored before the checkpoint was initiated - val key = savedOffsets.keys.toList.sortWith(_ < _).filter(_ < currentTime.millis).last - val latestOffsets = savedOffsets(key) - logInfo("Updating KafkaDStream checkpoint data: " + latestOffsets.toString) - checkpointData = KafkaDStreamCheckpointData(checkpointData.rdds, latestOffsets) - // TODO: This may throw out offsets that are created after the checkpoint, - // but it's unlikely we'll need them. - savedOffsets.clear() - } - } - - override protected[streaming] def restoreCheckpointData() { - super.restoreCheckpointData() - logInfo("Restoring KafkaDStream checkpoint data.") - checkpointData match { - case x : KafkaDStreamCheckpointData => - restoredOffsets = x.savedOffsets - logInfo("Restored KafkaDStream offsets: " + savedOffsets) - } - } */ - def createReceiver(): NetworkReceiver[T] = { - new KafkaReceiver(host, port, groupId, topics, initialOffsets, storageLevel) + def getReceiver(): NetworkReceiver[T] = { + new KafkaReceiver[T, D](kafkaParams, topics, storageLevel) .asInstanceOf[NetworkReceiver[T]] } } private[streaming] -class KafkaReceiver(host: String, port: Int, groupId: String, - topics: Map[String, Int], initialOffsets: Map[KafkaPartitionKey, Long], - storageLevel: StorageLevel) extends NetworkReceiver[Any] { - - // Timeout for establishing a connection to Zookeper in ms. - val ZK_TIMEOUT = 10000 +class KafkaReceiver[T: ClassManifest, D <: Decoder[_]: Manifest]( + kafkaParams: Map[String, String], + topics: Map[String, Int], + storageLevel: StorageLevel + ) extends NetworkReceiver[Any] { // Handles pushing data into the BlockManager lazy protected val blockGenerator = new BlockGenerator(storageLevel) - // Keeps track of the current offsets. Maps from (broker, topic, group, part) -> Offset - lazy val offsets = HashMap[KafkaPartitionKey, Long]() // Connection to Kafka - var consumerConnector : ZookeeperConsumerConnector = null + var consumerConnector : ConsumerConnector = null def onStop() { blockGenerator.stop() @@ -127,74 +66,59 @@ class KafkaReceiver(host: String, port: Int, groupId: String, // In case we are using multiple Threads to handle Kafka Messages val executorPool = Executors.newFixedThreadPool(topics.values.reduce(_ + _)) - val zooKeeperEndPoint = host + ":" + port - logInfo("Starting Kafka Consumer Stream with group: " + groupId) - logInfo("Initial offsets: " + initialOffsets.toString) - - // Zookeper connection properties + logInfo("Starting Kafka Consumer Stream with group: " + kafkaParams("groupid")) + + // Kafka connection properties val props = new Properties() - props.put("zk.connect", zooKeeperEndPoint) - props.put("zk.connectiontimeout.ms", ZK_TIMEOUT.toString) - props.put("groupid", groupId) + kafkaParams.foreach(param => props.put(param._1, param._2)) // Create the connection to the cluster - logInfo("Connecting to Zookeper: " + zooKeeperEndPoint) + logInfo("Connecting to Zookeper: " + kafkaParams("zk.connect")) val consumerConfig = new ConsumerConfig(props) - consumerConnector = Consumer.create(consumerConfig).asInstanceOf[ZookeeperConsumerConnector] - logInfo("Connected to " + zooKeeperEndPoint) + consumerConnector = Consumer.create(consumerConfig) + logInfo("Connected to " + kafkaParams("zk.connect")) - // Reset the Kafka offsets in case we are recovering from a failure - resetOffsets(initialOffsets) + // When autooffset.reset is defined, it is our responsibility to try and whack the + // consumer group zk node. + if (kafkaParams.contains("autooffset.reset")) { + tryZookeeperConsumerGroupCleanup(kafkaParams("zk.connect"), kafkaParams("groupid")) + } // Create Threads for each Topic/Message Stream we are listening - val topicMessageStreams = consumerConnector.createMessageStreams(topics, new StringDecoder()) + val decoder = manifest[D].erasure.newInstance.asInstanceOf[Decoder[T]] + val topicMessageStreams = consumerConnector.createMessageStreams(topics, decoder) // Start the messages handler for each partition topicMessageStreams.values.foreach { streams => streams.foreach { stream => executorPool.submit(new MessageHandler(stream)) } } - - } - - // Overwrites the offets in Zookeper. - private def resetOffsets(offsets: Map[KafkaPartitionKey, Long]) { - offsets.foreach { case(key, offset) => - val topicDirs = new ZKGroupTopicDirs(key.groupId, key.topic) - val partitionName = key.brokerId + "-" + key.partId - updatePersistentPath(consumerConnector.zkClient, - topicDirs.consumerOffsetDir + "/" + partitionName, offset.toString) - } } // Handles Kafka Messages - private class MessageHandler(stream: KafkaStream[String]) extends Runnable { + private class MessageHandler[T: ClassManifest](stream: KafkaStream[T]) extends Runnable { def run() { logInfo("Starting MessageHandler.") - stream.takeWhile { msgAndMetadata => + for (msgAndMetadata <- stream) { blockGenerator += msgAndMetadata.message - - // Updating the offet. The key is (broker, topic, group, partition). - val key = KafkaPartitionKey(msgAndMetadata.topicInfo.brokerId, msgAndMetadata.topic, - groupId, msgAndMetadata.topicInfo.partition.partId) - val offset = msgAndMetadata.topicInfo.getConsumeOffset - offsets.put(key, offset) - // logInfo("Handled message: " + (key, offset).toString) - - // Keep on handling messages - true - } + } } } - // NOT USED - Originally intended for fault-tolerance - // class KafkaDataHandler(receiver: KafkaReceiver, storageLevel: StorageLevel) - // extends BufferingBlockCreator[Any](receiver, storageLevel) { - - // override def createBlock(blockId: String, iterator: Iterator[Any]) : Block = { - // // Creates a new Block with Kafka-specific Metadata - // new Block(blockId, iterator, KafkaInputDStreamMetadata(System.currentTimeMillis, offsets.toMap)) - // } - - // } - + // It is our responsibility to delete the consumer group when specifying autooffset.reset. This is because + // Kafka 0.7.2 only honors this param when the group is not in zookeeper. + // + // The kafka high level consumer doesn't expose setting offsets currently, this is a trick copied from Kafkas' + // ConsoleConsumer. See code related to 'autooffset.reset' when it is set to 'smallest'/'largest': + // https://github.com/apache/kafka/blob/0.7.2/core/src/main/scala/kafka/consumer/ConsoleConsumer.scala + private def tryZookeeperConsumerGroupCleanup(zkUrl: String, groupId: String) { + try { + val dir = "/consumers/" + groupId + logInfo("Cleaning up temporary zookeeper data under " + dir + ".") + val zk = new ZkClient(zkUrl, 30*1000, 30*1000, ZKStringSerializer) + zk.deleteRecursive(dir) + zk.close() + } catch { + case _ => // swallow + } + } } diff --git a/streaming/src/main/scala/spark/streaming/dstream/NetworkInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/NetworkInputDStream.scala index 8c322dd698..26805e9621 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/NetworkInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/NetworkInputDStream.scala @@ -20,7 +20,7 @@ import java.util.concurrent.ArrayBlockingQueue /** * Abstract class for defining any InputDStream that has to start a receiver on worker * nodes to receive external data. Specific implementations of NetworkInputDStream must - * define the createReceiver() function that creates the receiver object of type + * define the getReceiver() function that gets the receiver object of type * [[spark.streaming.dstream.NetworkReceiver]] that will be sent to the workers to receive * data. * @param ssc_ Streaming context that will execute this input stream @@ -34,11 +34,11 @@ abstract class NetworkInputDStream[T: ClassManifest](@transient ssc_ : Streaming val id = ssc.getNewNetworkStreamId() /** - * Creates the receiver object that will be sent to the worker nodes + * Gets the receiver object that will be sent to the worker nodes * to receive data. This method needs to defined by any specific implementation * of a NetworkInputDStream. */ - def createReceiver(): NetworkReceiver[T] + def getReceiver(): NetworkReceiver[T] // Nothing to start or stop as both taken care of by the NetworkInputTracker. def start() {} @@ -46,8 +46,15 @@ abstract class NetworkInputDStream[T: ClassManifest](@transient ssc_ : Streaming def stop() {} override def compute(validTime: Time): Option[RDD[T]] = { - val blockIds = ssc.networkInputTracker.getBlockIds(id, validTime) - Some(new BlockRDD[T](ssc.sc, blockIds)) + // If this is called for any time before the start time of the context, + // then this returns an empty RDD. This may happen when recovering from a + // master failure + if (validTime >= graph.startTime) { + val blockIds = ssc.networkInputTracker.getBlockIds(id, validTime) + Some(new BlockRDD[T](ssc.sc, blockIds)) + } else { + Some(new BlockRDD[T](ssc.sc, Array[String]())) + } } } @@ -191,7 +198,7 @@ abstract class NetworkReceiver[T: ClassManifest]() extends Serializable with Log case class Block(id: String, iterator: Iterator[T], metadata: Any = null) val clock = new SystemClock() - val blockInterval = 200L + val blockInterval = System.getProperty("spark.streaming.blockInterval", "200").toLong val blockIntervalTimer = new RecurringTimer(clock, blockInterval, updateCurrentBuffer) val blockStorageLevel = storageLevel val blocksForPushing = new ArrayBlockingQueue[Block](1000) diff --git a/streaming/src/main/scala/spark/streaming/dstream/PluggableInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/PluggableInputDStream.scala new file mode 100644 index 0000000000..3c2a81947b --- /dev/null +++ b/streaming/src/main/scala/spark/streaming/dstream/PluggableInputDStream.scala @@ -0,0 +1,13 @@ +package spark.streaming.dstream + +import spark.streaming.StreamingContext + +private[streaming] +class PluggableInputDStream[T: ClassManifest]( + @transient ssc_ : StreamingContext, + receiver: NetworkReceiver[T]) extends NetworkInputDStream[T](ssc_) { + + def getReceiver(): NetworkReceiver[T] = { + receiver + } +} diff --git a/streaming/src/main/scala/spark/streaming/dstream/QueueInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/QueueInputDStream.scala index 024bf3bea4..da224ad6f7 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/QueueInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/QueueInputDStream.scala @@ -7,6 +7,7 @@ import scala.collection.mutable.Queue import scala.collection.mutable.ArrayBuffer import spark.streaming.{Time, StreamingContext} +private[streaming] class QueueInputDStream[T: ClassManifest]( @transient ssc: StreamingContext, val queue: Queue[RDD[T]], @@ -27,7 +28,7 @@ class QueueInputDStream[T: ClassManifest]( } if (buffer.size > 0) { if (oneAtATime) { - Some(buffer.first) + Some(buffer.head) } else { Some(new UnionRDD(ssc.sc, buffer.toSeq)) } diff --git a/streaming/src/main/scala/spark/streaming/dstream/RawInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/RawInputDStream.scala index 04e6b69b7b..1b2fa56779 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/RawInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/RawInputDStream.scala @@ -25,7 +25,7 @@ class RawInputDStream[T: ClassManifest]( storageLevel: StorageLevel ) extends NetworkInputDStream[T](ssc_ ) with Logging { - def createReceiver(): NetworkReceiver[T] = { + def getReceiver(): NetworkReceiver[T] = { new RawNetworkReceiver(host, port, storageLevel).asInstanceOf[NetworkReceiver[T]] } } diff --git a/streaming/src/main/scala/spark/streaming/dstream/ReducedWindowedDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/ReducedWindowedDStream.scala index 263655039c..343b6915e7 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/ReducedWindowedDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/ReducedWindowedDStream.scala @@ -3,7 +3,7 @@ package spark.streaming.dstream import spark.streaming.StreamingContext._ import spark.RDD -import spark.rdd.CoGroupedRDD +import spark.rdd.{CoGroupedRDD, MapPartitionsRDD} import spark.Partitioner import spark.SparkContext._ import spark.storage.StorageLevel @@ -15,7 +15,8 @@ private[streaming] class ReducedWindowedDStream[K: ClassManifest, V: ClassManifest]( parent: DStream[(K, V)], reduceFunc: (V, V) => V, - invReduceFunc: (V, V) => V, + invReduceFunc: (V, V) => V, + filterFunc: Option[((K, V)) => Boolean], _windowDuration: Duration, _slideDuration: Duration, partitioner: Partitioner @@ -87,15 +88,18 @@ class ReducedWindowedDStream[K: ClassManifest, V: ClassManifest]( // // Get the RDDs of the reduced values in "old time steps" - val oldRDDs = reducedStream.slice(previousWindow.beginTime, currentWindow.beginTime - parent.slideDuration) + val oldRDDs = + reducedStream.slice(previousWindow.beginTime, currentWindow.beginTime - parent.slideDuration) logDebug("# old RDDs = " + oldRDDs.size) // Get the RDDs of the reduced values in "new time steps" - val newRDDs = reducedStream.slice(previousWindow.endTime + parent.slideDuration, currentWindow.endTime) + val newRDDs = + reducedStream.slice(previousWindow.endTime + parent.slideDuration, currentWindow.endTime) logDebug("# new RDDs = " + newRDDs.size) // Get the RDD of the reduced value of the previous window - val previousWindowRDD = getOrCompute(previousWindow.endTime).getOrElse(ssc.sc.makeRDD(Seq[(K,V)]())) + val previousWindowRDD = + getOrCompute(previousWindow.endTime).getOrElse(ssc.sc.makeRDD(Seq[(K,V)]())) // Make the list of RDDs that needs to cogrouped together for reducing their reduced values val allRDDs = new ArrayBuffer[RDD[(K, V)]]() += previousWindowRDD ++= oldRDDs ++= newRDDs @@ -114,7 +118,9 @@ class ReducedWindowedDStream[K: ClassManifest, V: ClassManifest]( // Getting reduced values "old time steps" that will be removed from current window val oldValues = (1 to numOldValues).map(i => seqOfValues(i)).filter(!_.isEmpty).map(_.head) // Getting reduced values "new time steps" - val newValues = (1 to numNewValues).map(i => seqOfValues(numOldValues + i)).filter(!_.isEmpty).map(_.head) + val newValues = + (1 to numNewValues).map(i => seqOfValues(numOldValues + i)).filter(!_.isEmpty).map(_.head) + if (seqOfValues(0).isEmpty) { // If previous window's reduce value does not exist, then at least new values should exist if (newValues.isEmpty) { @@ -140,10 +146,12 @@ class ReducedWindowedDStream[K: ClassManifest, V: ClassManifest]( val mergedValuesRDD = cogroupedRDD.asInstanceOf[RDD[(K,Seq[Seq[V]])]].mapValues(mergeValues) - Some(mergedValuesRDD) + if (filterFunc.isDefined) { + Some(mergedValuesRDD.filter(filterFunc.get)) + } else { + Some(mergedValuesRDD) + } } - - } diff --git a/streaming/src/main/scala/spark/streaming/dstream/SocketInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/SocketInputDStream.scala index d42027092b..1408af0afa 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/SocketInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/SocketInputDStream.scala @@ -2,6 +2,7 @@ package spark.streaming.dstream import spark.streaming.StreamingContext import spark.storage.StorageLevel +import spark.util.NextIterator import java.io._ import java.net.Socket @@ -15,7 +16,7 @@ class SocketInputDStream[T: ClassManifest]( storageLevel: StorageLevel ) extends NetworkInputDStream[T](ssc_) { - def createReceiver(): NetworkReceiver[T] = { + def getReceiver(): NetworkReceiver[T] = { new SocketReceiver(host, port, bytesToObjects, storageLevel) } } @@ -59,45 +60,18 @@ object SocketReceiver { */ def bytesToLines(inputStream: InputStream): Iterator[String] = { val dataInputStream = new BufferedReader(new InputStreamReader(inputStream, "UTF-8")) - - val iterator = new Iterator[String] { - var gotNext = false - var finished = false - var nextValue: String = null - - private def getNext() { - try { - nextValue = dataInputStream.readLine() - if (nextValue == null) { - finished = true - } - } - gotNext = true - } - - override def hasNext: Boolean = { - if (!finished) { - if (!gotNext) { - getNext() - if (finished) { - dataInputStream.close() - } - } + new NextIterator[String] { + protected override def getNext() = { + val nextValue = dataInputStream.readLine() + if (nextValue == null) { + finished = true } - !finished + nextValue } - override def next(): String = { - if (finished) { - throw new NoSuchElementException("End of stream") - } - if (!gotNext) { - getNext() - } - gotNext = false - nextValue + protected override def close() { + dataInputStream.close() } } - iterator } } diff --git a/streaming/src/main/scala/spark/streaming/dstream/StateDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/StateDStream.scala index b4506c74aa..db62955036 100644 --- a/streaming/src/main/scala/spark/streaming/dstream/StateDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/StateDStream.scala @@ -48,8 +48,16 @@ class StateDStream[K: ClassManifest, V: ClassManifest, S: ClassManifest]( //logDebug("Generating state RDD for time " + validTime) return Some(stateRDD) } - case None => { // If parent RDD does not exist, then return old state RDD - return Some(prevStateRDD) + case None => { // If parent RDD does not exist + + // Re-apply the update function to the old state RDD + val updateFuncLocal = updateFunc + val finalFunc = (iterator: Iterator[(K, S)]) => { + val i = iterator.map(t => (t._1, Seq[V](), Option(t._2))) + updateFuncLocal(i) + } + val stateRDD = prevStateRDD.mapPartitions(finalFunc, preservePartitioning) + return Some(stateRDD) } } } diff --git a/examples/src/main/scala/spark/streaming/examples/twitter/TwitterInputDStream.scala b/streaming/src/main/scala/spark/streaming/dstream/TwitterInputDStream.scala index 99ed4cdc1c..c697498862 100644 --- a/examples/src/main/scala/spark/streaming/examples/twitter/TwitterInputDStream.scala +++ b/streaming/src/main/scala/spark/streaming/dstream/TwitterInputDStream.scala @@ -1,12 +1,11 @@ -package spark.streaming.examples.twitter +package spark.streaming.dstream import spark._ import spark.streaming._ -import dstream.{NetworkReceiver, NetworkInputDStream} import storage.StorageLevel + import twitter4j._ import twitter4j.auth.BasicAuthorization -import collection.JavaConversions._ /* A stream of Twitter statuses, potentially filtered by one or more keywords. * @@ -14,19 +13,21 @@ import collection.JavaConversions._ * An optional set of string filters can be used to restrict the set of tweets. The Twitter API is * such that this may return a sampled subset of all tweets during each interval. */ +private[streaming] class TwitterInputDStream( @transient ssc_ : StreamingContext, username: String, password: String, filters: Seq[String], storageLevel: StorageLevel - ) extends NetworkInputDStream[Status](ssc_) { + ) extends NetworkInputDStream[Status](ssc_) { - override def createReceiver(): NetworkReceiver[Status] = { + override def getReceiver(): NetworkReceiver[Status] = { new TwitterReceiver(username, password, filters, storageLevel) } } +private[streaming] class TwitterReceiver( username: String, password: String, @@ -50,7 +51,7 @@ class TwitterReceiver( def onTrackLimitationNotice(i: Int) {} def onScrubGeo(l: Long, l1: Long) {} def onStallWarning(stallWarning: StallWarning) {} - def onException(e: Exception) {} + def onException(e: Exception) { stopOnError(e) } }) val query: FilterQuery = new FilterQuery diff --git a/streaming/src/main/scala/spark/streaming/receivers/ActorReceiver.scala b/streaming/src/main/scala/spark/streaming/receivers/ActorReceiver.scala new file mode 100644 index 0000000000..b3201d0b28 --- /dev/null +++ b/streaming/src/main/scala/spark/streaming/receivers/ActorReceiver.scala @@ -0,0 +1,153 @@ +package spark.streaming.receivers + +import akka.actor.{ Actor, PoisonPill, Props, SupervisorStrategy } +import akka.actor.{ actorRef2Scala, ActorRef } +import akka.actor.{ PossiblyHarmful, OneForOneStrategy } + +import spark.storage.StorageLevel +import spark.streaming.dstream.NetworkReceiver + +import java.util.concurrent.atomic.AtomicInteger + +/** A helper with set of defaults for supervisor strategy **/ +object ReceiverSupervisorStrategy { + + import akka.util.duration._ + import akka.actor.SupervisorStrategy._ + + val defaultStrategy = OneForOneStrategy(maxNrOfRetries = 10, withinTimeRange = + 15 millis) { + case _: RuntimeException ⇒ Restart + case _: Exception ⇒ Escalate + } +} + +/** + * A receiver trait to be mixed in with your Actor to gain access to + * pushBlock API. + * + * @example {{{ + * class MyActor extends Actor with Receiver{ + * def receive { + * case anything :String ⇒ pushBlock(anything) + * } + * } + * //Can be plugged in actorStream as follows + * ssc.actorStream[String](Props(new MyActor),"MyActorReceiver") + * + * }}} + * + * @note An important point to note: + * Since Actor may exist outside the spark framework, It is thus user's responsibility + * to ensure the type safety, i.e parametrized type of push block and InputDStream + * should be same. + * + */ +trait Receiver { self: Actor ⇒ + def pushBlock[T: ClassManifest](iter: Iterator[T]) { + context.parent ! Data(iter) + } + + def pushBlock[T: ClassManifest](data: T) { + context.parent ! Data(data) + } + +} + +/** + * Statistics for querying the supervisor about state of workers + */ +case class Statistics(numberOfMsgs: Int, + numberOfWorkers: Int, + numberOfHiccups: Int, + otherInfo: String) + +/** Case class to receive data sent by child actors **/ +private[streaming] case class Data[T: ClassManifest](data: T) + +/** + * Provides Actors as receivers for receiving stream. + * + * As Actors can also be used to receive data from almost any stream source. + * A nice set of abstraction(s) for actors as receivers is already provided for + * a few general cases. It is thus exposed as an API where user may come with + * his own Actor to run as receiver for Spark Streaming input source. + * + * This starts a supervisor actor which starts workers and also provides + * [http://doc.akka.io/docs/akka/2.0.5/scala/fault-tolerance.html fault-tolerance]. + * + * Here's a way to start more supervisor/workers as its children. + * + * @example {{{ + * context.parent ! Props(new Supervisor) + * }}} OR {{{ + * context.parent ! Props(new Worker,"Worker") + * }}} + * + * + */ +private[streaming] class ActorReceiver[T: ClassManifest]( + props: Props, + name: String, + storageLevel: StorageLevel, + receiverSupervisorStrategy: SupervisorStrategy) + extends NetworkReceiver[T] { + + protected lazy val blocksGenerator: BlockGenerator = + new BlockGenerator(storageLevel) + + protected lazy val supervisor = env.actorSystem.actorOf(Props(new Supervisor), + "Supervisor" + streamId) + + private class Supervisor extends Actor { + + override val supervisorStrategy = receiverSupervisorStrategy + val worker = context.actorOf(props, name) + logInfo("Started receiver worker at:" + worker.path) + + val n: AtomicInteger = new AtomicInteger(0) + val hiccups: AtomicInteger = new AtomicInteger(0) + + def receive = { + + case Data(iter: Iterator[_]) ⇒ pushBlock(iter.asInstanceOf[Iterator[T]]) + + case Data(msg) ⇒ + blocksGenerator += msg.asInstanceOf[T] + n.incrementAndGet + + case props: Props ⇒ + val worker = context.actorOf(props) + logInfo("Started receiver worker at:" + worker.path) + sender ! worker + + case (props: Props, name: String) ⇒ + val worker = context.actorOf(props, name) + logInfo("Started receiver worker at:" + worker.path) + sender ! worker + + case _: PossiblyHarmful => hiccups.incrementAndGet() + + case _: Statistics ⇒ + val workers = context.children + sender ! Statistics(n.get, workers.size, hiccups.get, workers.mkString("\n")) + + } + } + + protected def pushBlock(iter: Iterator[T]) { + pushBlock("block-" + streamId + "-" + System.nanoTime(), + iter, null, storageLevel) + } + + protected def onStart() = { + blocksGenerator.start() + supervisor + logInfo("Supervision tree for receivers initialized at:" + supervisor.path) + } + + protected def onStop() = { + supervisor ! PoisonPill + } + +} diff --git a/streaming/src/main/scala/spark/streaming/receivers/ZeroMQReceiver.scala b/streaming/src/main/scala/spark/streaming/receivers/ZeroMQReceiver.scala new file mode 100644 index 0000000000..5533c3cf1e --- /dev/null +++ b/streaming/src/main/scala/spark/streaming/receivers/ZeroMQReceiver.scala @@ -0,0 +1,33 @@ +package spark.streaming.receivers + +import akka.actor.Actor +import akka.zeromq._ + +import spark.Logging + +/** + * A receiver to subscribe to ZeroMQ stream. + */ +private[streaming] class ZeroMQReceiver[T: ClassManifest](publisherUrl: String, + subscribe: Subscribe, + bytesToObjects: Seq[Seq[Byte]] ⇒ Iterator[T]) + extends Actor with Receiver with Logging { + + override def preStart() = context.system.newSocket(SocketType.Sub, Listener(self), + Connect(publisherUrl), subscribe) + + def receive: Receive = { + + case Connecting ⇒ logInfo("connecting ...") + + case m: ZMQMessage ⇒ + logDebug("Received message for:" + m.firstFrameAsString) + + //We ignore first frame for processing as it is the topic + val bytes = m.frames.tail.map(_.payload) + pushBlock(bytesToObjects(bytes)) + + case Closed ⇒ logInfo("received closed ") + + } +} diff --git a/streaming/src/main/scala/spark/streaming/util/MasterFailureTest.scala b/streaming/src/main/scala/spark/streaming/util/MasterFailureTest.scala new file mode 100644 index 0000000000..426a9b6f71 --- /dev/null +++ b/streaming/src/main/scala/spark/streaming/util/MasterFailureTest.scala @@ -0,0 +1,397 @@ +package spark.streaming.util + +import spark.{Logging, RDD} +import spark.streaming._ +import spark.streaming.dstream.ForEachDStream +import StreamingContext._ + +import scala.util.Random +import scala.collection.mutable.{SynchronizedBuffer, ArrayBuffer} + +import java.io.{File, ObjectInputStream, IOException} +import java.util.UUID + +import com.google.common.io.Files + +import org.apache.commons.io.FileUtils +import org.apache.hadoop.fs.{FileUtil, FileSystem, Path} +import org.apache.hadoop.conf.Configuration + + +private[streaming] +object MasterFailureTest extends Logging { + initLogging() + + @volatile var killed = false + @volatile var killCount = 0 + + def main(args: Array[String]) { + if (args.size < 2) { + println( + "Usage: MasterFailureTest <local/HDFS directory> <# batches> [<batch size in milliseconds>]") + System.exit(1) + } + val directory = args(0) + val numBatches = args(1).toInt + val batchDuration = if (args.size > 2) Milliseconds(args(2).toInt) else Seconds(1) + + println("\n\n========================= MAP TEST =========================\n\n") + testMap(directory, numBatches, batchDuration) + + println("\n\n================= UPDATE-STATE-BY-KEY TEST =================\n\n") + testUpdateStateByKey(directory, numBatches, batchDuration) + + println("\n\nSUCCESS\n\n") + } + + def testMap(directory: String, numBatches: Int, batchDuration: Duration) { + // Input: time=1 ==> [ 1 ] , time=2 ==> [ 2 ] , time=3 ==> [ 3 ] , ... + val input = (1 to numBatches).map(_.toString).toSeq + // Expected output: time=1 ==> [ 1 ] , time=2 ==> [ 2 ] , time=3 ==> [ 3 ] , ... + val expectedOutput = (1 to numBatches) + + val operation = (st: DStream[String]) => st.map(_.toInt) + + // Run streaming operation with multiple master failures + val output = testOperation(directory, batchDuration, input, operation, expectedOutput) + + logInfo("Expected output, size = " + expectedOutput.size) + logInfo(expectedOutput.mkString("[", ",", "]")) + logInfo("Output, size = " + output.size) + logInfo(output.mkString("[", ",", "]")) + + // Verify whether all the values of the expected output is present + // in the output + assert(output.distinct.toSet == expectedOutput.toSet) + } + + + def testUpdateStateByKey(directory: String, numBatches: Int, batchDuration: Duration) { + // Input: time=1 ==> [ a ] , time=2 ==> [ a, a ] , time=3 ==> [ a, a, a ] , ... + val input = (1 to numBatches).map(i => (1 to i).map(_ => "a").mkString(" ")).toSeq + // Expected output: time=1 ==> [ (a, 1) ] , time=2 ==> [ (a, 3) ] , time=3 ==> [ (a,6) ] , ... + val expectedOutput = (1L to numBatches).map(i => (1L to i).reduce(_ + _)).map(j => ("a", j)) + + val operation = (st: DStream[String]) => { + val updateFunc = (values: Seq[Long], state: Option[Long]) => { + Some(values.foldLeft(0L)(_ + _) + state.getOrElse(0L)) + } + st.flatMap(_.split(" ")) + .map(x => (x, 1L)) + .updateStateByKey[Long](updateFunc) + .checkpoint(batchDuration * 5) + } + + // Run streaming operation with multiple master failures + val output = testOperation(directory, batchDuration, input, operation, expectedOutput) + + logInfo("Expected output, size = " + expectedOutput.size + "\n" + expectedOutput) + logInfo("Output, size = " + output.size + "\n" + output) + + // Verify whether all the values in the output are among the expected output values + output.foreach(o => + assert(expectedOutput.contains(o), "Expected value " + o + " not found") + ) + + // Verify whether the last expected output value has been generated, there by + // confirming that none of the inputs have been missed + assert(output.last == expectedOutput.last) + } + + /** + * Tests stream operation with multiple master failures, and verifies whether the + * final set of output values is as expected or not. + */ + def testOperation[T: ClassManifest]( + directory: String, + batchDuration: Duration, + input: Seq[String], + operation: DStream[String] => DStream[T], + expectedOutput: Seq[T] + ): Seq[T] = { + + // Just making sure that the expected output does not have duplicates + assert(expectedOutput.distinct.toSet == expectedOutput.toSet) + + // Setup the stream computation with the given operation + val (ssc, checkpointDir, testDir) = setupStreams(directory, batchDuration, operation) + + // Start generating files in the a different thread + val fileGeneratingThread = new FileGeneratingThread(input, testDir, batchDuration.milliseconds) + fileGeneratingThread.start() + + // Run the streams and repeatedly kill it until the last expected output + // has been generated, or until it has run for twice the expected time + val lastExpectedOutput = expectedOutput.last + val maxTimeToRun = expectedOutput.size * batchDuration.milliseconds * 2 + val mergedOutput = runStreams(ssc, lastExpectedOutput, maxTimeToRun) + + // Delete directories + fileGeneratingThread.join() + val fs = checkpointDir.getFileSystem(new Configuration()) + fs.delete(checkpointDir, true) + fs.delete(testDir, true) + logInfo("Finished test after " + killCount + " failures") + mergedOutput + } + + /** + * Sets up the stream computation with the given operation, directory (local or HDFS), + * and batch duration. Returns the streaming context and the directory to which + * files should be written for testing. + */ + private def setupStreams[T: ClassManifest]( + directory: String, + batchDuration: Duration, + operation: DStream[String] => DStream[T] + ): (StreamingContext, Path, Path) = { + // Reset all state + reset() + + // Create the directories for this test + val uuid = UUID.randomUUID().toString + val rootDir = new Path(directory, uuid) + val fs = rootDir.getFileSystem(new Configuration()) + val checkpointDir = new Path(rootDir, "checkpoint") + val testDir = new Path(rootDir, "test") + fs.mkdirs(checkpointDir) + fs.mkdirs(testDir) + + // Setup the streaming computation with the given operation + System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") + var ssc = new StreamingContext("local[4]", "MasterFailureTest", batchDuration, null, Nil, Map()) + ssc.checkpoint(checkpointDir.toString) + val inputStream = ssc.textFileStream(testDir.toString) + val operatedStream = operation(inputStream) + val outputStream = new TestOutputStream(operatedStream) + ssc.registerOutputStream(outputStream) + (ssc, checkpointDir, testDir) + } + + + /** + * Repeatedly starts and kills the streaming context until timed out or + * the last expected output is generated. Finally, return + */ + private def runStreams[T: ClassManifest]( + ssc_ : StreamingContext, + lastExpectedOutput: T, + maxTimeToRun: Long + ): Seq[T] = { + + var ssc = ssc_ + var totalTimeRan = 0L + var isLastOutputGenerated = false + var isTimedOut = false + val mergedOutput = new ArrayBuffer[T]() + val checkpointDir = ssc.checkpointDir + var batchDuration = ssc.graph.batchDuration + + while(!isLastOutputGenerated && !isTimedOut) { + // Get the output buffer + val outputBuffer = ssc.graph.getOutputStreams.head.asInstanceOf[TestOutputStream[T]].output + def output = outputBuffer.flatMap(x => x) + + // Start the thread to kill the streaming after some time + killed = false + val killingThread = new KillingThread(ssc, batchDuration.milliseconds * 10) + killingThread.start() + + var timeRan = 0L + try { + // Start the streaming computation and let it run while ... + // (i) StreamingContext has not been shut down yet + // (ii) The last expected output has not been generated yet + // (iii) Its not timed out yet + System.clearProperty("spark.streaming.clock") + System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") + ssc.start() + val startTime = System.currentTimeMillis() + while (!killed && !isLastOutputGenerated && !isTimedOut) { + Thread.sleep(100) + timeRan = System.currentTimeMillis() - startTime + isLastOutputGenerated = (!output.isEmpty && output.last == lastExpectedOutput) + isTimedOut = (timeRan + totalTimeRan > maxTimeToRun) + } + } catch { + case e: Exception => logError("Error running streaming context", e) + } + if (killingThread.isAlive) killingThread.interrupt() + ssc.stop() + + logInfo("Has been killed = " + killed) + logInfo("Is last output generated = " + isLastOutputGenerated) + logInfo("Is timed out = " + isTimedOut) + + // Verify whether the output of each batch has only one element or no element + // and then merge the new output with all the earlier output + mergedOutput ++= output + totalTimeRan += timeRan + logInfo("New output = " + output) + logInfo("Merged output = " + mergedOutput) + logInfo("Time ran = " + timeRan) + logInfo("Total time ran = " + totalTimeRan) + + if (!isLastOutputGenerated && !isTimedOut) { + val sleepTime = Random.nextInt(batchDuration.milliseconds.toInt * 10) + logInfo( + "\n-------------------------------------------\n" + + " Restarting stream computation in " + sleepTime + " ms " + + "\n-------------------------------------------\n" + ) + Thread.sleep(sleepTime) + // Recreate the streaming context from checkpoint + ssc = new StreamingContext(checkpointDir) + } + } + mergedOutput + } + + /** + * Verifies the output value are the same as expected. Since failures can lead to + * a batch being processed twice, a batches output may appear more than once + * consecutively. To avoid getting confused with those, we eliminate consecutive + * duplicate batch outputs of values from the `output`. As a result, the + * expected output should not have consecutive batches with the same values as output. + */ + private def verifyOutput[T: ClassManifest](output: Seq[T], expectedOutput: Seq[T]) { + // Verify whether expected outputs do not consecutive batches with same output + for (i <- 0 until expectedOutput.size - 1) { + assert(expectedOutput(i) != expectedOutput(i+1), + "Expected output has consecutive duplicate sequence of values") + } + + // Log the output + println("Expected output, size = " + expectedOutput.size) + println(expectedOutput.mkString("[", ",", "]")) + println("Output, size = " + output.size) + println(output.mkString("[", ",", "]")) + + // Match the output with the expected output + output.foreach(o => + assert(expectedOutput.contains(o), "Expected value " + o + " not found") + ) + } + + /** Resets counter to prepare for the test */ + private def reset() { + killed = false + killCount = 0 + } +} + +/** + * This is a output stream just for testing. All the output is collected into a + * ArrayBuffer. This buffer is wiped clean on being restored from checkpoint. + */ +private[streaming] +class TestOutputStream[T: ClassManifest]( + parent: DStream[T], + val output: ArrayBuffer[Seq[T]] = new ArrayBuffer[Seq[T]] with SynchronizedBuffer[Seq[T]] + ) extends ForEachDStream[T]( + parent, + (rdd: RDD[T], t: Time) => { + val collected = rdd.collect() + output += collected + } + ) { + + // This is to clear the output buffer every it is read from a checkpoint + @throws(classOf[IOException]) + private def readObject(ois: ObjectInputStream) { + ois.defaultReadObject() + output.clear() + } +} + + +/** + * Thread to kill streaming context after a random period of time. + */ +private[streaming] +class KillingThread(ssc: StreamingContext, maxKillWaitTime: Long) extends Thread with Logging { + initLogging() + + override def run() { + try { + // If it is the first killing, then allow the first checkpoint to be created + var minKillWaitTime = if (MasterFailureTest.killCount == 0) 5000 else 2000 + val killWaitTime = minKillWaitTime + math.abs(Random.nextLong % maxKillWaitTime) + logInfo("Kill wait time = " + killWaitTime) + Thread.sleep(killWaitTime) + logInfo( + "\n---------------------------------------\n" + + "Killing streaming context after " + killWaitTime + " ms" + + "\n---------------------------------------\n" + ) + if (ssc != null) { + ssc.stop() + MasterFailureTest.killed = true + MasterFailureTest.killCount += 1 + } + logInfo("Killing thread finished normally") + } catch { + case ie: InterruptedException => logInfo("Killing thread interrupted") + case e: Exception => logWarning("Exception in killing thread", e) + } + + } +} + + +/** + * Thread to generate input files periodically with the desired text. + */ +private[streaming] +class FileGeneratingThread(input: Seq[String], testDir: Path, interval: Long) + extends Thread with Logging { + initLogging() + + override def run() { + val localTestDir = Files.createTempDir() + var fs = testDir.getFileSystem(new Configuration()) + val maxTries = 3 + try { + Thread.sleep(5000) // To make sure that all the streaming context has been set up + for (i <- 0 until input.size) { + // Write the data to a local file and then move it to the target test directory + val localFile = new File(localTestDir, (i+1).toString) + val hadoopFile = new Path(testDir, (i+1).toString) + val tempHadoopFile = new Path(testDir, ".tmp_" + (i+1).toString) + FileUtils.writeStringToFile(localFile, input(i).toString + "\n") + var tries = 0 + var done = false + while (!done && tries < maxTries) { + tries += 1 + try { + // fs.copyFromLocalFile(new Path(localFile.toString), hadoopFile) + fs.copyFromLocalFile(new Path(localFile.toString), tempHadoopFile) + fs.rename(tempHadoopFile, hadoopFile) + done = true + } catch { + case ioe: IOException => { + fs = testDir.getFileSystem(new Configuration()) + logWarning("Attempt " + tries + " at generating file " + hadoopFile + " failed.", ioe) + } + } + } + if (!done) + logError("Could not generate file " + hadoopFile) + else + logInfo("Generated file " + hadoopFile + " at " + System.currentTimeMillis) + Thread.sleep(interval) + localFile.delete() + } + logInfo("File generating thread finished normally") + } catch { + case ie: InterruptedException => logInfo("File generating thread interrupted") + case e: Exception => logWarning("File generating in killing thread", e) + } finally { + fs.close() + } + } +} + + diff --git a/streaming/src/main/scala/spark/streaming/util/RawTextSender.scala b/streaming/src/main/scala/spark/streaming/util/RawTextSender.scala index d8b987ec86..bd0b0e74c1 100644 --- a/streaming/src/main/scala/spark/streaming/util/RawTextSender.scala +++ b/streaming/src/main/scala/spark/streaming/util/RawTextSender.scala @@ -5,7 +5,7 @@ import spark.util.{RateLimitedOutputStream, IntParam} import java.net.ServerSocket import spark.{Logging, KryoSerializer} import it.unimi.dsi.fastutil.io.FastByteArrayOutputStream -import io.Source +import scala.io.Source import java.io.IOException /** diff --git a/streaming/src/main/scala/spark/streaming/util/RecurringTimer.scala b/streaming/src/main/scala/spark/streaming/util/RecurringTimer.scala index db715cc295..8e10276deb 100644 --- a/streaming/src/main/scala/spark/streaming/util/RecurringTimer.scala +++ b/streaming/src/main/scala/spark/streaming/util/RecurringTimer.scala @@ -3,9 +3,9 @@ package spark.streaming.util private[streaming] class RecurringTimer(val clock: Clock, val period: Long, val callback: (Long) => Unit) { - val minPollTime = 25L + private val minPollTime = 25L - val pollTime = { + private val pollTime = { if (period / 10.0 > minPollTime) { (period / 10.0).toLong } else { @@ -13,11 +13,20 @@ class RecurringTimer(val clock: Clock, val period: Long, val callback: (Long) => } } - val thread = new Thread() { + private val thread = new Thread() { override def run() { loop } } - var nextTime = 0L + private var nextTime = 0L + + def getStartTime(): Long = { + (math.floor(clock.currentTime.toDouble / period) + 1).toLong * period + } + + def getRestartTime(originalStartTime: Long): Long = { + val gap = clock.currentTime - originalStartTime + (math.floor(gap.toDouble / period).toLong + 1) * period + originalStartTime + } def start(startTime: Long): Long = { nextTime = startTime @@ -26,21 +35,14 @@ class RecurringTimer(val clock: Clock, val period: Long, val callback: (Long) => } def start(): Long = { - val startTime = (math.floor(clock.currentTime.toDouble / period) + 1).toLong * period - start(startTime) + start(getStartTime()) } - def restart(originalStartTime: Long): Long = { - val gap = clock.currentTime - originalStartTime - val newStartTime = (math.floor(gap.toDouble / period).toLong + 1) * period + originalStartTime - start(newStartTime) - } - - def stop() { + def stop() { thread.interrupt() } - def loop() { + private def loop() { try { while (true) { clock.waitTillTime(nextTime) diff --git a/streaming/src/test/java/spark/streaming/JavaAPISuite.java b/streaming/src/test/java/spark/streaming/JavaAPISuite.java index 7b385f609d..e5fdbe1b7a 100644 --- a/streaming/src/test/java/spark/streaming/JavaAPISuite.java +++ b/streaming/src/test/java/spark/streaming/JavaAPISuite.java @@ -4,6 +4,7 @@ import com.google.common.base.Optional; import com.google.common.collect.Lists; import com.google.common.collect.Maps; import com.google.common.io.Files; +import kafka.serializer.StringDecoder; import org.apache.hadoop.mapreduce.lib.input.SequenceFileInputFormat; import org.junit.After; import org.junit.Assert; @@ -13,6 +14,8 @@ import scala.Tuple2; import spark.HashPartitioner; import spark.api.java.JavaPairRDD; import spark.api.java.JavaRDD; +import spark.api.java.JavaRDDLike; +import spark.api.java.JavaPairRDD; import spark.api.java.JavaSparkContext; import spark.api.java.function.*; import spark.storage.StorageLevel; @@ -21,11 +24,16 @@ import spark.streaming.api.java.JavaPairDStream; import spark.streaming.api.java.JavaStreamingContext; import spark.streaming.JavaTestUtils; import spark.streaming.JavaCheckpointTestUtils; -import spark.streaming.dstream.KafkaPartitionKey; +import spark.streaming.InputStreamsSuite; import java.io.*; import java.util.*; +import akka.actor.Props; +import akka.zeromq.Subscribe; + + + // The test suite itself is Serializable so that anonymous Function implementations can be // serialized, as an alternative to converting these anonymous classes to static inner classes; // see http://stackoverflow.com/questions/758570/. @@ -34,8 +42,9 @@ public class JavaAPISuite implements Serializable { @Before public void setUp() { - ssc = new JavaStreamingContext("local[2]", "test", new Duration(1000)); - ssc.checkpoint("checkpoint", new Duration(1000)); + System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock"); + ssc = new JavaStreamingContext("local[2]", "test", new Duration(1000)); + ssc.checkpoint("checkpoint"); } @After @@ -135,29 +144,6 @@ public class JavaAPISuite implements Serializable { } @Test - public void testTumble() { - List<List<Integer>> inputData = Arrays.asList( - Arrays.asList(1,2,3), - Arrays.asList(4,5,6), - Arrays.asList(7,8,9), - Arrays.asList(10,11,12), - Arrays.asList(13,14,15), - Arrays.asList(16,17,18)); - - List<List<Integer>> expected = Arrays.asList( - Arrays.asList(1,2,3,4,5,6), - Arrays.asList(7,8,9,10,11,12), - Arrays.asList(13,14,15,16,17,18)); - - JavaDStream stream = JavaTestUtils.attachTestInputStream(ssc, inputData, 1); - JavaDStream windowed = stream.tumble(new Duration(2000)); - JavaTestUtils.attachTestOutputStream(windowed); - List<List<Integer>> result = JavaTestUtils.runStreams(ssc, 6, 3); - - assertOrderInvariantEquals(expected, result); - } - - @Test public void testFilter() { List<List<String>> inputData = Arrays.asList( Arrays.asList("giants", "dodgers"), @@ -316,8 +302,9 @@ public class JavaAPISuite implements Serializable { Arrays.asList(6,7,8), Arrays.asList(9,10,11)); - JavaDStream stream = JavaTestUtils.attachTestInputStream(ssc, inputData, 1); - JavaDStream transformed = stream.transform(new Function<JavaRDD<Integer>, JavaRDD<Integer>>() { + JavaDStream<Integer> stream = JavaTestUtils.attachTestInputStream(ssc, inputData, 1); + JavaDStream<Integer> transformed = + stream.transform(new Function<JavaRDD<Integer>, JavaRDD<Integer>>() { @Override public JavaRDD<Integer> call(JavaRDD<Integer> in) throws Exception { return in.map(new Function<Integer, Integer>() { @@ -719,50 +706,73 @@ public class JavaAPISuite implements Serializable { } @Test - public void testCountByKey() { - List<List<Tuple2<String, String>>> inputData = stringStringKVStream; + public void testCountByValue() { + List<List<String>> inputData = Arrays.asList( + Arrays.asList("hello", "world"), + Arrays.asList("hello", "moon"), + Arrays.asList("hello")); List<List<Tuple2<String, Long>>> expected = Arrays.asList( - Arrays.asList( - new Tuple2<String, Long>("california", 2L), - new Tuple2<String, Long>("new york", 2L)), - Arrays.asList( - new Tuple2<String, Long>("california", 2L), - new Tuple2<String, Long>("new york", 2L))); - - JavaDStream<Tuple2<String, String>> stream = JavaTestUtils.attachTestInputStream( - ssc, inputData, 1); - JavaPairDStream<String, String> pairStream = JavaPairDStream.fromJavaDStream(stream); + Arrays.asList( + new Tuple2<String, Long>("hello", 1L), + new Tuple2<String, Long>("world", 1L)), + Arrays.asList( + new Tuple2<String, Long>("hello", 1L), + new Tuple2<String, Long>("moon", 1L)), + Arrays.asList( + new Tuple2<String, Long>("hello", 1L))); - JavaPairDStream<String, Long> counted = pairStream.countByKey(); + JavaDStream<String> stream = JavaTestUtils.attachTestInputStream(ssc, inputData, 1); + JavaPairDStream<String, Long> counted = stream.countByValue(); JavaTestUtils.attachTestOutputStream(counted); - List<List<Tuple2<String, Long>>> result = JavaTestUtils.runStreams(ssc, 2, 2); + List<List<Tuple2<String, Long>>> result = JavaTestUtils.runStreams(ssc, 3, 3); Assert.assertEquals(expected, result); } @Test public void testGroupByKeyAndWindow() { - List<List<Tuple2<String, String>>> inputData = stringStringKVStream; + List<List<Tuple2<String, Integer>>> inputData = stringIntKVStream; - List<List<Tuple2<String, List<String>>>> expected = Arrays.asList( - Arrays.asList(new Tuple2<String, List<String>>("california", Arrays.asList("dodgers", "giants")), - new Tuple2<String, List<String>>("new york", Arrays.asList("yankees", "mets"))), - Arrays.asList(new Tuple2<String, List<String>>("california", - Arrays.asList("sharks", "ducks", "dodgers", "giants")), - new Tuple2<String, List<String>>("new york", Arrays.asList("rangers", "islanders", "yankees", "mets"))), - Arrays.asList(new Tuple2<String, List<String>>("california", Arrays.asList("sharks", "ducks")), - new Tuple2<String, List<String>>("new york", Arrays.asList("rangers", "islanders")))); + List<List<Tuple2<String, List<Integer>>>> expected = Arrays.asList( + Arrays.asList( + new Tuple2<String, List<Integer>>("california", Arrays.asList(1, 3)), + new Tuple2<String, List<Integer>>("new york", Arrays.asList(1, 4)) + ), + Arrays.asList( + new Tuple2<String, List<Integer>>("california", Arrays.asList(1, 3, 5, 5)), + new Tuple2<String, List<Integer>>("new york", Arrays.asList(1, 1, 3, 4)) + ), + Arrays.asList( + new Tuple2<String, List<Integer>>("california", Arrays.asList(5, 5)), + new Tuple2<String, List<Integer>>("new york", Arrays.asList(1, 3)) + ) + ); - JavaDStream<Tuple2<String, String>> stream = JavaTestUtils.attachTestInputStream(ssc, inputData, 1); - JavaPairDStream<String, String> pairStream = JavaPairDStream.fromJavaDStream(stream); + JavaDStream<Tuple2<String, Integer>> stream = JavaTestUtils.attachTestInputStream(ssc, inputData, 1); + JavaPairDStream<String, Integer> pairStream = JavaPairDStream.fromJavaDStream(stream); - JavaPairDStream<String, List<String>> groupWindowed = + JavaPairDStream<String, List<Integer>> groupWindowed = pairStream.groupByKeyAndWindow(new Duration(2000), new Duration(1000)); JavaTestUtils.attachTestOutputStream(groupWindowed); - List<List<Tuple2<String, List<String>>>> result = JavaTestUtils.runStreams(ssc, 3, 3); + List<List<Tuple2<String, List<Integer>>>> result = JavaTestUtils.runStreams(ssc, 3, 3); - Assert.assertEquals(expected, result); + assert(result.size() == expected.size()); + for (int i = 0; i < result.size(); i++) { + assert(convert(result.get(i)).equals(convert(expected.get(i)))); + } + } + + private HashSet<Tuple2<String, HashSet<Integer>>> convert(List<Tuple2<String, List<Integer>>> listOfTuples) { + List<Tuple2<String, HashSet<Integer>>> newListOfTuples = new ArrayList<Tuple2<String, HashSet<Integer>>>(); + for (Tuple2<String, List<Integer>> tuple: listOfTuples) { + newListOfTuples.add(convert(tuple)); + } + return new HashSet<Tuple2<String, HashSet<Integer>>>(newListOfTuples); + } + + private Tuple2<String, HashSet<Integer>> convert(Tuple2<String, List<Integer>> tuple) { + return new Tuple2<String, HashSet<Integer>>(tuple._1(), new HashSet<Integer>(tuple._2())); } @Test @@ -847,26 +857,28 @@ public class JavaAPISuite implements Serializable { } @Test - public void testCountByKeyAndWindow() { - List<List<Tuple2<String, String>>> inputData = stringStringKVStream; + public void testCountByValueAndWindow() { + List<List<String>> inputData = Arrays.asList( + Arrays.asList("hello", "world"), + Arrays.asList("hello", "moon"), + Arrays.asList("hello")); List<List<Tuple2<String, Long>>> expected = Arrays.asList( Arrays.asList( - new Tuple2<String, Long>("california", 2L), - new Tuple2<String, Long>("new york", 2L)), + new Tuple2<String, Long>("hello", 1L), + new Tuple2<String, Long>("world", 1L)), Arrays.asList( - new Tuple2<String, Long>("california", 4L), - new Tuple2<String, Long>("new york", 4L)), + new Tuple2<String, Long>("hello", 2L), + new Tuple2<String, Long>("world", 1L), + new Tuple2<String, Long>("moon", 1L)), Arrays.asList( - new Tuple2<String, Long>("california", 2L), - new Tuple2<String, Long>("new york", 2L))); + new Tuple2<String, Long>("hello", 2L), + new Tuple2<String, Long>("moon", 1L))); - JavaDStream<Tuple2<String, String>> stream = JavaTestUtils.attachTestInputStream( + JavaDStream<String> stream = JavaTestUtils.attachTestInputStream( ssc, inputData, 1); - JavaPairDStream<String, String> pairStream = JavaPairDStream.fromJavaDStream(stream); - JavaPairDStream<String, Long> counted = - pairStream.countByKeyAndWindow(new Duration(2000), new Duration(1000)); + stream.countByValueAndWindow(new Duration(2000), new Duration(1000)); JavaTestUtils.attachTestOutputStream(counted); List<List<Tuple2<String, Long>>> result = JavaTestUtils.runStreams(ssc, 3, 3); @@ -918,6 +930,46 @@ public class JavaAPISuite implements Serializable { } @Test + public void testPairToNormalRDDTransform() { + List<List<Tuple2<Integer, Integer>>> inputData = Arrays.asList( + Arrays.asList( + new Tuple2<Integer, Integer>(3, 5), + new Tuple2<Integer, Integer>(1, 5), + new Tuple2<Integer, Integer>(4, 5), + new Tuple2<Integer, Integer>(2, 5)), + Arrays.asList( + new Tuple2<Integer, Integer>(2, 5), + new Tuple2<Integer, Integer>(3, 5), + new Tuple2<Integer, Integer>(4, 5), + new Tuple2<Integer, Integer>(1, 5))); + + List<List<Integer>> expected = Arrays.asList( + Arrays.asList(3,1,4,2), + Arrays.asList(2,3,4,1)); + + JavaDStream<Tuple2<Integer, Integer>> stream = JavaTestUtils.attachTestInputStream( + ssc, inputData, 1); + JavaPairDStream<Integer, Integer> pairStream = JavaPairDStream.fromJavaDStream(stream); + + JavaDStream<Integer> firstParts = pairStream.transform( + new Function<JavaPairRDD<Integer, Integer>, JavaRDD<Integer>>() { + @Override + public JavaRDD<Integer> call(JavaPairRDD<Integer, Integer> in) throws Exception { + return in.map(new Function<Tuple2<Integer, Integer>, Integer>() { + @Override + public Integer call(Tuple2<Integer, Integer> in) { + return in._1(); + } + }); + } + }); + + JavaTestUtils.attachTestOutputStream(firstParts); + List<List<Integer>> result = JavaTestUtils.runStreams(ssc, 2, 2); + + Assert.assertEquals(expected, result); + } + public void testMapValues() { List<List<Tuple2<String, String>>> inputData = stringStringKVStream; @@ -1091,9 +1143,8 @@ public class JavaAPISuite implements Serializable { Arrays.asList(1,4), Arrays.asList(8,7)); - File tempDir = Files.createTempDir(); - ssc.checkpoint(tempDir.getAbsolutePath(), new Duration(1000)); + ssc.checkpoint(tempDir.getAbsolutePath()); JavaDStream stream = JavaCheckpointTestUtils.attachTestInputStream(ssc, inputData, 1); JavaDStream letterCount = stream.map(new Function<String, Integer>() { @@ -1107,14 +1158,16 @@ public class JavaAPISuite implements Serializable { assertOrderInvariantEquals(expectedInitial, initialResult); Thread.sleep(1000); - ssc.stop(); + ssc = new JavaStreamingContext(tempDir.getAbsolutePath()); - ssc.start(); - List<List<Integer>> finalResult = JavaCheckpointTestUtils.runStreams(ssc, 2, 2); - assertOrderInvariantEquals(expectedFinal, finalResult); + // Tweak to take into consideration that the last batch before failure + // will be re-processed after recovery + List<List<Integer>> finalResult = JavaCheckpointTestUtils.runStreams(ssc, 2, 3); + assertOrderInvariantEquals(expectedFinal, finalResult.subList(1, 3)); } + /** TEST DISABLED: Pending a discussion about checkpoint() semantics with TD @Test public void testCheckpointofIndividualStream() throws InterruptedException { @@ -1150,20 +1203,24 @@ public class JavaAPISuite implements Serializable { @Test public void testKafkaStream() { HashMap<String, Integer> topics = Maps.newHashMap(); - HashMap<KafkaPartitionKey, Long> offsets = Maps.newHashMap(); - JavaDStream test1 = ssc.kafkaStream("localhost", 12345, "group", topics); - JavaDStream test2 = ssc.kafkaStream("localhost", 12345, "group", topics, offsets); - JavaDStream test3 = ssc.kafkaStream("localhost", 12345, "group", topics, offsets, + JavaDStream test1 = ssc.kafkaStream("localhost:12345", "group", topics); + JavaDStream test2 = ssc.kafkaStream("localhost:12345", "group", topics, + StorageLevel.MEMORY_AND_DISK()); + + HashMap<String, String> kafkaParams = Maps.newHashMap(); + kafkaParams.put("zk.connect","localhost:12345"); + kafkaParams.put("groupid","consumer-group"); + JavaDStream test3 = ssc.kafkaStream(String.class, StringDecoder.class, kafkaParams, topics, StorageLevel.MEMORY_AND_DISK()); } @Test - public void testNetworkTextStream() { - JavaDStream test = ssc.networkTextStream("localhost", 12345); + public void testSocketTextStream() { + JavaDStream test = ssc.socketTextStream("localhost", 12345); } @Test - public void testNetworkString() { + public void testSocketString() { class Converter extends Function<InputStream, Iterable<String>> { public Iterable<String> call(InputStream in) { BufferedReader reader = new BufferedReader(new InputStreamReader(in)); @@ -1179,7 +1236,7 @@ public class JavaAPISuite implements Serializable { } } - JavaDStream test = ssc.networkStream( + JavaDStream test = ssc.socketStream( "localhost", 12345, new Converter(), @@ -1192,13 +1249,13 @@ public class JavaAPISuite implements Serializable { } @Test - public void testRawNetworkStream() { - JavaDStream test = ssc.rawNetworkStream("localhost", 12345); + public void testRawSocketStream() { + JavaDStream test = ssc.rawSocketStream("localhost", 12345); } @Test public void testFlumeStream() { - JavaDStream test = ssc.flumeStream("localhost", 12345); + JavaDStream test = ssc.flumeStream("localhost", 12345, StorageLevel.MEMORY_ONLY()); } @Test @@ -1206,4 +1263,25 @@ public class JavaAPISuite implements Serializable { JavaPairDStream<String, String> foo = ssc.<String, String, SequenceFileInputFormat>fileStream("/tmp/foo"); } + + @Test + public void testTwitterStream() { + String[] filters = new String[] { "good", "bad", "ugly" }; + JavaDStream test = ssc.twitterStream("username", "password", filters, StorageLevel.MEMORY_ONLY()); + } + + @Test + public void testActorStream() { + JavaDStream test = ssc.actorStream((Props)null, "TestActor", StorageLevel.MEMORY_ONLY()); + } + + @Test + public void testZeroMQStream() { + JavaDStream test = ssc.zeroMQStream("url", (Subscribe) null, new Function<byte[][], Iterable<String>>() { + @Override + public Iterable<String> call(byte[][] b) throws Exception { + return null; + } + }); + } } diff --git a/streaming/src/test/java/spark/streaming/JavaTestUtils.scala b/streaming/src/test/java/spark/streaming/JavaTestUtils.scala index 56349837e5..64a7e7cbf9 100644 --- a/streaming/src/test/java/spark/streaming/JavaTestUtils.scala +++ b/streaming/src/test/java/spark/streaming/JavaTestUtils.scala @@ -31,8 +31,9 @@ trait JavaTestBase extends TestSuiteBase { * Attach a provided stream to it's associated StreamingContext as a * [[spark.streaming.TestOutputStream]]. **/ - def attachTestOutputStream[T, This <: spark.streaming.api.java.JavaDStreamLike[T,This]]( - dstream: JavaDStreamLike[T, This]) = { + def attachTestOutputStream[T, This <: spark.streaming.api.java.JavaDStreamLike[T, This, R], + R <: spark.api.java.JavaRDDLike[T, R]]( + dstream: JavaDStreamLike[T, This, R]) = { implicit val cm: ClassManifest[T] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[T]] val ostream = new TestOutputStream(dstream.dstream, @@ -57,6 +58,7 @@ trait JavaTestBase extends TestSuiteBase { } object JavaTestUtils extends JavaTestBase { + override def maxWaitTimeMillis = 20000 } diff --git a/streaming/src/test/resources/log4j.properties b/streaming/src/test/resources/log4j.properties index edfa1243fa..59c445e63f 100644 --- a/streaming/src/test/resources/log4j.properties +++ b/streaming/src/test/resources/log4j.properties @@ -1,5 +1,6 @@ # Set everything to be logged to the file streaming/target/unit-tests.log log4j.rootCategory=INFO, file +# log4j.appender.file=org.apache.log4j.FileAppender log4j.appender.file=org.apache.log4j.FileAppender log4j.appender.file.append=false log4j.appender.file.file=streaming/target/unit-tests.log diff --git a/streaming/src/test/scala/spark/streaming/BasicOperationsSuite.scala b/streaming/src/test/scala/spark/streaming/BasicOperationsSuite.scala index 4a036f0710..565089a853 100644 --- a/streaming/src/test/scala/spark/streaming/BasicOperationsSuite.scala +++ b/streaming/src/test/scala/spark/streaming/BasicOperationsSuite.scala @@ -8,9 +8,14 @@ class BasicOperationsSuite extends TestSuiteBase { override def framework() = "BasicOperationsSuite" + before { + System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") + } + after { // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") } test("map") { @@ -22,7 +27,7 @@ class BasicOperationsSuite extends TestSuiteBase { ) } - test("flatmap") { + test("flatMap") { val input = Seq(1 to 4, 5 to 8, 9 to 12) testOperation( input, @@ -86,6 +91,23 @@ class BasicOperationsSuite extends TestSuiteBase { ) } + test("count") { + testOperation( + Seq(Seq(), 1 to 1, 1 to 2, 1 to 3, 1 to 4), + (s: DStream[Int]) => s.count(), + Seq(Seq(0L), Seq(1L), Seq(2L), Seq(3L), Seq(4L)) + ) + } + + test("countByValue") { + testOperation( + Seq(1 to 1, Seq(1, 1, 1), 1 to 2, Seq(1, 1, 2, 2)), + (s: DStream[Int]) => s.countByValue(), + Seq(Seq((1, 1L)), Seq((1, 3L)), Seq((1, 1L), (2, 1L)), Seq((2, 2L), (1, 2L))), + true + ) + } + test("mapValues") { testOperation( Seq( Seq("a", "a", "b"), Seq("", ""), Seq() ), @@ -165,6 +187,71 @@ class BasicOperationsSuite extends TestSuiteBase { testOperation(inputData, updateStateOperation, outputData, true) } + test("updateStateByKey - object lifecycle") { + val inputData = + Seq( + Seq("a","b"), + null, + Seq("a","c","a"), + Seq("c"), + null, + null + ) + + val outputData = + Seq( + Seq(("a", 1), ("b", 1)), + Seq(("a", 1), ("b", 1)), + Seq(("a", 3), ("c", 1)), + Seq(("a", 3), ("c", 2)), + Seq(("c", 2)), + Seq() + ) + + val updateStateOperation = (s: DStream[String]) => { + class StateObject(var counter: Int = 0, var expireCounter: Int = 0) extends Serializable + + // updateFunc clears a state when a StateObject is seen without new values twice in a row + val updateFunc = (values: Seq[Int], state: Option[StateObject]) => { + val stateObj = state.getOrElse(new StateObject) + values.foldLeft(0)(_ + _) match { + case 0 => stateObj.expireCounter += 1 // no new values + case n => { // has new values, increment and reset expireCounter + stateObj.counter += n + stateObj.expireCounter = 0 + } + } + stateObj.expireCounter match { + case 2 => None // seen twice with no new values, give it the boot + case _ => Option(stateObj) + } + } + s.map(x => (x, 1)).updateStateByKey[StateObject](updateFunc).mapValues(_.counter) + } + + testOperation(inputData, updateStateOperation, outputData, true) + } + + test("slice") { + val ssc = new StreamingContext("local[2]", "BasicOperationSuite", Seconds(1)) + val input = Seq(Seq(1), Seq(2), Seq(3), Seq(4)) + val stream = new TestInputStream[Int](ssc, input, 2) + ssc.registerInputStream(stream) + stream.foreach(_ => {}) // Dummy output stream + ssc.start() + Thread.sleep(2000) + def getInputFromSlice(fromMillis: Long, toMillis: Long) = { + stream.slice(new Time(fromMillis), new Time(toMillis)).flatMap(_.collect()).toSet + } + + assert(getInputFromSlice(0, 1000) == Set(1)) + assert(getInputFromSlice(0, 2000) == Set(1, 2)) + assert(getInputFromSlice(1000, 2000) == Set(1, 2)) + assert(getInputFromSlice(2000, 4000) == Set(2, 3, 4)) + ssc.stop() + Thread.sleep(1000) + } + test("forgetting of RDDs - map and window operations") { assert(batchDuration === Seconds(1), "Batch duration has changed from 1 second") diff --git a/streaming/src/test/scala/spark/streaming/CheckpointSuite.scala b/streaming/src/test/scala/spark/streaming/CheckpointSuite.scala index 563a7d1458..607dea77ec 100644 --- a/streaming/src/test/scala/spark/streaming/CheckpointSuite.scala +++ b/streaming/src/test/scala/spark/streaming/CheckpointSuite.scala @@ -1,5 +1,6 @@ package spark.streaming +import dstream.FileInputDStream import spark.streaming.StreamingContext._ import java.io.File import runtime.RichInt @@ -7,9 +8,19 @@ import org.scalatest.BeforeAndAfter import org.apache.commons.io.FileUtils import collection.mutable.{SynchronizedBuffer, ArrayBuffer} import util.{Clock, ManualClock} +import scala.util.Random +import com.google.common.io.Files + +/** + * This test suites tests the checkpointing functionality of DStreams - + * the checkpointing of a DStream's RDDs as well as the checkpointing of + * the whole DStream graph. + */ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { + System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") + before { FileUtils.deleteDirectory(new File(checkpointDir)) } @@ -20,6 +31,7 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") } var ssc: StreamingContext = null @@ -28,21 +40,18 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { override def batchDuration = Milliseconds(500) - override def checkpointInterval = batchDuration - override def actuallyWait = true - test("basic stream+rdd recovery") { + test("basic rdd checkpoints + dstream graph checkpoint recovery") { assert(batchDuration === Milliseconds(500), "batchDuration for this test must be 1 second") - assert(checkpointInterval === batchDuration, "checkpointInterval for this test much be same as batchDuration") System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") val stateStreamCheckpointInterval = Seconds(1) // this ensure checkpointing occurs at least once - val firstNumBatches = (stateStreamCheckpointInterval / batchDuration) * 2 + val firstNumBatches = (stateStreamCheckpointInterval / batchDuration).toLong * 2 val secondNumBatches = firstNumBatches // Setup the streams @@ -62,10 +71,10 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { // Run till a time such that at least one RDD in the stream should have been checkpointed, // then check whether some RDD has been checkpointed or not ssc.start() - runStreamsWithRealDelay(ssc, firstNumBatches) - logInfo("Checkpoint data of state stream = \n[" + stateStream.checkpointData.rdds.mkString(",\n") + "]") - assert(!stateStream.checkpointData.rdds.isEmpty, "No checkpointed RDDs in state stream before first failure") - stateStream.checkpointData.rdds.foreach { + advanceTimeWithRealDelay(ssc, firstNumBatches) + logInfo("Checkpoint data of state stream = \n" + stateStream.checkpointData) + assert(!stateStream.checkpointData.checkpointFiles.isEmpty, "No checkpointed RDDs in state stream before first failure") + stateStream.checkpointData.checkpointFiles.foreach { case (time, data) => { val file = new File(data.toString) assert(file.exists(), "Checkpoint file '" + file +"' for time " + time + " for state stream before first failure does not exist") @@ -74,8 +83,8 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { // Run till a further time such that previous checkpoint files in the stream would be deleted // and check whether the earlier checkpoint files are deleted - val checkpointFiles = stateStream.checkpointData.rdds.map(x => new File(x._2.toString)) - runStreamsWithRealDelay(ssc, secondNumBatches) + val checkpointFiles = stateStream.checkpointData.checkpointFiles.map(x => new File(x._2)) + advanceTimeWithRealDelay(ssc, secondNumBatches) checkpointFiles.foreach(file => assert(!file.exists, "Checkpoint file '" + file + "' was not deleted")) ssc.stop() @@ -90,9 +99,9 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { // Run one batch to generate a new checkpoint file and check whether some RDD // is present in the checkpoint data or not ssc.start() - runStreamsWithRealDelay(ssc, 1) - assert(!stateStream.checkpointData.rdds.isEmpty, "No checkpointed RDDs in state stream before second failure") - stateStream.checkpointData.rdds.foreach { + advanceTimeWithRealDelay(ssc, 1) + assert(!stateStream.checkpointData.checkpointFiles.isEmpty, "No checkpointed RDDs in state stream before second failure") + stateStream.checkpointData.checkpointFiles.foreach { case (time, data) => { val file = new File(data.toString) assert(file.exists(), @@ -111,13 +120,16 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { // Adjust manual clock time as if it is being restarted after a delay System.setProperty("spark.streaming.manualClock.jump", (batchDuration.milliseconds * 7).toString) ssc.start() - runStreamsWithRealDelay(ssc, 4) + advanceTimeWithRealDelay(ssc, 4) ssc.stop() System.clearProperty("spark.streaming.manualClock.jump") ssc = null } - test("map and reduceByKey") { + // This tests whether the systm can recover from a master failure with simple + // non-stateful operations. This assumes as reliable, replayable input + // source - TestInputDStream. + test("recovery with map and reduceByKey operations") { testCheckpointedOperation( Seq( Seq("a", "a", "b"), Seq("", ""), Seq(), Seq("a", "a", "b"), Seq("", ""), Seq() ), (s: DStream[String]) => s.map(x => (x, 1)).reduceByKey(_ + _), @@ -126,7 +138,11 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { ) } - test("reduceByKeyAndWindowInv") { + + // This tests whether the ReduceWindowedDStream's RDD checkpoints works correctly such + // that the system can recover from a master failure. This assumes as reliable, + // replayable input source - TestInputDStream. + test("recovery with invertible reduceByKeyAndWindow operation") { val n = 10 val w = 4 val input = (1 to n).map(_ => Seq("a")).toSeq @@ -139,7 +155,11 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { testCheckpointedOperation(input, operation, output, 7) } - test("updateStateByKey") { + + // This tests whether the StateDStream's RDD checkpoints works correctly such + // that the system can recover from a master failure. This assumes as reliable, + // replayable input source - TestInputDStream. + test("recovery with updateStateByKey operation") { val input = (1 to 10).map(_ => Seq("a")).toSeq val output = (1 to 10).map(x => Seq(("a", x))).toSeq val operation = (st: DStream[String]) => { @@ -154,11 +174,126 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { testCheckpointedOperation(input, operation, output, 7) } + // This tests whether file input stream remembers what files were seen before + // the master failure and uses them again to process a large window operation. + // It also tests whether batches, whose processing was incomplete due to the + // failure, are re-processed or not. + test("recovery with file input stream") { + // Disable manual clock as FileInputDStream does not work with manual clock + val clockProperty = System.getProperty("spark.streaming.clock") + System.clearProperty("spark.streaming.clock") + + // Set up the streaming context and input streams + val testDir = Files.createTempDir() + var ssc = new StreamingContext(master, framework, Seconds(1)) + ssc.checkpoint(checkpointDir) + val fileStream = ssc.textFileStream(testDir.toString) + // Making value 3 take large time to process, to ensure that the master + // shuts down in the middle of processing the 3rd batch + val mappedStream = fileStream.map(s => { + val i = s.toInt + if (i == 3) Thread.sleep(2000) + i + }) + + // Reducing over a large window to ensure that recovery from master failure + // requires reprocessing of all the files seen before the failure + val reducedStream = mappedStream.reduceByWindow(_ + _, Seconds(30), Seconds(1)) + val outputBuffer = new ArrayBuffer[Seq[Int]] + var outputStream = new TestOutputStream(reducedStream, outputBuffer) + ssc.registerOutputStream(outputStream) + ssc.start() + + // Create files and advance manual clock to process them + //var clock = ssc.scheduler.clock.asInstanceOf[ManualClock] + Thread.sleep(1000) + for (i <- Seq(1, 2, 3)) { + FileUtils.writeStringToFile(new File(testDir, i.toString), i.toString + "\n") + // wait to make sure that the file is written such that it gets shown in the file listings + Thread.sleep(1000) + } + logInfo("Output = " + outputStream.output.mkString(",")) + assert(outputStream.output.size > 0, "No files processed before restart") + ssc.stop() + + // Verify whether files created have been recorded correctly or not + var fileInputDStream = ssc.graph.getInputStreams().head.asInstanceOf[FileInputDStream[_, _, _]] + def recordedFiles = fileInputDStream.files.values.flatMap(x => x) + assert(!recordedFiles.filter(_.endsWith("1")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("2")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("3")).isEmpty) + + // Create files while the master is down + for (i <- Seq(4, 5, 6)) { + FileUtils.writeStringToFile(new File(testDir, i.toString), i.toString + "\n") + Thread.sleep(1000) + } + + // Recover context from checkpoint file and verify whether the files that were + // recorded before failure were saved and successfully recovered + logInfo("*********** RESTARTING ************") + ssc = new StreamingContext(checkpointDir) + fileInputDStream = ssc.graph.getInputStreams().head.asInstanceOf[FileInputDStream[_, _, _]] + assert(!recordedFiles.filter(_.endsWith("1")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("2")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("3")).isEmpty) + + // Restart stream computation + ssc.start() + for (i <- Seq(7, 8, 9)) { + FileUtils.writeStringToFile(new File(testDir, i.toString), i.toString + "\n") + Thread.sleep(1000) + } + Thread.sleep(1000) + logInfo("Output = " + outputStream.output.mkString("[", ", ", "]")) + assert(outputStream.output.size > 0, "No files processed after restart") + ssc.stop() + + // Verify whether files created while the driver was down have been recorded or not + assert(!recordedFiles.filter(_.endsWith("4")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("5")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("6")).isEmpty) + + // Verify whether new files created after recover have been recorded or not + assert(!recordedFiles.filter(_.endsWith("7")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("8")).isEmpty) + assert(!recordedFiles.filter(_.endsWith("9")).isEmpty) + + // Append the new output to the old buffer + outputStream = ssc.graph.getOutputStreams().head.asInstanceOf[TestOutputStream[Int]] + outputBuffer ++= outputStream.output + + val expectedOutput = Seq(1, 3, 6, 10, 15, 21, 28, 36, 45) + logInfo("--------------------------------") + logInfo("output, size = " + outputBuffer.size) + outputBuffer.foreach(x => logInfo("[" + x.mkString(",") + "]")) + logInfo("expected output, size = " + expectedOutput.size) + expectedOutput.foreach(x => logInfo("[" + x + "]")) + logInfo("--------------------------------") + + // Verify whether all the elements received are as expected + val output = outputBuffer.flatMap(x => x) + assert(output.contains(6)) // To ensure that the 3rd input (i.e., 3) was processed + output.foreach(o => // To ensure all the inputs are correctly added cumulatively + assert(expectedOutput.contains(o), "Expected value " + o + " not found") + ) + // To ensure that all the inputs were received correctly + assert(expectedOutput.last === output.last) + + // Enable manual clock back again for other tests + if (clockProperty != null) + System.setProperty("spark.streaming.clock", clockProperty) + } + + /** - * Tests a streaming operation under checkpointing, by restart the operation + * Tests a streaming operation under checkpointing, by restarting the operation * from checkpoint file and verifying whether the final output is correct. * The output is assumed to have come from a reliable queue which an replay * data as required. + * + * NOTE: This takes into consideration that the last batch processed before + * master failure will be re-processed after restart/recovery. */ def testCheckpointedOperation[U: ClassManifest, V: ClassManifest]( input: Seq[Seq[U]], @@ -172,11 +307,14 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { val totalNumBatches = input.size val nextNumBatches = totalNumBatches - initialNumBatches val initialNumExpectedOutputs = initialNumBatches - val nextNumExpectedOutputs = expectedOutput.size - initialNumExpectedOutputs + val nextNumExpectedOutputs = expectedOutput.size - initialNumExpectedOutputs + 1 + // because the last batch will be processed again // Do the computation for initial number of batches, create checkpoint file and quit ssc = setupStreams[U, V](input, operation) - val output = runStreams[V](ssc, initialNumBatches, initialNumExpectedOutputs) + ssc.start() + val output = advanceTimeWithRealDelay[V](ssc, initialNumBatches) + ssc.stop() verifyOutput[V](output, expectedOutput.take(initialNumBatches), true) Thread.sleep(1000) @@ -187,16 +325,21 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { "\n-------------------------------------------\n" ) ssc = new StreamingContext(checkpointDir) - val outputNew = runStreams[V](ssc, nextNumBatches, nextNumExpectedOutputs) + System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") + ssc.start() + val outputNew = advanceTimeWithRealDelay[V](ssc, nextNumBatches) + // the first element will be re-processed data of the last batch before restart verifyOutput[V](outputNew, expectedOutput.takeRight(nextNumExpectedOutputs), true) + ssc.stop() ssc = null } /** * Advances the manual clock on the streaming scheduler by given number of batches. - * It also wait for the expected amount of time for each batch. + * It also waits for the expected amount of time for each batch. */ - def runStreamsWithRealDelay(ssc: StreamingContext, numBatches: Long) { + def advanceTimeWithRealDelay[V: ClassManifest](ssc: StreamingContext, numBatches: Long): Seq[Seq[V]] = { val clock = ssc.scheduler.clock.asInstanceOf[ManualClock] logInfo("Manual clock before advancing = " + clock.time) for (i <- 1 to numBatches.toInt) { @@ -205,6 +348,8 @@ class CheckpointSuite extends TestSuiteBase with BeforeAndAfter { } logInfo("Manual clock after advancing = " + clock.time) Thread.sleep(batchDuration.milliseconds) - } -}
\ No newline at end of file + val outputStream = ssc.graph.getOutputStreams.head.asInstanceOf[TestOutputStream[V]] + outputStream.output + } +} diff --git a/streaming/src/test/scala/spark/streaming/FailureSuite.scala b/streaming/src/test/scala/spark/streaming/FailureSuite.scala index c4cfffbfc1..a5fa7ab92d 100644 --- a/streaming/src/test/scala/spark/streaming/FailureSuite.scala +++ b/streaming/src/test/scala/spark/streaming/FailureSuite.scala @@ -1,191 +1,40 @@ package spark.streaming -import org.scalatest.BeforeAndAfter -import org.apache.commons.io.FileUtils +import spark.Logging +import spark.streaming.util.MasterFailureTest +import StreamingContext._ + +import org.scalatest.{FunSuite, BeforeAndAfter} +import com.google.common.io.Files import java.io.File -import scala.runtime.RichInt -import scala.util.Random -import spark.streaming.StreamingContext._ +import org.apache.commons.io.FileUtils import collection.mutable.ArrayBuffer -import spark.Logging + /** * This testsuite tests master failures at random times while the stream is running using * the real clock. */ -class FailureSuite extends TestSuiteBase with BeforeAndAfter { +class FailureSuite extends FunSuite with BeforeAndAfter with Logging { + + var directory = "FailureSuite" + val numBatches = 30 + val batchDuration = Milliseconds(1000) before { - FileUtils.deleteDirectory(new File(checkpointDir)) + FileUtils.deleteDirectory(new File(directory)) } after { - FailureSuite.reset() - FileUtils.deleteDirectory(new File(checkpointDir)) - - // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown - System.clearProperty("spark.driver.port") - } - - override def framework = "CheckpointSuite" - - override def batchDuration = Milliseconds(500) - - override def checkpointDir = "checkpoint" - - override def checkpointInterval = batchDuration - - test("multiple failures with updateStateByKey") { - val n = 30 - // Input: time=1 ==> [ a ] , time=2 ==> [ a, a ] , time=3 ==> [ a, a, a ] , ... - val input = (1 to n).map(i => (1 to i).map(_ =>"a").toSeq).toSeq - // Last output: [ (a, 465) ] for n=30 - val lastOutput = Seq( ("a", (1 to n).reduce(_ + _)) ) - - val operation = (st: DStream[String]) => { - val updateFunc = (values: Seq[Int], state: Option[RichInt]) => { - Some(new RichInt(values.foldLeft(0)(_ + _) + state.map(_.self).getOrElse(0))) - } - st.map(x => (x, 1)) - .updateStateByKey[RichInt](updateFunc) - .checkpoint(Seconds(2)) - .map(t => (t._1, t._2.self)) - } - - testOperationWithMultipleFailures(input, operation, lastOutput, n, n) - } - - test("multiple failures with reduceByKeyAndWindow") { - val n = 30 - val w = 100 - assert(w > n, "Window should be much larger than the number of input sets in this test") - // Input: time=1 ==> [ a ] , time=2 ==> [ a, a ] , time=3 ==> [ a, a, a ] , ... - val input = (1 to n).map(i => (1 to i).map(_ =>"a").toSeq).toSeq - // Last output: [ (a, 465) ] - val lastOutput = Seq( ("a", (1 to n).reduce(_ + _)) ) - - val operation = (st: DStream[String]) => { - st.map(x => (x, 1)) - .reduceByKeyAndWindow(_ + _, _ - _, batchDuration * w, batchDuration) - .checkpoint(Seconds(2)) - } - - testOperationWithMultipleFailures(input, operation, lastOutput, n, n) - } - - - /** - * Tests stream operation with multiple master failures, and verifies whether the - * final set of output values is as expected or not. Checking the final value is - * proof that no intermediate data was lost due to master failures. - */ - def testOperationWithMultipleFailures[U: ClassManifest, V: ClassManifest]( - input: Seq[Seq[U]], - operation: DStream[U] => DStream[V], - lastExpectedOutput: Seq[V], - numBatches: Int, - numExpectedOutput: Int - ) { - var ssc = setupStreams[U, V](input, operation) - val mergedOutput = new ArrayBuffer[Seq[V]]() - - var totalTimeRan = 0L - while(totalTimeRan <= numBatches * batchDuration.milliseconds * 2) { - new KillingThread(ssc, numBatches * batchDuration.milliseconds.toInt / 4).start() - val (output, timeRan) = runStreamsWithRealClock[V](ssc, numBatches, numExpectedOutput) - - mergedOutput ++= output - totalTimeRan += timeRan - logInfo("New output = " + output) - logInfo("Merged output = " + mergedOutput) - logInfo("Total time spent = " + totalTimeRan) - val sleepTime = Random.nextInt(numBatches * batchDuration.milliseconds.toInt / 8) - logInfo( - "\n-------------------------------------------\n" + - " Restarting stream computation in " + sleepTime + " ms " + - "\n-------------------------------------------\n" - ) - Thread.sleep(sleepTime) - FailureSuite.failed = false - ssc = new StreamingContext(checkpointDir) - } - ssc.stop() - ssc = null - - // Verify whether the last output is the expected one - val lastOutput = mergedOutput(mergedOutput.lastIndexWhere(!_.isEmpty)) - assert(lastOutput.toSet === lastExpectedOutput.toSet) - logInfo("Finished computation after " + FailureSuite.failureCount + " failures") + FileUtils.deleteDirectory(new File(directory)) } - /** - * Runs the streams set up in `ssc` on real clock until the expected max number of - */ - def runStreamsWithRealClock[V: ClassManifest]( - ssc: StreamingContext, - numBatches: Int, - maxExpectedOutput: Int - ): (Seq[Seq[V]], Long) = { - - System.clearProperty("spark.streaming.clock") - - assert(numBatches > 0, "Number of batches to run stream computation is zero") - assert(maxExpectedOutput > 0, "Max expected outputs after " + numBatches + " is zero") - logInfo("numBatches = " + numBatches + ", maxExpectedOutput = " + maxExpectedOutput) - - // Get the output buffer - val outputStream = ssc.graph.getOutputStreams.head.asInstanceOf[TestOutputStream[V]] - val output = outputStream.output - val waitTime = (batchDuration.milliseconds * (numBatches.toDouble + 0.5)).toLong - val startTime = System.currentTimeMillis() - - try { - // Start computation - ssc.start() - - // Wait until expected number of output items have been generated - while (output.size < maxExpectedOutput && System.currentTimeMillis() - startTime < waitTime && !FailureSuite.failed) { - logInfo("output.size = " + output.size + ", maxExpectedOutput = " + maxExpectedOutput) - Thread.sleep(100) - } - } catch { - case e: Exception => logInfo("Exception while running streams: " + e) - } finally { - ssc.stop() - } - val timeTaken = System.currentTimeMillis() - startTime - logInfo("" + output.size + " sets of output generated in " + timeTaken + " ms") - (output, timeTaken) + test("multiple failures with map") { + MasterFailureTest.testMap(directory, numBatches, batchDuration) } - -} - -object FailureSuite { - var failed = false - var failureCount = 0 - - def reset() { - failed = false - failureCount = 0 + test("multiple failures with updateStateByKey") { + MasterFailureTest.testUpdateStateByKey(directory, numBatches, batchDuration) } } -class KillingThread(ssc: StreamingContext, maxKillWaitTime: Int) extends Thread with Logging { - initLogging() - - override def run() { - var minKillWaitTime = if (FailureSuite.failureCount == 0) 3000 else 1000 // to allow the first checkpoint - val killWaitTime = minKillWaitTime + Random.nextInt(maxKillWaitTime) - logInfo("Kill wait time = " + killWaitTime) - Thread.sleep(killWaitTime.toLong) - logInfo( - "\n---------------------------------------\n" + - "Killing streaming context after " + killWaitTime + " ms" + - "\n---------------------------------------\n" - ) - if (ssc != null) ssc.stop() - FailureSuite.failed = true - FailureSuite.failureCount += 1 - } -} diff --git a/streaming/src/test/scala/spark/streaming/InputStreamsSuite.scala b/streaming/src/test/scala/spark/streaming/InputStreamsSuite.scala index 70ae6e3934..b024fc9dcc 100644 --- a/streaming/src/test/scala/spark/streaming/InputStreamsSuite.scala +++ b/streaming/src/test/scala/spark/streaming/InputStreamsSuite.scala @@ -1,5 +1,11 @@ package spark.streaming +import akka.actor.Actor +import akka.actor.IO +import akka.actor.IOManager +import akka.actor.Props +import akka.util.ByteString + import dstream.SparkFlumeEvent import java.net.{InetSocketAddress, SocketException, Socket, ServerSocket} import java.io.{File, BufferedWriter, OutputStreamWriter} @@ -7,6 +13,7 @@ import java.util.concurrent.{TimeUnit, ArrayBlockingQueue} import collection.mutable.{SynchronizedBuffer, ArrayBuffer} import util.ManualClock import spark.storage.StorageLevel +import spark.streaming.receivers.Receiver import spark.Logging import scala.util.Random import org.apache.commons.io.FileUtils @@ -19,40 +26,33 @@ import org.apache.avro.ipc.specific.SpecificRequestor import java.nio.ByteBuffer import collection.JavaConversions._ import java.nio.charset.Charset +import com.google.common.io.Files class InputStreamsSuite extends TestSuiteBase with BeforeAndAfter { - - System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") val testPort = 9999 - var testServer: TestServer = null - var testDir: File = null override def checkpointDir = "checkpoint" - after { - FileUtils.deleteDirectory(new File(checkpointDir)) - if (testServer != null) { - testServer.stop() - testServer = null - } - if (testDir != null && testDir.exists()) { - FileUtils.deleteDirectory(testDir) - testDir = null - } + before { + System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") + } + after { // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") } - test("network input stream") { + + test("socket input stream") { // Start the server - testServer = new TestServer(testPort) + val testServer = new TestServer() testServer.start() // Set up the streaming context and input streams val ssc = new StreamingContext(master, framework, batchDuration) - val networkStream = ssc.networkTextStream("localhost", testPort, StorageLevel.MEMORY_AND_DISK) + val networkStream = ssc.socketTextStream("localhost", testServer.port, StorageLevel.MEMORY_AND_DISK) val outputBuffer = new ArrayBuffer[Seq[String]] with SynchronizedBuffer[Seq[String ]] val outputStream = new TestOutputStream(networkStream, outputBuffer) def output = outputBuffer.flatMap(x => x) @@ -93,51 +93,11 @@ class InputStreamsSuite extends TestSuiteBase with BeforeAndAfter { } } - test("network input stream with checkpoint") { - // Start the server - testServer = new TestServer(testPort) - testServer.start() - - // Set up the streaming context and input streams - var ssc = new StreamingContext(master, framework, batchDuration) - ssc.checkpoint(checkpointDir, checkpointInterval) - val networkStream = ssc.networkTextStream("localhost", testPort, StorageLevel.MEMORY_AND_DISK) - var outputStream = new TestOutputStream(networkStream, new ArrayBuffer[Seq[String]]) - ssc.registerOutputStream(outputStream) - ssc.start() - - // Feed data to the server to send to the network receiver - var clock = ssc.scheduler.clock.asInstanceOf[ManualClock] - for (i <- Seq(1, 2, 3)) { - testServer.send(i.toString + "\n") - Thread.sleep(100) - clock.addToTime(batchDuration.milliseconds) - } - Thread.sleep(500) - assert(outputStream.output.size > 0) - ssc.stop() - - // Restart stream computation from checkpoint and feed more data to see whether - // they are being received and processed - logInfo("*********** RESTARTING ************") - ssc = new StreamingContext(checkpointDir) - ssc.start() - clock = ssc.scheduler.clock.asInstanceOf[ManualClock] - for (i <- Seq(4, 5, 6)) { - testServer.send(i.toString + "\n") - Thread.sleep(100) - clock.addToTime(batchDuration.milliseconds) - } - Thread.sleep(500) - outputStream = ssc.graph.getOutputStreams().head.asInstanceOf[TestOutputStream[String]] - assert(outputStream.output.size > 0) - ssc.stop() - } test("flume input stream") { // Set up the streaming context and input streams val ssc = new StreamingContext(master, framework, batchDuration) - val flumeStream = ssc.flumeStream("localhost", 33333, StorageLevel.MEMORY_AND_DISK) + val flumeStream = ssc.flumeStream("localhost", testPort, StorageLevel.MEMORY_AND_DISK) val outputBuffer = new ArrayBuffer[Seq[SparkFlumeEvent]] with SynchronizedBuffer[Seq[SparkFlumeEvent]] val outputStream = new TestOutputStream(flumeStream, outputBuffer) @@ -146,8 +106,8 @@ class InputStreamsSuite extends TestSuiteBase with BeforeAndAfter { val clock = ssc.scheduler.clock.asInstanceOf[ManualClock] val input = Seq(1, 2, 3, 4, 5) - - val transceiver = new NettyTransceiver(new InetSocketAddress("localhost", 33333)); + Thread.sleep(1000) + val transceiver = new NettyTransceiver(new InetSocketAddress("localhost", testPort)); val client = SpecificRequestor.getClient( classOf[AvroSourceProtocol], transceiver); @@ -182,42 +142,33 @@ class InputStreamsSuite extends TestSuiteBase with BeforeAndAfter { } } - test("file input stream") { - // Create a temporary directory - testDir = { - var temp = File.createTempFile(".temp.", Random.nextInt().toString) - temp.delete() - temp.mkdirs() - logInfo("Created temp dir " + temp) - temp - } + test("file input stream") { + // Disable manual clock as FileInputDStream does not work with manual clock + System.clearProperty("spark.streaming.clock") // Set up the streaming context and input streams + val testDir = Files.createTempDir() val ssc = new StreamingContext(master, framework, batchDuration) - val filestream = ssc.textFileStream(testDir.toString) + val fileStream = ssc.textFileStream(testDir.toString) val outputBuffer = new ArrayBuffer[Seq[String]] with SynchronizedBuffer[Seq[String]] def output = outputBuffer.flatMap(x => x) - val outputStream = new TestOutputStream(filestream, outputBuffer) + val outputStream = new TestOutputStream(fileStream, outputBuffer) ssc.registerOutputStream(outputStream) ssc.start() // Create files in the temporary directory so that Spark Streaming can read data from it - val clock = ssc.scheduler.clock.asInstanceOf[ManualClock] val input = Seq(1, 2, 3, 4, 5) val expectedOutput = input.map(_.toString) Thread.sleep(1000) for (i <- 0 until input.size) { - FileUtils.writeStringToFile(new File(testDir, i.toString), input(i).toString + "\n") - Thread.sleep(500) - clock.addToTime(batchDuration.milliseconds) - //Thread.sleep(100) + val file = new File(testDir, i.toString) + FileUtils.writeStringToFile(file, input(i).toString + "\n") + logInfo("Created file " + file) + Thread.sleep(batchDuration.milliseconds) + Thread.sleep(1000) } val startTime = System.currentTimeMillis() - /*while (output.size < expectedOutput.size && System.currentTimeMillis() - startTime < maxWaitTimeMillis) { - logInfo("output.size = " + output.size + ", expectedOutput.size = " + expectedOutput.size) - Thread.sleep(100) - }*/ Thread.sleep(1000) val timeTaken = System.currentTimeMillis() - startTime assert(timeTaken < maxWaitTimeMillis, "Operation timed out after " + timeTaken + " ms") @@ -226,80 +177,92 @@ class InputStreamsSuite extends TestSuiteBase with BeforeAndAfter { // Verify whether data received by Spark Streaming was as expected logInfo("--------------------------------") - logInfo("output.size = " + outputBuffer.size) - logInfo("output") + logInfo("output, size = " + outputBuffer.size) outputBuffer.foreach(x => logInfo("[" + x.mkString(",") + "]")) - logInfo("expected output.size = " + expectedOutput.size) - logInfo("expected output") + logInfo("expected output, size = " + expectedOutput.size) expectedOutput.foreach(x => logInfo("[" + x.mkString(",") + "]")) logInfo("--------------------------------") // Verify whether all the elements received are as expected // (whether the elements were received one in each interval is not verified) - assert(output.size === expectedOutput.size) - for (i <- 0 until output.size) { - assert(output(i).size === 1) - assert(output(i).head.toString === expectedOutput(i)) - } + assert(output.toList === expectedOutput.toList) + + FileUtils.deleteDirectory(testDir) + + // Enable manual clock back again for other tests + System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") } - test("file input stream with checkpoint") { - // Create a temporary directory - testDir = { - var temp = File.createTempFile(".temp.", Random.nextInt().toString) - temp.delete() - temp.mkdirs() - logInfo("Created temp dir " + temp) - temp - } + + test("actor input stream") { + // Start the server + val testServer = new TestServer() + val port = testServer.port + testServer.start() // Set up the streaming context and input streams - var ssc = new StreamingContext(master, framework, batchDuration) - ssc.checkpoint(checkpointDir, checkpointInterval) - val filestream = ssc.textFileStream(testDir.toString) - var outputStream = new TestOutputStream(filestream, new ArrayBuffer[Seq[String]]) + val ssc = new StreamingContext(master, framework, batchDuration) + val networkStream = ssc.actorStream[String](Props(new TestActor(port)), "TestActor", + StorageLevel.MEMORY_AND_DISK) //Had to pass the local value of port to prevent from closing over entire scope + val outputBuffer = new ArrayBuffer[Seq[String]] with SynchronizedBuffer[Seq[String]] + val outputStream = new TestOutputStream(networkStream, outputBuffer) + def output = outputBuffer.flatMap(x => x) ssc.registerOutputStream(outputStream) ssc.start() - // Create files and advance manual clock to process them - var clock = ssc.scheduler.clock.asInstanceOf[ManualClock] + // Feed data to the server to send to the network receiver + val clock = ssc.scheduler.clock.asInstanceOf[ManualClock] + val input = 1 to 9 + val expectedOutput = input.map(x => x.toString) Thread.sleep(1000) - for (i <- Seq(1, 2, 3)) { - FileUtils.writeStringToFile(new File(testDir, i.toString), i.toString + "\n") - Thread.sleep(100) + for (i <- 0 until input.size) { + testServer.send(input(i).toString) + Thread.sleep(500) clock.addToTime(batchDuration.milliseconds) } - Thread.sleep(500) - logInfo("Output = " + outputStream.output.mkString(",")) - assert(outputStream.output.size > 0) + Thread.sleep(1000) + logInfo("Stopping server") + testServer.stop() + logInfo("Stopping context") ssc.stop() - // Restart stream computation from checkpoint and create more files to see whether - // they are being processed - logInfo("*********** RESTARTING ************") - ssc = new StreamingContext(checkpointDir) - ssc.start() - clock = ssc.scheduler.clock.asInstanceOf[ManualClock] - Thread.sleep(500) - for (i <- Seq(4, 5, 6)) { - FileUtils.writeStringToFile(new File(testDir, i.toString), i.toString + "\n") - Thread.sleep(100) - clock.addToTime(batchDuration.milliseconds) + // Verify whether data received was as expected + logInfo("--------------------------------") + logInfo("output.size = " + outputBuffer.size) + logInfo("output") + outputBuffer.foreach(x => logInfo("[" + x.mkString(",") + "]")) + logInfo("expected output.size = " + expectedOutput.size) + logInfo("expected output") + expectedOutput.foreach(x => logInfo("[" + x.mkString(",") + "]")) + logInfo("--------------------------------") + + // Verify whether all the elements received are as expected + // (whether the elements were received one in each interval is not verified) + assert(output.size === expectedOutput.size) + for (i <- 0 until output.size) { + assert(output(i) === expectedOutput(i)) } - Thread.sleep(500) - outputStream = ssc.graph.getOutputStreams().head.asInstanceOf[TestOutputStream[String]] - logInfo("Output = " + outputStream.output.mkString(",")) - assert(outputStream.output.size > 0) - ssc.stop() + } + + test("kafka input stream") { + val ssc = new StreamingContext(master, framework, batchDuration) + val topics = Map("my-topic" -> 1) + val test1 = ssc.kafkaStream("localhost:12345", "group", topics) + val test2 = ssc.kafkaStream("localhost:12345", "group", topics, StorageLevel.MEMORY_AND_DISK) + + // Test specifying decoder + val kafkaParams = Map("zk.connect"->"localhost:12345","groupid"->"consumer-group") + val test3 = ssc.kafkaStream[String, kafka.serializer.StringDecoder](kafkaParams, topics, StorageLevel.MEMORY_AND_DISK) } } -class TestServer(port: Int) extends Logging { +/** This is server to test the network input stream */ +class TestServer() extends Logging { val queue = new ArrayBlockingQueue[String](100) - val serverSocket = new ServerSocket(port) + val serverSocket = new ServerSocket(0) val servingThread = new Thread() { override def run() { @@ -341,11 +304,13 @@ class TestServer(port: Int) extends Logging { def send(msg: String) { queue.add(msg) } def stop() { servingThread.interrupt() } + + def port = serverSocket.getLocalPort } object TestServer { def main(args: Array[String]) { - val s = new TestServer(9999) + val s = new TestServer() s.start() while(true) { Thread.sleep(1000) @@ -353,3 +318,15 @@ object TestServer { } } } + +class TestActor(port: Int) extends Actor with Receiver { + + def bytesToString(byteString: ByteString) = byteString.utf8String + + override def preStart = IOManager(context.system).connect(new InetSocketAddress(port)) + + def receive = { + case IO.Read(socket, bytes) => + pushBlock(bytesToString(bytes)) + } +} diff --git a/streaming/src/test/scala/spark/streaming/TestSuiteBase.scala b/streaming/src/test/scala/spark/streaming/TestSuiteBase.scala index 49129f3964..ad6aa79d10 100644 --- a/streaming/src/test/scala/spark/streaming/TestSuiteBase.scala +++ b/streaming/src/test/scala/spark/streaming/TestSuiteBase.scala @@ -28,6 +28,11 @@ class TestInputStream[T: ClassManifest](ssc_ : StreamingContext, input: Seq[Seq[ logInfo("Computing RDD for time " + validTime) val index = ((validTime - zeroTime) / slideDuration - 1).toInt val selectedInput = if (index < input.size) input(index) else Seq[T]() + + // lets us test cases where RDDs are not created + if (selectedInput == null) + return None + val rdd = ssc.sc.makeRDD(selectedInput, numPartitions) logInfo("Created RDD " + rdd.id + " with " + selectedInput) Some(rdd) @@ -58,20 +63,25 @@ class TestOutputStream[T: ClassManifest](parent: DStream[T], val output: ArrayBu */ trait TestSuiteBase extends FunSuite with BeforeAndAfter with Logging { + // Name of the framework for Spark context def framework = "TestSuiteBase" + // Master for Spark context def master = "local[2]" + // Batch duration def batchDuration = Seconds(1) + // Directory where the checkpoint data will be saved def checkpointDir = "checkpoint" - def checkpointInterval = batchDuration - + // Number of partitions of the input parallel collections created for testing def numInputPartitions = 2 + // Maximum time to wait before the test times out def maxWaitTimeMillis = 10000 + // Whether to actually wait in real time before changing manual clock def actuallyWait = false /** @@ -86,7 +96,7 @@ trait TestSuiteBase extends FunSuite with BeforeAndAfter with Logging { // Create StreamingContext val ssc = new StreamingContext(master, framework, batchDuration) if (checkpointDir != null) { - ssc.checkpoint(checkpointDir, checkpointInterval) + ssc.checkpoint(checkpointDir) } // Setup the stream computation @@ -111,7 +121,7 @@ trait TestSuiteBase extends FunSuite with BeforeAndAfter with Logging { // Create StreamingContext val ssc = new StreamingContext(master, framework, batchDuration) if (checkpointDir != null) { - ssc.checkpoint(checkpointDir, checkpointInterval) + ssc.checkpoint(checkpointDir) } // Setup the stream computation @@ -135,9 +145,6 @@ trait TestSuiteBase extends FunSuite with BeforeAndAfter with Logging { numBatches: Int, numExpectedOutput: Int ): Seq[Seq[V]] = { - - System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") - assert(numBatches > 0, "Number of batches to run stream computation is zero") assert(numExpectedOutput > 0, "Number of expected outputs after " + numBatches + " is zero") logInfo("numBatches = " + numBatches + ", numExpectedOutput = " + numExpectedOutput) @@ -181,7 +188,6 @@ trait TestSuiteBase extends FunSuite with BeforeAndAfter with Logging { } finally { ssc.stop() } - output } diff --git a/streaming/src/test/scala/spark/streaming/WindowOperationsSuite.scala b/streaming/src/test/scala/spark/streaming/WindowOperationsSuite.scala index cd9608df53..80d827706f 100644 --- a/streaming/src/test/scala/spark/streaming/WindowOperationsSuite.scala +++ b/streaming/src/test/scala/spark/streaming/WindowOperationsSuite.scala @@ -5,6 +5,8 @@ import collection.mutable.ArrayBuffer class WindowOperationsSuite extends TestSuiteBase { + System.setProperty("spark.streaming.clock", "spark.streaming.util.ManualClock") + override def framework = "WindowOperationsSuite" override def maxWaitTimeMillis = 20000 @@ -14,6 +16,7 @@ class WindowOperationsSuite extends TestSuiteBase { after { // To avoid Akka rebinding to the same port, since it doesn't unbind immediately on shutdown System.clearProperty("spark.driver.port") + System.clearProperty("spark.hostPort") } val largerSlideInput = Seq( @@ -82,12 +85,9 @@ class WindowOperationsSuite extends TestSuiteBase { ) /* - The output of the reduceByKeyAndWindow with inverse reduce function is - different from the naive reduceByKeyAndWindow. Even if the count of a - particular key is 0, the key does not get eliminated from the RDDs of - ReducedWindowedDStream. This causes the number of keys in these RDDs to - increase forever. A more generalized version that allows elimination of - keys should be considered. + The output of the reduceByKeyAndWindow with inverse function but without a filter + function will be different from the naive reduceByKeyAndWindow, as no keys get + eliminated from the ReducedWindowedDStream even if the value of a key becomes 0. */ val bigReduceInvOutput = Seq( @@ -175,31 +175,31 @@ class WindowOperationsSuite extends TestSuiteBase { // Testing reduceByKeyAndWindow (with invertible reduce function) - testReduceByKeyAndWindowInv( + testReduceByKeyAndWindowWithInverse( "basic reduction", Seq(Seq(("a", 1), ("a", 3)) ), Seq(Seq(("a", 4)) ) ) - testReduceByKeyAndWindowInv( + testReduceByKeyAndWindowWithInverse( "key already in window and new value added into window", Seq( Seq(("a", 1)), Seq(("a", 1)) ), Seq( Seq(("a", 1)), Seq(("a", 2)) ) ) - testReduceByKeyAndWindowInv( + testReduceByKeyAndWindowWithInverse( "new key added into window", Seq( Seq(("a", 1)), Seq(("a", 1), ("b", 1)) ), Seq( Seq(("a", 1)), Seq(("a", 2), ("b", 1)) ) ) - testReduceByKeyAndWindowInv( + testReduceByKeyAndWindowWithInverse( "key removed from window", Seq( Seq(("a", 1)), Seq(("a", 1)), Seq(), Seq() ), Seq( Seq(("a", 1)), Seq(("a", 2)), Seq(("a", 1)), Seq(("a", 0)) ) ) - testReduceByKeyAndWindowInv( + testReduceByKeyAndWindowWithInverse( "larger slide time", largerSlideInput, largerSlideReduceOutput, @@ -207,7 +207,9 @@ class WindowOperationsSuite extends TestSuiteBase { Seconds(2) ) - testReduceByKeyAndWindowInv("big test", bigInput, bigReduceInvOutput) + testReduceByKeyAndWindowWithInverse("big test", bigInput, bigReduceInvOutput) + + testReduceByKeyAndWindowWithFilteredInverse("big test", bigInput, bigReduceOutput) test("groupByKeyAndWindow") { val input = bigInput @@ -235,14 +237,14 @@ class WindowOperationsSuite extends TestSuiteBase { testOperation(input, operation, expectedOutput, numBatches, true) } - test("countByKeyAndWindow") { - val input = Seq(Seq(("a", 1)), Seq(("b", 1), ("b", 2)), Seq(("a", 10), ("b", 20))) + test("countByValueAndWindow") { + val input = Seq(Seq("a"), Seq("b", "b"), Seq("a", "b")) val expectedOutput = Seq( Seq(("a", 1)), Seq(("a", 1), ("b", 2)), Seq(("a", 1), ("b", 3))) val windowDuration = Seconds(2) val slideDuration = Seconds(1) val numBatches = expectedOutput.size * (slideDuration / batchDuration).toInt - val operation = (s: DStream[(String, Int)]) => { - s.countByKeyAndWindow(windowDuration, slideDuration).map(x => (x._1, x._2.toInt)) + val operation = (s: DStream[String]) => { + s.countByValueAndWindow(windowDuration, slideDuration).map(x => (x._1, x._2.toInt)) } testOperation(input, operation, expectedOutput, numBatches, true) } @@ -272,29 +274,50 @@ class WindowOperationsSuite extends TestSuiteBase { slideDuration: Duration = Seconds(1) ) { test("reduceByKeyAndWindow - " + name) { + logInfo("reduceByKeyAndWindow - " + name) val numBatches = expectedOutput.size * (slideDuration / batchDuration).toInt val operation = (s: DStream[(String, Int)]) => { - s.reduceByKeyAndWindow(_ + _, windowDuration, slideDuration).persist() + s.reduceByKeyAndWindow((x: Int, y: Int) => x + y, windowDuration, slideDuration) } testOperation(input, operation, expectedOutput, numBatches, true) } } - def testReduceByKeyAndWindowInv( + def testReduceByKeyAndWindowWithInverse( name: String, input: Seq[Seq[(String, Int)]], expectedOutput: Seq[Seq[(String, Int)]], windowDuration: Duration = Seconds(2), slideDuration: Duration = Seconds(1) ) { - test("reduceByKeyAndWindowInv - " + name) { + test("reduceByKeyAndWindow with inverse function - " + name) { + logInfo("reduceByKeyAndWindow with inverse function - " + name) val numBatches = expectedOutput.size * (slideDuration / batchDuration).toInt val operation = (s: DStream[(String, Int)]) => { s.reduceByKeyAndWindow(_ + _, _ - _, windowDuration, slideDuration) - .persist() .checkpoint(Seconds(100)) // Large value to avoid effect of RDD checkpointing } testOperation(input, operation, expectedOutput, numBatches, true) } } + + def testReduceByKeyAndWindowWithFilteredInverse( + name: String, + input: Seq[Seq[(String, Int)]], + expectedOutput: Seq[Seq[(String, Int)]], + windowDuration: Duration = Seconds(2), + slideDuration: Duration = Seconds(1) + ) { + test("reduceByKeyAndWindow with inverse and filter functions - " + name) { + logInfo("reduceByKeyAndWindow with inverse and filter functions - " + name) + val numBatches = expectedOutput.size * (slideDuration / batchDuration).toInt + val filterFunc = (p: (String, Int)) => p._2 != 0 + val operation = (s: DStream[(String, Int)]) => { + s.reduceByKeyAndWindow(_ + _, _ - _, windowDuration, slideDuration, filterFunc = filterFunc) + .persist() + .checkpoint(Seconds(100)) // Large value to avoid effect of RDD checkpointing + } + testOperation(input, operation, expectedOutput, numBatches, true) + } + } } |