diff options
113 files changed, 6587 insertions, 1198 deletions
@@ -55,7 +55,7 @@ versions without YARN, use: # Cloudera CDH 4.2.0 with MapReduce v1 $ SPARK_HADOOP_VERSION=2.0.0-mr1-cdh4.2.0 sbt/sbt assembly -For Apache Hadoop 2.x, 0.23.x, Cloudera CDH MRv2, and other Hadoop versions +For Apache Hadoop 2.0.X, 2.1.X, 0.23.x, Cloudera CDH MRv2, and other Hadoop versions with YARN, also set `SPARK_YARN=true`: # Apache Hadoop 2.0.5-alpha @@ -64,6 +64,11 @@ with YARN, also set `SPARK_YARN=true`: # Cloudera CDH 4.2.0 with MapReduce v2 $ SPARK_HADOOP_VERSION=2.0.0-cdh4.2.0 SPARK_YARN=true sbt/sbt assembly +When building for Hadoop 2.2.X and newer, you'll need to include the additional `new-yarn` profile: + + # Apache Hadoop 2.2.X and newer + $ mvn -Dyarn.version=2.2.0 -Dhadoop.version=2.2.0 -Pnew-yarn + For convenience, these variables may also be set through the `conf/spark-env.sh` file described below. diff --git a/conf/metrics.properties.template b/conf/metrics.properties.template index ae10f615d1..1c3d94e1b0 100644 --- a/conf/metrics.properties.template +++ b/conf/metrics.properties.template @@ -80,6 +80,14 @@ # /metrics/aplications/json # App information # /metrics/master/json # Master information +# org.apache.spark.metrics.sink.GraphiteSink +# Name: Default: Description: +# host NONE Hostname of Graphite server +# port NONE Port of Graphite server +# period 10 Poll period +# unit seconds Units of poll period +# prefix EMPTY STRING Prefix to prepend to metric name + ## Examples # Enable JmxSink for all instances by class name #*.sink.jmx.class=org.apache.spark.metrics.sink.JmxSink diff --git a/core/pom.xml b/core/pom.xml index 8621d257e5..38f4be1280 100644 --- a/core/pom.xml +++ b/core/pom.xml @@ -95,15 +95,15 @@ <version>0.3.1</version> </dependency> <dependency> - <groupId>com.typesafe.akka</groupId> + <groupId>${akka.group}</groupId> <artifactId>akka-actor</artifactId> </dependency> <dependency> - <groupId>com.typesafe.akka</groupId> + <groupId>${akka.group}</groupId> <artifactId>akka-remote</artifactId> </dependency> <dependency> - <groupId>com.typesafe.akka</groupId> + <groupId>${akka.group}</groupId> <artifactId>akka-slf4j</artifactId> </dependency> <dependency> @@ -127,10 +127,6 @@ <artifactId>colt</artifactId> </dependency> <dependency> - <groupId>com.github.scala-incubator.io</groupId> - <artifactId>scala-io-file_2.9.2</artifactId> - </dependency> - <dependency> <groupId>org.apache.mesos</groupId> <artifactId>mesos</artifactId> </dependency> @@ -159,6 +155,10 @@ <artifactId>metrics-ganglia</artifactId> </dependency> <dependency> + <groupId>com.codahale.metrics</groupId> + <artifactId>metrics-graphite</artifactId> + </dependency> + <dependency> <groupId>org.apache.derby</groupId> <artifactId>derby</artifactId> <scope>test</scope> diff --git a/core/src/main/scala/org/apache/spark/FutureAction.scala b/core/src/main/scala/org/apache/spark/FutureAction.scala index 1ad9240cfa..c6b4ac5192 100644 --- a/core/src/main/scala/org/apache/spark/FutureAction.scala +++ b/core/src/main/scala/org/apache/spark/FutureAction.scala @@ -99,7 +99,7 @@ class SimpleFutureAction[T] private[spark](jobWaiter: JobWaiter[_], resultFunc: override def ready(atMost: Duration)(implicit permit: CanAwait): SimpleFutureAction.this.type = { if (!atMost.isFinite()) { awaitResult() - } else { + } else jobWaiter.synchronized { val finishTime = System.currentTimeMillis() + atMost.toMillis while (!isCompleted) { val time = System.currentTimeMillis() diff --git a/core/src/main/scala/org/apache/spark/MapOutputTracker.scala b/core/src/main/scala/org/apache/spark/MapOutputTracker.scala index 5e465fa22c..b4d0b7017c 100644 --- a/core/src/main/scala/org/apache/spark/MapOutputTracker.scala +++ b/core/src/main/scala/org/apache/spark/MapOutputTracker.scala @@ -244,12 +244,12 @@ private[spark] class MapOutputTrackerMaster extends MapOutputTracker { case Some(bytes) => return bytes case None => - statuses = mapStatuses(shuffleId) + statuses = mapStatuses.getOrElse(shuffleId, Array[MapStatus]()) epochGotten = epoch } } // If we got here, we failed to find the serialized locations in the cache, so we pulled - // out a snapshot of the locations as "locs"; let's serialize and return that + // out a snapshot of the locations as "statuses"; let's serialize and return that val bytes = MapOutputTracker.serializeMapStatuses(statuses) logInfo("Size of output statuses for shuffle %d is %d bytes".format(shuffleId, bytes.length)) // Add them into the table only if the epoch hasn't changed while we were working @@ -274,6 +274,10 @@ private[spark] class MapOutputTrackerMaster extends MapOutputTracker { override def updateEpoch(newEpoch: Long) { // This might be called on the MapOutputTrackerMaster if we're running in local mode. } + + def has(shuffleId: Int): Boolean = { + cachedSerializedStatuses.get(shuffleId).isDefined || mapStatuses.contains(shuffleId) + } } private[spark] object MapOutputTracker { diff --git a/core/src/main/scala/org/apache/spark/SparkContext.scala b/core/src/main/scala/org/apache/spark/SparkContext.scala index 42b2985b50..66006bf212 100644 --- a/core/src/main/scala/org/apache/spark/SparkContext.scala +++ b/core/src/main/scala/org/apache/spark/SparkContext.scala @@ -81,7 +81,7 @@ class SparkContext( val sparkHome: String = null, val jars: Seq[String] = Nil, val environment: Map[String, String] = Map(), - // This is used only by yarn for now, but should be relevant to other cluster types (mesos, etc) + // 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]] = @@ -153,98 +153,11 @@ class SparkContext( executorEnvs("SPARK_USER") = sparkUser // Create and start the scheduler - private[spark] var taskScheduler: TaskScheduler = { - // Regular expression used for local[N] master format - val LOCAL_N_REGEX = """local\[([0-9]+)\]""".r - // Regular expression for local[N, maxRetries], used in tests with failing tasks - val LOCAL_N_FAILURES_REGEX = """local\[([0-9]+)\s*,\s*([0-9]+)\]""".r - // Regular expression for simulating a Spark cluster of [N, cores, memory] locally - val LOCAL_CLUSTER_REGEX = """local-cluster\[\s*([0-9]+)\s*,\s*([0-9]+)\s*,\s*([0-9]+)\s*]""".r - // Regular expression for connecting to Spark deploy clusters - val SPARK_REGEX = """spark://(.*)""".r - // Regular expression for connection to Mesos cluster - val MESOS_REGEX = """mesos://(.*)""".r - // Regular expression for connection to Simr cluster - val SIMR_REGEX = """simr://(.*)""".r - - master match { - case "local" => - new LocalScheduler(1, 0, this) - - case LOCAL_N_REGEX(threads) => - new LocalScheduler(threads.toInt, 0, this) - - case LOCAL_N_FAILURES_REGEX(threads, maxFailures) => - new LocalScheduler(threads.toInt, maxFailures.toInt, this) - - case SPARK_REGEX(sparkUrl) => - val scheduler = new ClusterScheduler(this) - val masterUrls = sparkUrl.split(",").map("spark://" + _) - val backend = new SparkDeploySchedulerBackend(scheduler, this, masterUrls, appName) - scheduler.initialize(backend) - scheduler - - case SIMR_REGEX(simrUrl) => - val scheduler = new ClusterScheduler(this) - val backend = new SimrSchedulerBackend(scheduler, this, simrUrl) - scheduler.initialize(backend) - scheduler - - case LOCAL_CLUSTER_REGEX(numSlaves, coresPerSlave, memoryPerSlave) => - // Check to make sure memory requested <= memoryPerSlave. Otherwise Spark will just hang. - val memoryPerSlaveInt = memoryPerSlave.toInt - if (SparkContext.executorMemoryRequested > memoryPerSlaveInt) { - throw new SparkException( - "Asked to launch cluster with %d MB RAM / worker but requested %d MB/worker".format( - memoryPerSlaveInt, SparkContext.executorMemoryRequested)) - } - - val scheduler = new ClusterScheduler(this) - val localCluster = new LocalSparkCluster( - numSlaves.toInt, coresPerSlave.toInt, memoryPerSlaveInt) - val masterUrls = localCluster.start() - val backend = new SparkDeploySchedulerBackend(scheduler, this, masterUrls, appName) - scheduler.initialize(backend) - backend.shutdownCallback = (backend: SparkDeploySchedulerBackend) => { - localCluster.stop() - } - scheduler - - case "yarn-standalone" => - val scheduler = try { - val clazz = Class.forName("org.apache.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 CoarseGrainedSchedulerBackend(scheduler, this.env.actorSystem) - scheduler.initialize(backend) - scheduler - - case MESOS_REGEX(mesosUrl) => - MesosNativeLibrary.load() - val scheduler = new ClusterScheduler(this) - val coarseGrained = System.getProperty("spark.mesos.coarse", "false").toBoolean - val backend = if (coarseGrained) { - new CoarseMesosSchedulerBackend(scheduler, this, mesosUrl, appName) - } else { - new MesosSchedulerBackend(scheduler, this, mesosUrl, appName) - } - scheduler.initialize(backend) - scheduler - - case _ => - throw new SparkException("Could not parse Master URL: '" + master + "'") - } - } + private[spark] var taskScheduler = SparkContext.createTaskScheduler(this, master, appName) taskScheduler.start() @volatile private[spark] var dagScheduler = new DAGScheduler(taskScheduler) + dagScheduler.start() ui.start() @@ -1111,6 +1024,124 @@ object SparkContext { .map(Utils.memoryStringToMb) .getOrElse(512) } + + // Creates a task scheduler based on a given master URL. Extracted for testing. + private + def createTaskScheduler(sc: SparkContext, master: String, appName: String): TaskScheduler = { + // Regular expression used for local[N] master format + val LOCAL_N_REGEX = """local\[([0-9]+)\]""".r + // Regular expression for local[N, maxRetries], used in tests with failing tasks + val LOCAL_N_FAILURES_REGEX = """local\[([0-9]+)\s*,\s*([0-9]+)\]""".r + // Regular expression for simulating a Spark cluster of [N, cores, memory] locally + val LOCAL_CLUSTER_REGEX = """local-cluster\[\s*([0-9]+)\s*,\s*([0-9]+)\s*,\s*([0-9]+)\s*]""".r + // Regular expression for connecting to Spark deploy clusters + val SPARK_REGEX = """spark://(.*)""".r + // Regular expression for connection to Mesos cluster by mesos:// or zk:// url + val MESOS_REGEX = """(mesos|zk)://.*""".r + // Regular expression for connection to Simr cluster + val SIMR_REGEX = """simr://(.*)""".r + + master match { + case "local" => + new LocalScheduler(1, 0, sc) + + case LOCAL_N_REGEX(threads) => + new LocalScheduler(threads.toInt, 0, sc) + + case LOCAL_N_FAILURES_REGEX(threads, maxFailures) => + new LocalScheduler(threads.toInt, maxFailures.toInt, sc) + + case SPARK_REGEX(sparkUrl) => + val scheduler = new ClusterScheduler(sc) + val masterUrls = sparkUrl.split(",").map("spark://" + _) + val backend = new SparkDeploySchedulerBackend(scheduler, sc, masterUrls, appName) + scheduler.initialize(backend) + scheduler + + case LOCAL_CLUSTER_REGEX(numSlaves, coresPerSlave, memoryPerSlave) => + // Check to make sure memory requested <= memoryPerSlave. Otherwise Spark will just hang. + val memoryPerSlaveInt = memoryPerSlave.toInt + if (SparkContext.executorMemoryRequested > memoryPerSlaveInt) { + throw new SparkException( + "Asked to launch cluster with %d MB RAM / worker but requested %d MB/worker".format( + memoryPerSlaveInt, SparkContext.executorMemoryRequested)) + } + + val scheduler = new ClusterScheduler(sc) + val localCluster = new LocalSparkCluster( + numSlaves.toInt, coresPerSlave.toInt, memoryPerSlaveInt) + val masterUrls = localCluster.start() + val backend = new SparkDeploySchedulerBackend(scheduler, sc, masterUrls, appName) + scheduler.initialize(backend) + backend.shutdownCallback = (backend: SparkDeploySchedulerBackend) => { + localCluster.stop() + } + scheduler + + case "yarn-standalone" => + val scheduler = try { + val clazz = Class.forName("org.apache.spark.scheduler.cluster.YarnClusterScheduler") + val cons = clazz.getConstructor(classOf[SparkContext]) + cons.newInstance(sc).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 CoarseGrainedSchedulerBackend(scheduler, sc.env.actorSystem) + scheduler.initialize(backend) + scheduler + + case "yarn-client" => + val scheduler = try { + val clazz = Class.forName("org.apache.spark.scheduler.cluster.YarnClientClusterScheduler") + val cons = clazz.getConstructor(classOf[SparkContext]) + cons.newInstance(sc).asInstanceOf[ClusterScheduler] + + } catch { + case th: Throwable => { + throw new SparkException("YARN mode not available ?", th) + } + } + + val backend = try { + val clazz = Class.forName("org.apache.spark.scheduler.cluster.YarnClientSchedulerBackend") + val cons = clazz.getConstructor(classOf[ClusterScheduler], classOf[SparkContext]) + cons.newInstance(scheduler, sc).asInstanceOf[CoarseGrainedSchedulerBackend] + } catch { + case th: Throwable => { + throw new SparkException("YARN mode not available ?", th) + } + } + + scheduler.initialize(backend) + scheduler + + case mesosUrl @ MESOS_REGEX(_) => + MesosNativeLibrary.load() + val scheduler = new ClusterScheduler(sc) + val coarseGrained = System.getProperty("spark.mesos.coarse", "false").toBoolean + val url = mesosUrl.stripPrefix("mesos://") // strip scheme from raw Mesos URLs + val backend = if (coarseGrained) { + new CoarseMesosSchedulerBackend(scheduler, sc, url, appName) + } else { + new MesosSchedulerBackend(scheduler, sc, url, appName) + } + scheduler.initialize(backend) + scheduler + + case SIMR_REGEX(simrUrl) => + val scheduler = new ClusterScheduler(sc) + val backend = new SimrSchedulerBackend(scheduler, sc, simrUrl) + scheduler.initialize(backend) + scheduler + + case _ => + throw new SparkException("Could not parse Master URL: '" + master + "'") + } + } } /** diff --git a/core/src/main/scala/org/apache/spark/api/java/JavaDoubleRDD.scala b/core/src/main/scala/org/apache/spark/api/java/JavaDoubleRDD.scala index 043cb183ba..9f02a9b7d3 100644 --- a/core/src/main/scala/org/apache/spark/api/java/JavaDoubleRDD.scala +++ b/core/src/main/scala/org/apache/spark/api/java/JavaDoubleRDD.scala @@ -26,6 +26,8 @@ import org.apache.spark.storage.StorageLevel import java.lang.Double import org.apache.spark.Partitioner +import scala.collection.JavaConverters._ + class JavaDoubleRDD(val srdd: RDD[scala.Double]) extends JavaRDDLike[Double, JavaDoubleRDD] { override val classManifest: ClassManifest[Double] = implicitly[ClassManifest[Double]] @@ -182,6 +184,44 @@ class JavaDoubleRDD(val srdd: RDD[scala.Double]) extends JavaRDDLike[Double, Jav /** (Experimental) Approximate operation to return the sum within a timeout. */ def sumApprox(timeout: Long): PartialResult[BoundedDouble] = srdd.sumApprox(timeout) + + /** + * Compute a histogram of the data using bucketCount number of buckets evenly + * spaced between the minimum and maximum of the RDD. For example if the min + * value is 0 and the max is 100 and there are two buckets the resulting + * buckets will be [0,50) [50,100]. bucketCount must be at least 1 + * If the RDD contains infinity, NaN throws an exception + * If the elements in RDD do not vary (max == min) always returns a single bucket. + */ + def histogram(bucketCount: Int): Pair[Array[scala.Double], Array[Long]] = { + val result = srdd.histogram(bucketCount) + (result._1, result._2) + } + + /** + * Compute a histogram using the provided buckets. The buckets are all open + * to the left except for the last which is closed + * e.g. for the array + * [1,10,20,50] the buckets are [1,10) [10,20) [20,50] + * e.g 1<=x<10 , 10<=x<20, 20<=x<50 + * And on the input of 1 and 50 we would have a histogram of 1,0,0 + * + * Note: if your histogram is evenly spaced (e.g. [0, 10, 20, 30]) this can be switched + * from an O(log n) inseration to O(1) per element. (where n = # buckets) if you set evenBuckets + * to true. + * buckets must be sorted and not contain any duplicates. + * buckets array must be at least two elements + * All NaN entries are treated the same. If you have a NaN bucket it must be + * the maximum value of the last position and all NaN entries will be counted + * in that bucket. + */ + def histogram(buckets: Array[scala.Double]): Array[Long] = { + srdd.histogram(buckets, false) + } + + def histogram(buckets: Array[Double], evenBuckets: Boolean): Array[Long] = { + srdd.histogram(buckets.map(_.toDouble), evenBuckets) + } } object JavaDoubleRDD { diff --git a/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala b/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala index 12b4d94a56..132e4fb0d2 100644 --- a/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala +++ b/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala @@ -27,13 +27,12 @@ import org.apache.spark.api.java.{JavaSparkContext, JavaPairRDD, JavaRDD} import org.apache.spark.broadcast.Broadcast import org.apache.spark._ import org.apache.spark.rdd.RDD -import org.apache.spark.rdd.PipedRDD import org.apache.spark.util.Utils private[spark] class PythonRDD[T: ClassManifest]( parent: RDD[T], - command: Seq[String], + command: Array[Byte], envVars: JMap[String, String], pythonIncludes: JList[String], preservePartitoning: Boolean, @@ -44,21 +43,10 @@ private[spark] class PythonRDD[T: ClassManifest]( val bufferSize = System.getProperty("spark.buffer.size", "65536").toInt - // 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: JMap[String, String], - pythonIncludes: JList[String], - preservePartitoning: Boolean, pythonExec: String, - broadcastVars: JList[Broadcast[Array[Byte]]], - accumulator: Accumulator[JList[Array[Byte]]]) = - this(parent, PipedRDD.tokenize(command), envVars, pythonIncludes, preservePartitoning, pythonExec, - broadcastVars, accumulator) - override def getPartitions = parent.partitions override val partitioner = if (preservePartitoning) parent.partitioner else None - override def compute(split: Partition, context: TaskContext): Iterator[Array[Byte]] = { val startTime = System.currentTimeMillis val env = SparkEnv.get @@ -71,11 +59,10 @@ private[spark] class PythonRDD[T: ClassManifest]( SparkEnv.set(env) val stream = new BufferedOutputStream(worker.getOutputStream, bufferSize) val dataOut = new DataOutputStream(stream) - val printOut = new PrintWriter(stream) // Partition index dataOut.writeInt(split.index) // sparkFilesDir - PythonRDD.writeAsPickle(SparkFiles.getRootDirectory, dataOut) + dataOut.writeUTF(SparkFiles.getRootDirectory) // Broadcast variables dataOut.writeInt(broadcastVars.length) for (broadcast <- broadcastVars) { @@ -85,21 +72,16 @@ private[spark] class PythonRDD[T: ClassManifest]( } // Python includes (*.zip and *.egg files) dataOut.writeInt(pythonIncludes.length) - for (f <- pythonIncludes) { - PythonRDD.writeAsPickle(f, dataOut) - } + pythonIncludes.foreach(dataOut.writeUTF) dataOut.flush() - // Serialized user code - for (elem <- command) { - printOut.println(elem) - } - printOut.flush() + // Serialized command: + dataOut.writeInt(command.length) + dataOut.write(command) // Data values for (elem <- parent.iterator(split, context)) { - PythonRDD.writeAsPickle(elem, dataOut) + PythonRDD.writeToStream(elem, dataOut) } dataOut.flush() - printOut.flush() worker.shutdownOutput() } catch { case e: IOException => @@ -132,7 +114,7 @@ private[spark] class PythonRDD[T: ClassManifest]( val obj = new Array[Byte](length) stream.readFully(obj) obj - case -3 => + case SpecialLengths.TIMING_DATA => // Timing data from worker val bootTime = stream.readLong() val initTime = stream.readLong() @@ -143,24 +125,24 @@ private[spark] class PythonRDD[T: ClassManifest]( val total = finishTime - startTime logInfo("Times: total = %s, boot = %s, init = %s, finish = %s".format(total, boot, init, finish)) read - case -2 => + case SpecialLengths.PYTHON_EXCEPTION_THROWN => // Signals that an exception has been thrown in python val exLength = stream.readInt() val obj = new Array[Byte](exLength) stream.readFully(obj) throw new PythonException(new String(obj)) - case -1 => + case SpecialLengths.END_OF_DATA_SECTION => // 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) + // read some accumulator updates: + val numAccumulatorUpdates = stream.readInt() + (1 to numAccumulatorUpdates).foreach { _ => + val updateLen = stream.readInt() + val update = new Array[Byte](updateLen) stream.readFully(update) accumulator += Collections.singletonList(update) - len2 = stream.readInt() + } - new Array[Byte](0) + Array.empty[Byte] } } catch { case eof: EOFException => { @@ -197,62 +179,15 @@ private class PairwiseRDD(prev: RDD[Array[Byte]]) extends val asJavaPairRDD : JavaPairRDD[Long, Array[Byte]] = JavaPairRDD.fromRDD(this) } -private[spark] object PythonRDD { - - /** Strips the pickle PROTO and STOP opcodes from the start and end of a pickle */ - def stripPickle(arr: Array[Byte]) : Array[Byte] = { - arr.slice(2, arr.length - 1) - } +private object SpecialLengths { + val END_OF_DATA_SECTION = -1 + val PYTHON_EXCEPTION_THROWN = -2 + val TIMING_DATA = -3 +} - /** - * Write strings, pickled Python objects, or pairs of pickled objects to a data output stream. - * The data format is a 32-bit integer representing the pickled object's length (in bytes), - * followed by the pickled data. - * - * Pickle module: - * - * http://docs.python.org/2/library/pickle.html - * - * The pickle protocol is documented in the source of the `pickle` and `pickletools` modules: - * - * http://hg.python.org/cpython/file/2.6/Lib/pickle.py - * http://hg.python.org/cpython/file/2.6/Lib/pickletools.py - * - * @param elem the object to write - * @param dOut a data output stream - */ - def writeAsPickle(elem: Any, dOut: DataOutputStream) { - if (elem.isInstanceOf[Array[Byte]]) { - val arr = elem.asInstanceOf[Array[Byte]] - dOut.writeInt(arr.length) - dOut.write(arr) - } else if (elem.isInstanceOf[scala.Tuple2[Array[Byte], Array[Byte]]]) { - val t = elem.asInstanceOf[scala.Tuple2[Array[Byte], Array[Byte]]] - val length = t._1.length + t._2.length - 3 - 3 + 4 // stripPickle() removes 3 bytes - dOut.writeInt(length) - dOut.writeByte(Pickle.PROTO) - dOut.writeByte(Pickle.TWO) - dOut.write(PythonRDD.stripPickle(t._1)) - dOut.write(PythonRDD.stripPickle(t._2)) - dOut.writeByte(Pickle.TUPLE2) - dOut.writeByte(Pickle.STOP) - } else if (elem.isInstanceOf[String]) { - // For uniformity, strings are wrapped into Pickles. - val s = elem.asInstanceOf[String].getBytes("UTF-8") - val length = 2 + 1 + 4 + s.length + 1 - dOut.writeInt(length) - dOut.writeByte(Pickle.PROTO) - dOut.writeByte(Pickle.TWO) - dOut.write(Pickle.BINUNICODE) - dOut.writeInt(Integer.reverseBytes(s.length)) - dOut.write(s) - dOut.writeByte(Pickle.STOP) - } else { - throw new SparkException("Unexpected RDD type") - } - } +private[spark] object PythonRDD { - def readRDDFromPickleFile(sc: JavaSparkContext, filename: String, parallelism: Int) : + def readRDDFromFile(sc: JavaSparkContext, filename: String, parallelism: Int): JavaRDD[Array[Byte]] = { val file = new DataInputStream(new FileInputStream(filename)) val objs = new collection.mutable.ArrayBuffer[Array[Byte]] @@ -270,15 +205,32 @@ private[spark] object PythonRDD { JavaRDD.fromRDD(sc.sc.parallelize(objs, parallelism)) } - def writeIteratorToPickleFile[T](items: java.util.Iterator[T], filename: String) { + def writeToStream(elem: Any, dataOut: DataOutputStream) { + elem match { + case bytes: Array[Byte] => + dataOut.writeInt(bytes.length) + dataOut.write(bytes) + case pair: (Array[Byte], Array[Byte]) => + dataOut.writeInt(pair._1.length) + dataOut.write(pair._1) + dataOut.writeInt(pair._2.length) + dataOut.write(pair._2) + case str: String => + dataOut.writeUTF(str) + case other => + throw new SparkException("Unexpected element type " + other.getClass) + } + } + + def writeToFile[T](items: java.util.Iterator[T], filename: String) { import scala.collection.JavaConverters._ - writeIteratorToPickleFile(items.asScala, filename) + writeToFile(items.asScala, filename) } - def writeIteratorToPickleFile[T](items: Iterator[T], filename: String) { + def writeToFile[T](items: Iterator[T], filename: String) { val file = new DataOutputStream(new FileOutputStream(filename)) for (item <- items) { - writeAsPickle(item, file) + writeToStream(item, file) } file.close() } @@ -289,17 +241,6 @@ private[spark] object PythonRDD { } } -private object Pickle { - val PROTO: Byte = 0x80.toByte - val TWO: Byte = 0x02.toByte - val BINUNICODE: Byte = 'X' - val STOP: Byte = '.' - val TUPLE2: Byte = 0x86.toByte - val EMPTY_LIST: Byte = ']' - val MARK: Byte = '(' - val APPENDS: Byte = 'e' -} - private class BytesToString extends org.apache.spark.api.java.function.Function[Array[Byte], String] { override def call(arr: Array[Byte]) : String = new String(arr, "UTF-8") } diff --git a/core/src/main/scala/org/apache/spark/broadcast/HttpBroadcast.scala b/core/src/main/scala/org/apache/spark/broadcast/HttpBroadcast.scala index 609464e38d..47db720416 100644 --- a/core/src/main/scala/org/apache/spark/broadcast/HttpBroadcast.scala +++ b/core/src/main/scala/org/apache/spark/broadcast/HttpBroadcast.scala @@ -19,6 +19,7 @@ package org.apache.spark.broadcast import java.io.{File, FileOutputStream, ObjectInputStream, OutputStream} import java.net.URL +import java.util.concurrent.TimeUnit import it.unimi.dsi.fastutil.io.FastBufferedInputStream import it.unimi.dsi.fastutil.io.FastBufferedOutputStream @@ -83,6 +84,8 @@ private object HttpBroadcast extends Logging { private val files = new TimeStampedHashSet[String] private val cleaner = new MetadataCleaner(MetadataCleanerType.HTTP_BROADCAST, cleanup) + private val httpReadTimeout = TimeUnit.MILLISECONDS.convert(5,TimeUnit.MINUTES).toInt + private lazy val compressionCodec = CompressionCodec.createCodec() def initialize(isDriver: Boolean) { @@ -138,10 +141,13 @@ private object HttpBroadcast extends Logging { def read[T](id: Long): T = { val url = serverUri + "/" + BroadcastBlockId(id).name val in = { + val httpConnection = new URL(url).openConnection() + httpConnection.setReadTimeout(httpReadTimeout) + val inputStream = httpConnection.getInputStream() if (compress) { - compressionCodec.compressedInputStream(new URL(url).openStream()) + compressionCodec.compressedInputStream(inputStream) } else { - new FastBufferedInputStream(new URL(url).openStream(), bufferSize) + new FastBufferedInputStream(inputStream, bufferSize) } } val ser = SparkEnv.get.serializer.newInstance() diff --git a/core/src/main/scala/org/apache/spark/deploy/FaultToleranceTest.scala b/core/src/main/scala/org/apache/spark/deploy/FaultToleranceTest.scala index 668032a3a2..0aa8852649 100644 --- a/core/src/main/scala/org/apache/spark/deploy/FaultToleranceTest.scala +++ b/core/src/main/scala/org/apache/spark/deploy/FaultToleranceTest.scala @@ -1,19 +1,19 @@ /* * - * * Licensed to the Apache Software Foundation (ASF) under one or more - * * contributor license agreements. See the NOTICE file distributed with - * * this work for additional information regarding copyright ownership. - * * The ASF licenses this file to You under the Apache License, Version 2.0 - * * (the "License"); you may not use this file except in compliance with - * * the License. You may obtain a copy of the License at - * * - * * http://www.apache.org/licenses/LICENSE-2.0 - * * - * * Unless required by applicable law or agreed to in writing, software - * * distributed under the License is distributed on an "AS IS" BASIS, - * * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * * See the License for the specific language governing permissions and - * * limitations under the License. + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. * */ diff --git a/core/src/main/scala/org/apache/spark/executor/TaskMetrics.scala b/core/src/main/scala/org/apache/spark/executor/TaskMetrics.scala index 0b4892f98f..c0ce46e379 100644 --- a/core/src/main/scala/org/apache/spark/executor/TaskMetrics.scala +++ b/core/src/main/scala/org/apache/spark/executor/TaskMetrics.scala @@ -61,50 +61,53 @@ object TaskMetrics { class ShuffleReadMetrics extends Serializable { /** - * Time when shuffle finishs + * Absolute time when this task finished reading shuffle data */ var shuffleFinishTime: Long = _ /** - * Total number of blocks fetched in a shuffle (remote or local) + * Number of blocks fetched in this shuffle by this task (remote or local) */ var totalBlocksFetched: Int = _ /** - * Number of remote blocks fetched in a shuffle + * Number of remote blocks fetched in this shuffle by this task */ var remoteBlocksFetched: Int = _ /** - * Local blocks fetched in a shuffle + * Number of local blocks fetched in this shuffle by this task */ var localBlocksFetched: Int = _ /** - * Total time that is spent blocked waiting for shuffle to fetch data + * Time the task spent waiting for remote shuffle blocks. This only includes the time + * blocking on shuffle input data. For instance if block B is being fetched while the task is + * still not finished processing block A, it is not considered to be blocking on block B. */ 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 + * Total time spent fetching remote shuffle blocks. This aggregates the time spent fetching all + * input blocks. Since block fetches are both pipelined and parallelized, this can + * exceed fetchWaitTime and executorRunTime. */ var remoteFetchTime: Long = _ /** - * Total number of remote bytes read from a shuffle + * Total number of remote bytes read from the shuffle by this task */ var remoteBytesRead: Long = _ } class ShuffleWriteMetrics extends Serializable { /** - * Number of bytes written for a shuffle + * Number of bytes written for the shuffle by this task */ var shuffleBytesWritten: Long = _ /** - * Time spent blocking on writes to disk or buffer cache, in nanoseconds. + * Time the task spent blocking on writes to disk or buffer cache, in nanoseconds */ var shuffleWriteTime: Long = _ } diff --git a/core/src/main/scala/org/apache/spark/metrics/sink/GraphiteSink.scala b/core/src/main/scala/org/apache/spark/metrics/sink/GraphiteSink.scala new file mode 100644 index 0000000000..cdcfec8ca7 --- /dev/null +++ b/core/src/main/scala/org/apache/spark/metrics/sink/GraphiteSink.scala @@ -0,0 +1,82 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.metrics.sink + +import java.util.Properties +import java.util.concurrent.TimeUnit +import java.net.InetSocketAddress + +import com.codahale.metrics.MetricRegistry +import com.codahale.metrics.graphite.{GraphiteReporter, Graphite} + +import org.apache.spark.metrics.MetricsSystem + +class GraphiteSink(val property: Properties, val registry: MetricRegistry) extends Sink { + val GRAPHITE_DEFAULT_PERIOD = 10 + val GRAPHITE_DEFAULT_UNIT = "SECONDS" + val GRAPHITE_DEFAULT_PREFIX = "" + + val GRAPHITE_KEY_HOST = "host" + val GRAPHITE_KEY_PORT = "port" + val GRAPHITE_KEY_PERIOD = "period" + val GRAPHITE_KEY_UNIT = "unit" + val GRAPHITE_KEY_PREFIX = "prefix" + + def propertyToOption(prop: String) = Option(property.getProperty(prop)) + + if (!propertyToOption(GRAPHITE_KEY_HOST).isDefined) { + throw new Exception("Graphite sink requires 'host' property.") + } + + if (!propertyToOption(GRAPHITE_KEY_PORT).isDefined) { + throw new Exception("Graphite sink requires 'port' property.") + } + + val host = propertyToOption(GRAPHITE_KEY_HOST).get + val port = propertyToOption(GRAPHITE_KEY_PORT).get.toInt + + val pollPeriod = propertyToOption(GRAPHITE_KEY_PERIOD) match { + case Some(s) => s.toInt + case None => GRAPHITE_DEFAULT_PERIOD + } + + val pollUnit = propertyToOption(GRAPHITE_KEY_UNIT) match { + case Some(s) => TimeUnit.valueOf(s.toUpperCase()) + case None => TimeUnit.valueOf(GRAPHITE_DEFAULT_UNIT) + } + + val prefix = propertyToOption(GRAPHITE_KEY_PREFIX).getOrElse(GRAPHITE_DEFAULT_PREFIX) + + MetricsSystem.checkMinimalPollingPeriod(pollUnit, pollPeriod) + + val graphite: Graphite = new Graphite(new InetSocketAddress(host, port)) + + val reporter: GraphiteReporter = GraphiteReporter.forRegistry(registry) + .convertDurationsTo(TimeUnit.MILLISECONDS) + .convertRatesTo(TimeUnit.SECONDS) + .prefixedWith(prefix) + .build(graphite) + + override def start() { + reporter.start(pollPeriod, pollUnit) + } + + override def stop() { + reporter.stop() + } +} diff --git a/core/src/main/scala/org/apache/spark/network/netty/ShuffleCopier.scala b/core/src/main/scala/org/apache/spark/network/netty/ShuffleCopier.scala index 481ff8c3e0..b1e1576dad 100644 --- a/core/src/main/scala/org/apache/spark/network/netty/ShuffleCopier.scala +++ b/core/src/main/scala/org/apache/spark/network/netty/ShuffleCopier.scala @@ -76,7 +76,7 @@ private[spark] object ShuffleCopier extends Logging { extends FileClientHandler with Logging { override def handle(ctx: ChannelHandlerContext, in: ByteBuf, header: FileHeader) { - logDebug("Received Block: " + header.blockId + " (" + header.fileLen + "B)"); + logDebug("Received Block: " + header.blockId + " (" + header.fileLen + "B)") resultCollectCallBack(header.blockId, header.fileLen.toLong, in.readBytes(header.fileLen)) } diff --git a/core/src/main/scala/org/apache/spark/rdd/CartesianRDD.scala b/core/src/main/scala/org/apache/spark/rdd/CartesianRDD.scala index 9b0c882481..0de22f0e06 100644 --- a/core/src/main/scala/org/apache/spark/rdd/CartesianRDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/CartesianRDD.scala @@ -70,7 +70,7 @@ class CartesianRDD[T: ClassManifest, U:ClassManifest]( override def compute(split: Partition, context: TaskContext) = { val currSplit = split.asInstanceOf[CartesianPartition] for (x <- rdd1.iterator(currSplit.s1, context); - y <- rdd2.iterator(currSplit.s2, context)) yield (x, y) + y <- rdd2.iterator(currSplit.s2, context)) yield (x, y) } override def getDependencies: Seq[Dependency[_]] = List( diff --git a/core/src/main/scala/org/apache/spark/rdd/DoubleRDDFunctions.scala b/core/src/main/scala/org/apache/spark/rdd/DoubleRDDFunctions.scala index a4bec41752..02d75eccc5 100644 --- a/core/src/main/scala/org/apache/spark/rdd/DoubleRDDFunctions.scala +++ b/core/src/main/scala/org/apache/spark/rdd/DoubleRDDFunctions.scala @@ -24,6 +24,8 @@ import org.apache.spark.partial.SumEvaluator import org.apache.spark.util.StatCounter import org.apache.spark.{TaskContext, Logging} +import scala.collection.immutable.NumericRange + /** * Extra functions available on RDDs of Doubles through an implicit conversion. * Import `org.apache.spark.SparkContext._` at the top of your program to use these functions. @@ -76,4 +78,128 @@ class DoubleRDDFunctions(self: RDD[Double]) extends Logging with Serializable { val evaluator = new SumEvaluator(self.partitions.size, confidence) self.context.runApproximateJob(self, processPartition, evaluator, timeout) } + + /** + * Compute a histogram of the data using bucketCount number of buckets evenly + * spaced between the minimum and maximum of the RDD. For example if the min + * value is 0 and the max is 100 and there are two buckets the resulting + * buckets will be [0, 50) [50, 100]. bucketCount must be at least 1 + * If the RDD contains infinity, NaN throws an exception + * If the elements in RDD do not vary (max == min) always returns a single bucket. + */ + def histogram(bucketCount: Int): Pair[Array[Double], Array[Long]] = { + // Compute the minimum and the maxium + val (max: Double, min: Double) = self.mapPartitions { items => + Iterator(items.foldRight(-1/0.0, Double.NaN)((e: Double, x: Pair[Double, Double]) => + (x._1.max(e), x._2.min(e)))) + }.reduce { (maxmin1, maxmin2) => + (maxmin1._1.max(maxmin2._1), maxmin1._2.min(maxmin2._2)) + } + if (max.isNaN() || max.isInfinity || min.isInfinity ) { + throw new UnsupportedOperationException( + "Histogram on either an empty RDD or RDD containing +/-infinity or NaN") + } + val increment = (max-min)/bucketCount.toDouble + val range = if (increment != 0) { + Range.Double.inclusive(min, max, increment) + } else { + List(min, min) + } + val buckets = range.toArray + (buckets, histogram(buckets, true)) + } + + /** + * Compute a histogram using the provided buckets. The buckets are all open + * to the left except for the last which is closed + * e.g. for the array + * [1, 10, 20, 50] the buckets are [1, 10) [10, 20) [20, 50] + * e.g 1<=x<10 , 10<=x<20, 20<=x<50 + * And on the input of 1 and 50 we would have a histogram of 1, 0, 0 + * + * Note: if your histogram is evenly spaced (e.g. [0, 10, 20, 30]) this can be switched + * from an O(log n) inseration to O(1) per element. (where n = # buckets) if you set evenBuckets + * to true. + * buckets must be sorted and not contain any duplicates. + * buckets array must be at least two elements + * All NaN entries are treated the same. If you have a NaN bucket it must be + * the maximum value of the last position and all NaN entries will be counted + * in that bucket. + */ + def histogram(buckets: Array[Double], evenBuckets: Boolean = false): Array[Long] = { + if (buckets.length < 2) { + throw new IllegalArgumentException("buckets array must have at least two elements") + } + // The histogramPartition function computes the partail histogram for a given + // partition. The provided bucketFunction determines which bucket in the array + // to increment or returns None if there is no bucket. This is done so we can + // specialize for uniformly distributed buckets and save the O(log n) binary + // search cost. + def histogramPartition(bucketFunction: (Double) => Option[Int])(iter: Iterator[Double]): + Iterator[Array[Long]] = { + val counters = new Array[Long](buckets.length - 1) + while (iter.hasNext) { + bucketFunction(iter.next()) match { + case Some(x: Int) => {counters(x) += 1} + case _ => {} + } + } + Iterator(counters) + } + // Merge the counters. + def mergeCounters(a1: Array[Long], a2: Array[Long]): Array[Long] = { + a1.indices.foreach(i => a1(i) += a2(i)) + a1 + } + // Basic bucket function. This works using Java's built in Array + // binary search. Takes log(size(buckets)) + def basicBucketFunction(e: Double): Option[Int] = { + val location = java.util.Arrays.binarySearch(buckets, e) + if (location < 0) { + // If the location is less than 0 then the insertion point in the array + // to keep it sorted is -location-1 + val insertionPoint = -location-1 + // If we have to insert before the first element or after the last one + // its out of bounds. + // We do this rather than buckets.lengthCompare(insertionPoint) + // because Array[Double] fails to override it (for now). + if (insertionPoint > 0 && insertionPoint < buckets.length) { + Some(insertionPoint-1) + } else { + None + } + } else if (location < buckets.length - 1) { + // Exact match, just insert here + Some(location) + } else { + // Exact match to the last element + Some(location - 1) + } + } + // Determine the bucket function in constant time. Requires that buckets are evenly spaced + def fastBucketFunction(min: Double, increment: Double, count: Int)(e: Double): Option[Int] = { + // If our input is not a number unless the increment is also NaN then we fail fast + if (e.isNaN()) { + return None + } + val bucketNumber = (e - min)/(increment) + // We do this rather than buckets.lengthCompare(bucketNumber) + // because Array[Double] fails to override it (for now). + if (bucketNumber > count || bucketNumber < 0) { + None + } else { + Some(bucketNumber.toInt.min(count - 1)) + } + } + // Decide which bucket function to pass to histogramPartition. We decide here + // rather than having a general function so that the decission need only be made + // once rather than once per shard + val bucketFunction = if (evenBuckets) { + fastBucketFunction(buckets(0), buckets(1)-buckets(0), buckets.length-1) _ + } else { + basicBucketFunction _ + } + self.mapPartitions(histogramPartition(bucketFunction)).reduce(mergeCounters) + } + } diff --git a/core/src/main/scala/org/apache/spark/rdd/MapPartitionsRDD.scala b/core/src/main/scala/org/apache/spark/rdd/MapPartitionsRDD.scala index 203179c4ea..ae70d55951 100644 --- a/core/src/main/scala/org/apache/spark/rdd/MapPartitionsRDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/MapPartitionsRDD.scala @@ -20,18 +20,16 @@ package org.apache.spark.rdd import org.apache.spark.{Partition, TaskContext} -private[spark] -class MapPartitionsRDD[U: ClassManifest, T: ClassManifest]( +private[spark] class MapPartitionsRDD[U: ClassManifest, T: ClassManifest]( prev: RDD[T], - f: Iterator[T] => Iterator[U], + f: (TaskContext, Int, Iterator[T]) => Iterator[U], // (TaskContext, partition index, iterator) preservesPartitioning: Boolean = false) extends RDD[U](prev) { - override val partitioner = - if (preservesPartitioning) firstParent[T].partitioner else None + override val partitioner = if (preservesPartitioning) firstParent[T].partitioner else None override def getPartitions: Array[Partition] = firstParent[T].partitions override def compute(split: Partition, context: TaskContext) = - f(firstParent[T].iterator(split, context)) + f(context, split.index, firstParent[T].iterator(split, context)) } diff --git a/core/src/main/scala/org/apache/spark/rdd/MapPartitionsWithContextRDD.scala b/core/src/main/scala/org/apache/spark/rdd/MapPartitionsWithContextRDD.scala deleted file mode 100644 index aea08ff81b..0000000000 --- a/core/src/main/scala/org/apache/spark/rdd/MapPartitionsWithContextRDD.scala +++ /dev/null @@ -1,41 +0,0 @@ -/* - * Licensed to the Apache Software Foundation (ASF) under one or more - * contributor license agreements. See the NOTICE file distributed with - * this work for additional information regarding copyright ownership. - * The ASF licenses this file to You under the Apache License, Version 2.0 - * (the "License"); you may not use this file except in compliance with - * the License. You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ - -package org.apache.spark.rdd - -import org.apache.spark.{Partition, TaskContext} - - -/** - * A variant of the MapPartitionsRDD that passes the TaskContext into the closure. From the - * TaskContext, the closure can either get access to the interruptible flag or get the index - * of the partition in the RDD. - */ -private[spark] -class MapPartitionsWithContextRDD[U: ClassManifest, T: ClassManifest]( - prev: RDD[T], - f: (TaskContext, Iterator[T]) => Iterator[U], - preservesPartitioning: Boolean - ) extends RDD[U](prev) { - - override def getPartitions: Array[Partition] = firstParent[T].partitions - - override val partitioner = if (preservesPartitioning) prev.partitioner else None - - override def compute(split: Partition, context: TaskContext) = - f(context, firstParent[T].iterator(split, context)) -} diff --git a/core/src/main/scala/org/apache/spark/rdd/PartitionPruningRDD.scala b/core/src/main/scala/org/apache/spark/rdd/PartitionPruningRDD.scala index 165cd412fc..574dd4233f 100644 --- a/core/src/main/scala/org/apache/spark/rdd/PartitionPruningRDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/PartitionPruningRDD.scala @@ -33,11 +33,13 @@ class PruneDependency[T](rdd: RDD[T], @transient partitionFilterFunc: Int => Boo extends NarrowDependency[T](rdd) { @transient - val partitions: Array[Partition] = rdd.partitions.zipWithIndex - .filter(s => partitionFilterFunc(s._2)) + val partitions: Array[Partition] = rdd.partitions + .filter(s => partitionFilterFunc(s.index)).zipWithIndex .map { case(split, idx) => new PartitionPruningRDDPartition(idx, split) : Partition } - override def getParents(partitionId: Int) = List(partitions(partitionId).index) + override def getParents(partitionId: Int) = { + List(partitions(partitionId).asInstanceOf[PartitionPruningRDDPartition].parentSplit.index) + } } diff --git a/core/src/main/scala/org/apache/spark/rdd/RDD.scala b/core/src/main/scala/org/apache/spark/rdd/RDD.scala index 6e88be6f6a..893708f8f2 100644 --- a/core/src/main/scala/org/apache/spark/rdd/RDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/RDD.scala @@ -101,7 +101,7 @@ abstract class RDD[T: ClassManifest]( protected def getPreferredLocations(split: Partition): Seq[String] = Nil /** Optionally overridden by subclasses to specify how they are partitioned. */ - val partitioner: Option[Partitioner] = None + @transient val partitioner: Option[Partitioner] = None // ======================================================================= // Methods and fields available on all RDDs @@ -114,7 +114,7 @@ abstract class RDD[T: ClassManifest]( val id: Int = sc.newRddId() /** A friendly name for this RDD */ - var name: String = null + @transient var name: String = null /** Assign a name to this RDD */ def setName(_name: String) = { @@ -123,7 +123,7 @@ abstract class RDD[T: ClassManifest]( } /** User-defined generator of this RDD*/ - var generator = Utils.getCallSiteInfo.firstUserClass + @transient var generator = Utils.getCallSiteInfo.firstUserClass /** Reset generator*/ def setGenerator(_generator: String) = { @@ -408,7 +408,6 @@ abstract class RDD[T: ClassManifest]( 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. * The print behavior can be customized by providing two functions. @@ -442,7 +441,8 @@ abstract class RDD[T: ClassManifest]( */ def mapPartitions[U: ClassManifest]( f: Iterator[T] => Iterator[U], preservesPartitioning: Boolean = false): RDD[U] = { - new MapPartitionsRDD(this, sc.clean(f), preservesPartitioning) + val func = (context: TaskContext, index: Int, iter: Iterator[T]) => f(iter) + new MapPartitionsRDD(this, sc.clean(func), preservesPartitioning) } /** @@ -451,8 +451,8 @@ abstract class RDD[T: ClassManifest]( */ def mapPartitionsWithIndex[U: ClassManifest]( f: (Int, Iterator[T]) => Iterator[U], preservesPartitioning: Boolean = false): RDD[U] = { - val func = (context: TaskContext, iter: Iterator[T]) => f(context.partitionId, iter) - new MapPartitionsWithContextRDD(this, sc.clean(func), preservesPartitioning) + val func = (context: TaskContext, index: Int, iter: Iterator[T]) => f(index, iter) + new MapPartitionsRDD(this, sc.clean(func), preservesPartitioning) } /** @@ -462,7 +462,8 @@ abstract class RDD[T: ClassManifest]( def mapPartitionsWithContext[U: ClassManifest]( f: (TaskContext, Iterator[T]) => Iterator[U], preservesPartitioning: Boolean = false): RDD[U] = { - new MapPartitionsWithContextRDD(this, sc.clean(f), preservesPartitioning) + val func = (context: TaskContext, index: Int, iter: Iterator[T]) => f(context, iter) + new MapPartitionsRDD(this, sc.clean(func), preservesPartitioning) } /** @@ -483,11 +484,10 @@ abstract class RDD[T: ClassManifest]( def mapWith[A: ClassManifest, U: ClassManifest] (constructA: Int => A, preservesPartitioning: Boolean = false) (f: (T, A) => U): RDD[U] = { - def iterF(context: TaskContext, iter: Iterator[T]): Iterator[U] = { - val a = constructA(context.partitionId) + mapPartitionsWithIndex((index, iter) => { + val a = constructA(index) iter.map(t => f(t, a)) - } - new MapPartitionsWithContextRDD(this, sc.clean(iterF _), preservesPartitioning) + }, preservesPartitioning) } /** @@ -498,11 +498,10 @@ abstract class RDD[T: ClassManifest]( def flatMapWith[A: ClassManifest, U: ClassManifest] (constructA: Int => A, preservesPartitioning: Boolean = false) (f: (T, A) => Seq[U]): RDD[U] = { - def iterF(context: TaskContext, iter: Iterator[T]): Iterator[U] = { - val a = constructA(context.partitionId) + mapPartitionsWithIndex((index, iter) => { + val a = constructA(index) iter.flatMap(t => f(t, a)) - } - new MapPartitionsWithContextRDD(this, sc.clean(iterF _), preservesPartitioning) + }, preservesPartitioning) } /** @@ -511,11 +510,10 @@ abstract class RDD[T: ClassManifest]( * partition with the index of that partition. */ def foreachWith[A: ClassManifest](constructA: Int => A)(f: (T, A) => Unit) { - def iterF(context: TaskContext, iter: Iterator[T]): Iterator[T] = { - val a = constructA(context.partitionId) + mapPartitionsWithIndex { (index, iter) => + val a = constructA(index) iter.map(t => {f(t, a); t}) - } - new MapPartitionsWithContextRDD(this, sc.clean(iterF _), true).foreach(_ => {}) + }.foreach(_ => {}) } /** @@ -524,11 +522,10 @@ abstract class RDD[T: ClassManifest]( * partition with the index of that partition. */ def filterWith[A: ClassManifest](constructA: Int => A)(p: (T, A) => Boolean): RDD[T] = { - def iterF(context: TaskContext, iter: Iterator[T]): Iterator[T] = { - val a = constructA(context.partitionId) + mapPartitionsWithIndex((index, iter) => { + val a = constructA(index) iter.filter(t => p(t, a)) - } - new MapPartitionsWithContextRDD(this, sc.clean(iterF _), true) + }, preservesPartitioning = true) } /** @@ -546,19 +543,34 @@ abstract class RDD[T: ClassManifest]( * of elements in each partition. */ def zipPartitions[B: ClassManifest, V: ClassManifest] + (rdd2: RDD[B], preservesPartitioning: Boolean) + (f: (Iterator[T], Iterator[B]) => Iterator[V]): RDD[V] = + new ZippedPartitionsRDD2(sc, sc.clean(f), this, rdd2, preservesPartitioning) + + def zipPartitions[B: ClassManifest, V: ClassManifest] (rdd2: RDD[B]) (f: (Iterator[T], Iterator[B]) => Iterator[V]): RDD[V] = - new ZippedPartitionsRDD2(sc, sc.clean(f), this, rdd2) + new ZippedPartitionsRDD2(sc, sc.clean(f), this, rdd2, false) + + def zipPartitions[B: ClassManifest, C: ClassManifest, V: ClassManifest] + (rdd2: RDD[B], rdd3: RDD[C], preservesPartitioning: Boolean) + (f: (Iterator[T], Iterator[B], Iterator[C]) => Iterator[V]): RDD[V] = + new ZippedPartitionsRDD3(sc, sc.clean(f), this, rdd2, rdd3, preservesPartitioning) def zipPartitions[B: ClassManifest, C: ClassManifest, V: ClassManifest] (rdd2: RDD[B], rdd3: RDD[C]) (f: (Iterator[T], Iterator[B], Iterator[C]) => Iterator[V]): RDD[V] = - new ZippedPartitionsRDD3(sc, sc.clean(f), this, rdd2, rdd3) + new ZippedPartitionsRDD3(sc, sc.clean(f), this, rdd2, rdd3, false) + + def zipPartitions[B: ClassManifest, C: ClassManifest, D: ClassManifest, V: ClassManifest] + (rdd2: RDD[B], rdd3: RDD[C], rdd4: RDD[D], preservesPartitioning: Boolean) + (f: (Iterator[T], Iterator[B], Iterator[C], Iterator[D]) => Iterator[V]): RDD[V] = + new ZippedPartitionsRDD4(sc, sc.clean(f), this, rdd2, rdd3, rdd4, preservesPartitioning) def zipPartitions[B: ClassManifest, C: ClassManifest, D: ClassManifest, V: ClassManifest] (rdd2: RDD[B], rdd3: RDD[C], rdd4: RDD[D]) (f: (Iterator[T], Iterator[B], Iterator[C], Iterator[D]) => Iterator[V]): RDD[V] = - new ZippedPartitionsRDD4(sc, sc.clean(f), this, rdd2, rdd3, rdd4) + new ZippedPartitionsRDD4(sc, sc.clean(f), this, rdd2, rdd3, rdd4, false) // Actions (launch a job to return a value to the user program) @@ -928,7 +940,7 @@ abstract class RDD[T: ClassManifest]( private var storageLevel: StorageLevel = StorageLevel.NONE /** Record user function generating this RDD. */ - private[spark] val origin = Utils.formatSparkCallSite + @transient private[spark] val origin = Utils.formatSparkCallSite private[spark] def elementClassManifest: ClassManifest[T] = classManifest[T] @@ -943,7 +955,7 @@ abstract class RDD[T: ClassManifest]( def context = sc // Avoid handling doCheckpoint multiple times to prevent excessive recursion - private var doCheckpointCalled = false + @transient private var doCheckpointCalled = false /** * Performs the checkpointing of this RDD by saving this. It is called by the DAGScheduler diff --git a/core/src/main/scala/org/apache/spark/rdd/ZippedPartitionsRDD.scala b/core/src/main/scala/org/apache/spark/rdd/ZippedPartitionsRDD.scala index 31e6fd519d..a97d2a01c8 100644 --- a/core/src/main/scala/org/apache/spark/rdd/ZippedPartitionsRDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/ZippedPartitionsRDD.scala @@ -22,7 +22,8 @@ import java.io.{ObjectOutputStream, IOException} private[spark] class ZippedPartitionsPartition( idx: Int, - @transient rdds: Seq[RDD[_]]) + @transient rdds: Seq[RDD[_]], + @transient val preferredLocations: Seq[String]) extends Partition { override val index: Int = idx @@ -39,31 +40,29 @@ private[spark] class ZippedPartitionsPartition( abstract class ZippedPartitionsBaseRDD[V: ClassManifest]( sc: SparkContext, - var rdds: Seq[RDD[_]]) + var rdds: Seq[RDD[_]], + preservesPartitioning: Boolean = false) extends RDD[V](sc, rdds.map(x => new OneToOneDependency(x))) { + override val partitioner = + if (preservesPartitioning) firstParent[Any].partitioner else None + override def getPartitions: Array[Partition] = { - val sizes = rdds.map(x => x.partitions.size) - if (!sizes.forall(x => x == sizes(0))) { + val numParts = rdds.head.partitions.size + if (!rdds.forall(rdd => rdd.partitions.size == numParts)) { 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.tabulate[Partition](numParts) { i => + val prefs = rdds.map(rdd => rdd.preferredLocations(rdd.partitions(i))) + // Check whether there are any hosts that match all RDDs; otherwise return the union + val exactMatchLocations = prefs.reduce((x, y) => x.intersect(y)) + val locs = if (!exactMatchLocations.isEmpty) exactMatchLocations else prefs.flatten.distinct + new ZippedPartitionsPartition(i, rdds, locs) } - array } override def getPreferredLocations(s: Partition): Seq[String] = { - val parts = s.asInstanceOf[ZippedPartitionsPartition].partitions - val prefs = rdds.zip(parts).map { case (rdd, p) => rdd.preferredLocations(p) } - // Check whether there are any hosts that match all RDDs; otherwise return the union - val exactMatchLocations = prefs.reduce((x, y) => x.intersect(y)) - if (!exactMatchLocations.isEmpty) { - exactMatchLocations - } else { - prefs.flatten.distinct - } + s.asInstanceOf[ZippedPartitionsPartition].preferredLocations } override def clearDependencies() { @@ -76,8 +75,9 @@ 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)) { + var rdd2: RDD[B], + preservesPartitioning: Boolean = false) + extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2), preservesPartitioning) { override def compute(s: Partition, context: TaskContext): Iterator[V] = { val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions @@ -97,8 +97,9 @@ class ZippedPartitionsRDD3 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)) { + var rdd3: RDD[C], + preservesPartitioning: Boolean = false) + extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2, rdd3), preservesPartitioning) { override def compute(s: Partition, context: TaskContext): Iterator[V] = { val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions @@ -122,8 +123,9 @@ class ZippedPartitionsRDD4 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)) { + var rdd4: RDD[D], + preservesPartitioning: Boolean = false) + extends ZippedPartitionsBaseRDD[V](sc, List(rdd1, rdd2, rdd3, rdd4), preservesPartitioning) { override def compute(s: Partition, context: TaskContext): Iterator[V] = { val partitions = s.asInstanceOf[ZippedPartitionsPartition].partitions diff --git a/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala b/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala index 42bb3884c8..f9cd021dd3 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala @@ -110,30 +110,10 @@ class DAGScheduler( // resubmit failed stages val POLL_TIMEOUT = 10L - private val eventProcessActor: ActorRef = env.actorSystem.actorOf(Props(new Actor { - override def preStart() { - context.system.scheduler.schedule(RESUBMIT_TIMEOUT milliseconds, RESUBMIT_TIMEOUT milliseconds) { - if (failed.size > 0) { - resubmitFailedStages() - } - } - } - - /** - * The main event loop of the DAG scheduler, which waits for new-job / task-finished / failure - * events and responds by launching tasks. This runs in a dedicated thread and receives events - * via the eventQueue. - */ - def receive = { - case event: DAGSchedulerEvent => - logDebug("Got event of type " + event.getClass.getName) + // Warns the user if a stage contains a task with size greater than this value (in KB) + val TASK_SIZE_TO_WARN = 100 - if (!processEvent(event)) - submitWaitingStages() - else - context.stop(self) - } - })) + private var eventProcessActor: ActorRef = _ private[scheduler] val nextJobId = new AtomicInteger(0) @@ -141,9 +121,13 @@ class DAGScheduler( private val nextStageId = new AtomicInteger(0) - private val stageIdToStage = new TimeStampedHashMap[Int, Stage] + private[scheduler] val jobIdToStageIds = new TimeStampedHashMap[Int, HashSet[Int]] - private val shuffleToMapStage = new TimeStampedHashMap[Int, Stage] + private[scheduler] val stageIdToJobIds = new TimeStampedHashMap[Int, HashSet[Int]] + + private[scheduler] val stageIdToStage = new TimeStampedHashMap[Int, Stage] + + private[scheduler] val shuffleToMapStage = new TimeStampedHashMap[Int, Stage] private[spark] val stageToInfos = new TimeStampedHashMap[Stage, StageInfo] @@ -174,6 +158,56 @@ class DAGScheduler( val metadataCleaner = new MetadataCleaner(MetadataCleanerType.DAG_SCHEDULER, this.cleanup) + /** + * Starts the event processing actor. The actor has two responsibilities: + * + * 1. Waits for events like job submission, task finished, task failure etc., and calls + * [[org.apache.spark.scheduler.DAGScheduler.processEvent()]] to process them. + * 2. Schedules a periodical task to resubmit failed stages. + * + * NOTE: the actor cannot be started in the constructor, because the periodical task references + * some internal states of the enclosing [[org.apache.spark.scheduler.DAGScheduler]] object, thus + * cannot be scheduled until the [[org.apache.spark.scheduler.DAGScheduler]] is fully constructed. + */ + def start() { + eventProcessActor = env.actorSystem.actorOf(Props(new Actor { + /** + * A handle to the periodical task, used to cancel the task when the actor is stopped. + */ + var resubmissionTask: Cancellable = _ + + override def preStart() { + /** + * A message is sent to the actor itself periodically to remind the actor to resubmit failed + * stages. In this way, stage resubmission can be done within the same thread context of + * other event processing logic to avoid unnecessary synchronization overhead. + */ + resubmissionTask = context.system.scheduler.schedule( + RESUBMIT_TIMEOUT.millis, RESUBMIT_TIMEOUT.millis, self, ResubmitFailedStages) + } + + /** + * The main event loop of the DAG scheduler. + */ + def receive = { + case event: DAGSchedulerEvent => + logDebug("Got event of type " + event.getClass.getName) + + /** + * All events are forwarded to `processEvent()`, so that the event processing logic can + * easily tested without starting a dedicated actor. Please refer to `DAGSchedulerSuite` + * for details. + */ + if (!processEvent(event)) { + submitWaitingStages() + } else { + resubmissionTask.cancel() + context.stop(self) + } + } + })) + } + def addSparkListener(listener: SparkListener) { listenerBus.addListener(listener) } @@ -202,16 +236,16 @@ class DAGScheduler( shuffleToMapStage.get(shuffleDep.shuffleId) match { case Some(stage) => stage case None => - val stage = newStage(shuffleDep.rdd, shuffleDep.rdd.partitions.size, Some(shuffleDep), jobId) + val stage = newOrUsedStage(shuffleDep.rdd, shuffleDep.rdd.partitions.size, shuffleDep, jobId) shuffleToMapStage(shuffleDep.shuffleId) = stage stage } } /** - * Create a Stage for the given RDD, either as a shuffle map stage (for a ShuffleDependency) or - * as a result stage for the final RDD used directly in an action. The stage will also be - * associated with the provided jobId. + * Create a Stage -- either directly for use as a result stage, or as part of the (re)-creation + * of a shuffle map stage in newOrUsedStage. The stage will be associated with the provided + * jobId. Production of shuffle map stages should always use newOrUsedStage, not newStage directly. */ private def newStage( rdd: RDD[_], @@ -221,21 +255,45 @@ class DAGScheduler( callSite: Option[String] = None) : Stage = { - if (shuffleDep != None) { - // Kind of ugly: need to register RDDs with the cache and map output tracker here - // since we can't do it in the RDD constructor because # of partitions is unknown - logInfo("Registering RDD " + rdd.id + " (" + rdd.origin + ")") - mapOutputTracker.registerShuffle(shuffleDep.get.shuffleId, rdd.partitions.size) - } val id = nextStageId.getAndIncrement() val stage = new Stage(id, rdd, numTasks, shuffleDep, getParentStages(rdd, jobId), jobId, callSite) stageIdToStage(id) = stage + updateJobIdStageIdMaps(jobId, stage) stageToInfos(stage) = new StageInfo(stage) stage } /** + * Create a shuffle map Stage for the given RDD. The stage will also be associated with the + * provided jobId. If a stage for the shuffleId existed previously so that the shuffleId is + * present in the MapOutputTracker, then the number and location of available outputs are + * recovered from the MapOutputTracker + */ + private def newOrUsedStage( + rdd: RDD[_], + numTasks: Int, + shuffleDep: ShuffleDependency[_,_], + jobId: Int, + callSite: Option[String] = None) + : Stage = + { + val stage = newStage(rdd, numTasks, Some(shuffleDep), jobId, callSite) + if (mapOutputTracker.has(shuffleDep.shuffleId)) { + val serLocs = mapOutputTracker.getSerializedMapOutputStatuses(shuffleDep.shuffleId) + val locs = MapOutputTracker.deserializeMapStatuses(serLocs) + for (i <- 0 until locs.size) stage.outputLocs(i) = List(locs(i)) + stage.numAvailableOutputs = locs.size + } else { + // Kind of ugly: need to register RDDs with the cache and map output tracker here + // since we can't do it in the RDD constructor because # of partitions is unknown + logInfo("Registering RDD " + rdd.id + " (" + rdd.origin + ")") + mapOutputTracker.registerShuffle(shuffleDep.shuffleId, rdd.partitions.size) + } + stage + } + + /** * Get or create the list of parent stages for a given RDD. The stages will be assigned the * provided jobId if they haven't already been created with a lower jobId. */ @@ -287,6 +345,89 @@ class DAGScheduler( } /** + * Registers the given jobId among the jobs that need the given stage and + * all of that stage's ancestors. + */ + private def updateJobIdStageIdMaps(jobId: Int, stage: Stage) { + def updateJobIdStageIdMapsList(stages: List[Stage]) { + if (!stages.isEmpty) { + val s = stages.head + stageIdToJobIds.getOrElseUpdate(s.id, new HashSet[Int]()) += jobId + jobIdToStageIds.getOrElseUpdate(jobId, new HashSet[Int]()) += s.id + val parents = getParentStages(s.rdd, jobId) + val parentsWithoutThisJobId = parents.filter(p => !stageIdToJobIds.get(p.id).exists(_.contains(jobId))) + updateJobIdStageIdMapsList(parentsWithoutThisJobId ++ stages.tail) + } + } + updateJobIdStageIdMapsList(List(stage)) + } + + /** + * Removes job and any stages that are not needed by any other job. Returns the set of ids for stages that + * were removed. The associated tasks for those stages need to be cancelled if we got here via job cancellation. + */ + private def removeJobAndIndependentStages(jobId: Int): Set[Int] = { + val registeredStages = jobIdToStageIds(jobId) + val independentStages = new HashSet[Int]() + if (registeredStages.isEmpty) { + logError("No stages registered for job " + jobId) + } else { + stageIdToJobIds.filterKeys(stageId => registeredStages.contains(stageId)).foreach { + case (stageId, jobSet) => + if (!jobSet.contains(jobId)) { + logError("Job %d not registered for stage %d even though that stage was registered for the job" + .format(jobId, stageId)) + } else { + def removeStage(stageId: Int) { + // data structures based on Stage + stageIdToStage.get(stageId).foreach { s => + if (running.contains(s)) { + logDebug("Removing running stage %d".format(stageId)) + running -= s + } + stageToInfos -= s + shuffleToMapStage.keys.filter(shuffleToMapStage(_) == s).foreach(shuffleToMapStage.remove) + if (pendingTasks.contains(s) && !pendingTasks(s).isEmpty) { + logDebug("Removing pending status for stage %d".format(stageId)) + } + pendingTasks -= s + if (waiting.contains(s)) { + logDebug("Removing stage %d from waiting set.".format(stageId)) + waiting -= s + } + if (failed.contains(s)) { + logDebug("Removing stage %d from failed set.".format(stageId)) + failed -= s + } + } + // data structures based on StageId + stageIdToStage -= stageId + stageIdToJobIds -= stageId + + logDebug("After removal of stage %d, remaining stages = %d".format(stageId, stageIdToStage.size)) + } + + jobSet -= jobId + if (jobSet.isEmpty) { // no other job needs this stage + independentStages += stageId + removeStage(stageId) + } + } + } + } + independentStages.toSet + } + + private def jobIdToStageIdsRemove(jobId: Int) { + if (!jobIdToStageIds.contains(jobId)) { + logDebug("Trying to remove unregistered job " + jobId) + } else { + removeJobAndIndependentStages(jobId) + jobIdToStageIds -= jobId + } + } + + /** * Submit a job to the job scheduler and get a JobWaiter object back. The JobWaiter object * can be used to block until the the job finishes executing or can be used to cancel the job. */ @@ -375,13 +516,25 @@ class DAGScheduler( } /** - * Process one event retrieved from the event queue. - * Returns true if we should stop the event loop. + * Process one event retrieved from the event processing actor. + * + * @param event The event to be processed. + * @return `true` if we should stop the event loop. */ private[scheduler] def processEvent(event: DAGSchedulerEvent): Boolean = { event match { case JobSubmitted(jobId, rdd, func, partitions, allowLocal, callSite, listener, properties) => - val finalStage = newStage(rdd, partitions.size, None, jobId, Some(callSite)) + var finalStage: Stage = null + try { + // New stage creation at times and if its not protected, the scheduler thread is killed. + // e.g. it can fail when jobs are run on HadoopRDD whose underlying hdfs files have been deleted + finalStage = newStage(rdd, partitions.size, None, jobId, Some(callSite)) + } catch { + case e: Exception => + logWarning("Creating new stage failed due to exception - job: " + jobId, e) + listener.jobFailed(e) + return false + } val job = new ActiveJob(jobId, finalStage, func, partitions, callSite, listener, properties) clearCacheLocs() logInfo("Got job " + job.jobId + " (" + callSite + ") with " + partitions.length + @@ -391,37 +544,31 @@ class DAGScheduler( logInfo("Missing parents: " + getMissingParentStages(finalStage)) if (allowLocal && finalStage.parents.size == 0 && partitions.length == 1) { // Compute very short actions like first() or take() with no parent stages locally. + listenerBus.post(SparkListenerJobStart(job, Array(), properties)) runLocally(job) } else { - listenerBus.post(SparkListenerJobStart(job, properties)) idToActiveJob(jobId) = job activeJobs += job resultStageToJob(finalStage) = job + listenerBus.post(SparkListenerJobStart(job, jobIdToStageIds(jobId).toArray, properties)) submitStage(finalStage) } case JobCancelled(jobId) => - // Cancel a job: find all the running stages that are linked to this job, and cancel them. - running.filter(_.jobId == jobId).foreach { stage => - taskSched.cancelTasks(stage.id) - } + handleJobCancellation(jobId) case JobGroupCancelled(groupId) => // Cancel all jobs belonging to this job group. // First finds all active jobs with this group id, and then kill stages for them. - val jobIds = activeJobs.filter(groupId == _.properties.get(SparkContext.SPARK_JOB_GROUP_ID)) - .map(_.jobId) - if (!jobIds.isEmpty) { - running.filter(stage => jobIds.contains(stage.jobId)).foreach { stage => - taskSched.cancelTasks(stage.id) - } - } + val activeInGroup = activeJobs.filter(groupId == _.properties.get(SparkContext.SPARK_JOB_GROUP_ID)) + val jobIds = activeInGroup.map(_.jobId) + jobIds.foreach { handleJobCancellation } case AllJobsCancelled => // Cancel all running jobs. - running.foreach { stage => - taskSched.cancelTasks(stage.id) - } + running.map(_.jobId).foreach { handleJobCancellation } + activeJobs.clear() // These should already be empty by this point, + idToActiveJob.clear() // but just in case we lost track of some jobs... case ExecutorGained(execId, host) => handleExecutorGained(execId, host) @@ -430,6 +577,18 @@ class DAGScheduler( handleExecutorLost(execId) case BeginEvent(task, taskInfo) => + for ( + job <- idToActiveJob.get(task.stageId); + stage <- stageIdToStage.get(task.stageId); + stageInfo <- stageToInfos.get(stage) + ) { + if (taskInfo.serializedSize > TASK_SIZE_TO_WARN * 1024 && !stageInfo.emittedTaskSizeWarning) { + stageInfo.emittedTaskSizeWarning = true + logWarning(("Stage %d (%s) contains a task of very large " + + "size (%d KB). The maximum recommended task size is %d KB.").format( + task.stageId, stageInfo.name, taskInfo.serializedSize / 1024, TASK_SIZE_TO_WARN)) + } + } listenerBus.post(SparkListenerTaskStart(task, taskInfo)) case GettingResultEvent(task, taskInfo) => @@ -440,7 +599,12 @@ class DAGScheduler( handleTaskCompletion(completion) case TaskSetFailed(taskSet, reason) => - abortStage(stageIdToStage(taskSet.stageId), reason) + stageIdToStage.get(taskSet.stageId).foreach { abortStage(_, reason) } + + case ResubmitFailedStages => + if (failed.size > 0) { + resubmitFailedStages() + } case StopDAGScheduler => // Cancel any active jobs @@ -502,6 +666,7 @@ class DAGScheduler( // Broken out for easier testing in DAGSchedulerSuite. protected def runLocallyWithinThread(job: ActiveJob) { + var jobResult: JobResult = JobSucceeded try { SparkEnv.set(env) val rdd = job.finalStage.rdd @@ -516,31 +681,59 @@ class DAGScheduler( } } catch { case e: Exception => + jobResult = JobFailed(e, Some(job.finalStage)) job.listener.jobFailed(e) + } finally { + val s = job.finalStage + stageIdToJobIds -= s.id // clean up data structures that were populated for a local job, + stageIdToStage -= s.id // but that won't get cleaned up via the normal paths through + stageToInfos -= s // completion events or stage abort + jobIdToStageIds -= job.jobId + listenerBus.post(SparkListenerJobEnd(job, jobResult)) + } + } + + /** Finds the earliest-created active job that needs the stage */ + // TODO: Probably should actually find among the active jobs that need this + // stage the one with the highest priority (highest-priority pool, earliest created). + // That should take care of at least part of the priority inversion problem with + // cross-job dependencies. + private def activeJobForStage(stage: Stage): Option[Int] = { + if (stageIdToJobIds.contains(stage.id)) { + val jobsThatUseStage: Array[Int] = stageIdToJobIds(stage.id).toArray.sorted + jobsThatUseStage.find(idToActiveJob.contains(_)) + } else { + None } } /** Submits stage, but first recursively submits any missing parents. */ private def submitStage(stage: Stage) { - logDebug("submitStage(" + stage + ")") - if (!waiting(stage) && !running(stage) && !failed(stage)) { - val missing = getMissingParentStages(stage).sortBy(_.id) - logDebug("missing: " + missing) - if (missing == Nil) { - logInfo("Submitting " + stage + " (" + stage.rdd + "), which has no missing parents") - submitMissingTasks(stage) - running += stage - } else { - for (parent <- missing) { - submitStage(parent) + val jobId = activeJobForStage(stage) + if (jobId.isDefined) { + logDebug("submitStage(" + stage + ")") + if (!waiting(stage) && !running(stage) && !failed(stage)) { + val missing = getMissingParentStages(stage).sortBy(_.id) + logDebug("missing: " + missing) + if (missing == Nil) { + logInfo("Submitting " + stage + " (" + stage.rdd + "), which has no missing parents") + submitMissingTasks(stage, jobId.get) + running += stage + } else { + for (parent <- missing) { + submitStage(parent) + } + waiting += stage } - waiting += stage } + } else { + abortStage(stage, "No active job for stage " + stage.id) } } + /** Called when stage's parents are available and we can now do its task. */ - private def submitMissingTasks(stage: Stage) { + private def submitMissingTasks(stage: Stage, jobId: Int) { logDebug("submitMissingTasks(" + stage + ")") // Get our pending tasks and remember them in our pendingTasks entry val myPending = pendingTasks.getOrElseUpdate(stage, new HashSet) @@ -561,7 +754,7 @@ class DAGScheduler( } } - val properties = if (idToActiveJob.contains(stage.jobId)) { + val properties = if (idToActiveJob.contains(jobId)) { idToActiveJob(stage.jobId).properties } else { //this stage will be assigned to "default" pool @@ -643,6 +836,7 @@ class DAGScheduler( activeJobs -= job resultStageToJob -= stage markStageAsFinished(stage) + jobIdToStageIdsRemove(job.jobId) listenerBus.post(SparkListenerJobEnd(job, JobSucceeded)) } job.listener.taskSucceeded(rt.outputId, event.result) @@ -679,7 +873,7 @@ class DAGScheduler( changeEpoch = true) } clearCacheLocs() - if (stage.outputLocs.count(_ == Nil) != 0) { + if (stage.outputLocs.exists(_ == Nil)) { // Some tasks had failed; let's resubmit this stage // TODO: Lower-level scheduler should also deal with this logInfo("Resubmitting " + stage + " (" + stage.name + @@ -696,9 +890,12 @@ class DAGScheduler( } waiting --= newlyRunnable running ++= newlyRunnable - for (stage <- newlyRunnable.sortBy(_.id)) { + for { + stage <- newlyRunnable.sortBy(_.id) + jobId <- activeJobForStage(stage) + } { logInfo("Submitting " + stage + " (" + stage.rdd + "), which is now runnable") - submitMissingTasks(stage) + submitMissingTasks(stage, jobId) } } } @@ -782,21 +979,42 @@ class DAGScheduler( } } + private def handleJobCancellation(jobId: Int) { + if (!jobIdToStageIds.contains(jobId)) { + logDebug("Trying to cancel unregistered job " + jobId) + } else { + val independentStages = removeJobAndIndependentStages(jobId) + independentStages.foreach { taskSched.cancelTasks } + val error = new SparkException("Job %d cancelled".format(jobId)) + val job = idToActiveJob(jobId) + job.listener.jobFailed(error) + jobIdToStageIds -= jobId + activeJobs -= job + idToActiveJob -= jobId + listenerBus.post(SparkListenerJobEnd(job, JobFailed(error, Some(job.finalStage)))) + } + } + /** * Aborts all jobs depending on a particular Stage. This is called in response to a task set * being canceled by the TaskScheduler. Use taskSetFailed() to inject this event from outside. */ private def abortStage(failedStage: Stage, reason: String) { + if (!stageIdToStage.contains(failedStage.id)) { + // Skip all the actions if the stage has been removed. + return + } val dependentStages = resultStageToJob.keys.filter(x => stageDependsOn(x, failedStage)).toSeq stageToInfos(failedStage).completionTime = Some(System.currentTimeMillis()) for (resultStage <- dependentStages) { val job = resultStageToJob(resultStage) val error = new SparkException("Job aborted: " + reason) job.listener.jobFailed(error) - listenerBus.post(SparkListenerJobEnd(job, JobFailed(error, Some(failedStage)))) + jobIdToStageIdsRemove(job.jobId) idToActiveJob -= resultStage.jobId activeJobs -= job resultStageToJob -= resultStage + listenerBus.post(SparkListenerJobEnd(job, JobFailed(error, Some(failedStage)))) } if (dependentStages.isEmpty) { logInfo("Ignoring failure of " + failedStage + " because all jobs depending on it are done") @@ -867,25 +1085,24 @@ class DAGScheduler( } private def cleanup(cleanupTime: Long) { - var sizeBefore = stageIdToStage.size - stageIdToStage.clearOldValues(cleanupTime) - logInfo("stageIdToStage " + sizeBefore + " --> " + stageIdToStage.size) - - sizeBefore = shuffleToMapStage.size - shuffleToMapStage.clearOldValues(cleanupTime) - logInfo("shuffleToMapStage " + sizeBefore + " --> " + shuffleToMapStage.size) - - sizeBefore = pendingTasks.size - pendingTasks.clearOldValues(cleanupTime) - logInfo("pendingTasks " + sizeBefore + " --> " + pendingTasks.size) - - sizeBefore = stageToInfos.size - stageToInfos.clearOldValues(cleanupTime) - logInfo("stageToInfos " + sizeBefore + " --> " + stageToInfos.size) + Map( + "stageIdToStage" -> stageIdToStage, + "shuffleToMapStage" -> shuffleToMapStage, + "pendingTasks" -> pendingTasks, + "stageToInfos" -> stageToInfos, + "jobIdToStageIds" -> jobIdToStageIds, + "stageIdToJobIds" -> stageIdToJobIds). + foreach { case(s, t) => { + val sizeBefore = t.size + t.clearOldValues(cleanupTime) + logInfo("%s %d --> %d".format(s, sizeBefore, t.size)) + }} } def stop() { - eventProcessActor ! StopDAGScheduler + if (eventProcessActor != null) { + eventProcessActor ! StopDAGScheduler + } metadataCleaner.cancel() taskSched.stop() } diff --git a/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala b/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala index 708d221d60..add1187613 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala @@ -65,12 +65,13 @@ private[scheduler] case class CompletionEvent( taskMetrics: TaskMetrics) extends DAGSchedulerEvent -private[scheduler] -case class ExecutorGained(execId: String, host: String) extends DAGSchedulerEvent +private[scheduler] case class ExecutorGained(execId: String, host: String) extends DAGSchedulerEvent private[scheduler] case class ExecutorLost(execId: String) extends DAGSchedulerEvent private[scheduler] case class TaskSetFailed(taskSet: TaskSet, reason: String) extends DAGSchedulerEvent +private[scheduler] case object ResubmitFailedStages extends DAGSchedulerEvent + private[scheduler] case object StopDAGScheduler extends DAGSchedulerEvent diff --git a/core/src/main/scala/org/apache/spark/scheduler/JobWaiter.scala b/core/src/main/scala/org/apache/spark/scheduler/JobWaiter.scala index 58f238d8cf..b026f860a8 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/JobWaiter.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/JobWaiter.scala @@ -31,6 +31,7 @@ private[spark] class JobWaiter[T]( private var finishedTasks = 0 // Is the job as a whole finished (succeeded or failed)? + @volatile private var _jobFinished = totalTasks == 0 def jobFinished = _jobFinished diff --git a/core/src/main/scala/org/apache/spark/scheduler/ShuffleMapTask.scala b/core/src/main/scala/org/apache/spark/scheduler/ShuffleMapTask.scala index 1dc71a0428..0f2deb4bcb 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/ShuffleMapTask.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/ShuffleMapTask.scala @@ -167,6 +167,7 @@ private[spark] class ShuffleMapTask( var totalTime = 0L val compressedSizes: Array[Byte] = shuffle.writers.map { writer: BlockObjectWriter => writer.commit() + writer.close() val size = writer.fileSegment().length totalBytes += size totalTime += writer.timeWriting() @@ -184,14 +185,16 @@ private[spark] class ShuffleMapTask( } 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 (shuffle != null) { - shuffle.writers.foreach(_.revertPartialWrites()) + if (shuffle != null && shuffle.writers != null) { + for (writer <- shuffle.writers) { + writer.revertPartialWrites() + writer.close() + } } throw e } finally { // Release the writers back to the shuffle block manager. if (shuffle != null && shuffle.writers != null) { - shuffle.writers.foreach(_.close()) shuffle.releaseWriters(success) } // Execute the callbacks on task completion. diff --git a/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala b/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala index a35081f7b1..3841b5616d 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala @@ -37,7 +37,7 @@ case class SparkListenerTaskGettingResult( case class SparkListenerTaskEnd(task: Task[_], reason: TaskEndReason, taskInfo: TaskInfo, taskMetrics: TaskMetrics) extends SparkListenerEvents -case class SparkListenerJobStart(job: ActiveJob, properties: Properties = null) +case class SparkListenerJobStart(job: ActiveJob, stageIds: Array[Int], properties: Properties = null) extends SparkListenerEvents case class SparkListenerJobEnd(job: ActiveJob, jobResult: JobResult) diff --git a/core/src/main/scala/org/apache/spark/scheduler/StageInfo.scala b/core/src/main/scala/org/apache/spark/scheduler/StageInfo.scala index 93599dfdc8..e9f2198a00 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/StageInfo.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/StageInfo.scala @@ -33,4 +33,5 @@ class StageInfo( val name = stage.name val numPartitions = stage.numPartitions val numTasks = stage.numTasks + var emittedTaskSizeWarning = false } diff --git a/core/src/main/scala/org/apache/spark/scheduler/TaskInfo.scala b/core/src/main/scala/org/apache/spark/scheduler/TaskInfo.scala index 4bae26f3a6..3c22edd524 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/TaskInfo.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/TaskInfo.scala @@ -46,6 +46,8 @@ class TaskInfo( var failed = false + var serializedSize: Int = 0 + def markGettingResult(time: Long = System.currentTimeMillis) { gettingResultTime = time } diff --git a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala index c1e65a3c48..4d82430b97 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala @@ -100,8 +100,8 @@ private[spark] class ClusterScheduler(val sc: SparkContext) this.dagScheduler = dagScheduler } - def initialize(context: SchedulerBackend) { - backend = context + def initialize(backend: SchedulerBackend) { + this.backend = backend // temporarily set rootPool name to empty rootPool = new Pool("", schedulingMode, 0, 0) schedulableBuilder = { @@ -173,7 +173,9 @@ private[spark] class ClusterScheduler(val sc: SparkContext) backend.killTask(tid, execId) } } - tsm.error("Stage %d was cancelled".format(stageId)) + logInfo("Stage %d was cancelled".format(stageId)) + tsm.removeAllRunningTasks() + taskSetFinished(tsm) } } diff --git a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala index 4c5eca8537..94961790df 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala @@ -377,6 +377,7 @@ private[spark] class ClusterTaskSetManager( 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) + info.serializedSize = serializedTask.limit if (taskAttempts(index).size == 1) taskStarted(task,info) return Some(new TaskDescription(taskId, execId, taskName, index, serializedTask)) @@ -573,7 +574,7 @@ private[spark] class ClusterTaskSetManager( runningTasks = runningTasksSet.size } - private def removeAllRunningTasks() { + private[cluster] def removeAllRunningTasks() { val numRunningTasks = runningTasksSet.size runningTasksSet.clear() if (parent != null) { diff --git a/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala b/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala index 2699f0b33e..01e95162c0 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala @@ -74,7 +74,7 @@ class LocalActor(localScheduler: LocalScheduler, private var freeCores: Int) } } -private[spark] class LocalScheduler(threads: Int, val maxFailures: Int, val sc: SparkContext) +private[spark] class LocalScheduler(val threads: Int, val maxFailures: Int, val sc: SparkContext) extends TaskScheduler with ExecutorBackend with Logging { @@ -144,7 +144,8 @@ private[spark] class LocalScheduler(threads: Int, val maxFailures: Int, val sc: localActor ! KillTask(tid) } } - tsm.error("Stage %d was cancelled".format(stageId)) + logInfo("Stage %d was cancelled".format(stageId)) + taskSetFinished(tsm) } } @@ -192,17 +193,19 @@ private[spark] class LocalScheduler(threads: Int, val maxFailures: Int, val sc: synchronized { taskIdToTaskSetId.get(taskId) match { case Some(taskSetId) => - val taskSetManager = activeTaskSets(taskSetId) - taskSetTaskIds(taskSetId) -= taskId - - state match { - case TaskState.FINISHED => - taskSetManager.taskEnded(taskId, state, serializedData) - case TaskState.FAILED => - taskSetManager.taskFailed(taskId, state, serializedData) - case TaskState.KILLED => - taskSetManager.error("Task %d was killed".format(taskId)) - case _ => {} + val taskSetManager = activeTaskSets.get(taskSetId) + taskSetManager.foreach { tsm => + taskSetTaskIds(taskSetId) -= taskId + + state match { + case TaskState.FINISHED => + tsm.taskEnded(taskId, state, serializedData) + case TaskState.FAILED => + tsm.taskFailed(taskId, state, serializedData) + case TaskState.KILLED => + tsm.error("Task %d was killed".format(taskId)) + case _ => {} + } } case None => logInfo("Ignoring update from TID " + taskId + " because its task set is gone") diff --git a/core/src/main/scala/org/apache/spark/storage/BlockObjectWriter.scala b/core/src/main/scala/org/apache/spark/storage/BlockObjectWriter.scala index 469e68fed7..b4451fc7b8 100644 --- a/core/src/main/scala/org/apache/spark/storage/BlockObjectWriter.scala +++ b/core/src/main/scala/org/apache/spark/storage/BlockObjectWriter.scala @@ -93,6 +93,8 @@ class DiskBlockObjectWriter( def write(i: Int): Unit = callWithTiming(out.write(i)) override def write(b: Array[Byte]) = callWithTiming(out.write(b)) override def write(b: Array[Byte], off: Int, len: Int) = callWithTiming(out.write(b, off, len)) + override def close() = out.close() + override def flush() = out.flush() } private val syncWrites = System.getProperty("spark.shuffle.sync", "false").toBoolean diff --git a/core/src/main/scala/org/apache/spark/storage/ShuffleBlockManager.scala b/core/src/main/scala/org/apache/spark/storage/ShuffleBlockManager.scala index 2f1b049ce4..e828e1d1c5 100644 --- a/core/src/main/scala/org/apache/spark/storage/ShuffleBlockManager.scala +++ b/core/src/main/scala/org/apache/spark/storage/ShuffleBlockManager.scala @@ -62,7 +62,7 @@ class ShuffleBlockManager(blockManager: BlockManager) { // Turning off shuffle file consolidation causes all shuffle Blocks to get their own file. // TODO: Remove this once the shuffle file consolidation feature is stable. val consolidateShuffleFiles = - System.getProperty("spark.shuffle.consolidateFiles", "true").toBoolean + System.getProperty("spark.shuffle.consolidateFiles", "false").toBoolean private val bufferSize = System.getProperty("spark.shuffle.file.buffer.kb", "100").toInt * 1024 diff --git a/core/src/main/scala/org/apache/spark/storage/StorageLevel.scala b/core/src/main/scala/org/apache/spark/storage/StorageLevel.scala index 632ff047d1..b5596dffd3 100644 --- a/core/src/main/scala/org/apache/spark/storage/StorageLevel.scala +++ b/core/src/main/scala/org/apache/spark/storage/StorageLevel.scala @@ -101,7 +101,7 @@ class StorageLevel private( var result = "" result += (if (useDisk) "Disk " else "") result += (if (useMemory) "Memory " else "") - result += (if (deserialized) "Deserialized " else "Serialized") + result += (if (deserialized) "Deserialized " else "Serialized ") result += "%sx Replicated".format(replication) result } diff --git a/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala b/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala index c1c7aa70e6..69f9446bab 100644 --- a/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala +++ b/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala @@ -60,11 +60,13 @@ private[spark] class StagePage(parent: JobProgressUI) { var activeTime = 0L listener.stageIdToTasksActive(stageId).foreach(activeTime += _.timeRunning(now)) + val finishedTasks = listener.stageIdToTaskInfos(stageId).filter(_._1.finished) + val summary = <div> <ul class="unstyled"> <li> - <strong>CPU time: </strong> + <strong>Total duration across all tasks: </strong> {parent.formatDuration(listener.stageIdToTime.getOrElse(stageId, 0L) + activeTime)} </li> {if (hasShuffleRead) @@ -104,6 +106,33 @@ private[spark] class StagePage(parent: JobProgressUI) { val serviceQuantiles = "Duration" +: Distribution(serviceTimes).get.getQuantiles().map( ms => parent.formatDuration(ms.toLong)) + val gettingResultTimes = validTasks.map{case (info, metrics, exception) => + if (info.gettingResultTime > 0) { + (info.finishTime - info.gettingResultTime).toDouble + } else { + 0.0 + } + } + val gettingResultQuantiles = ("Time spent fetching task results" +: + Distribution(gettingResultTimes).get.getQuantiles().map( + millis => parent.formatDuration(millis.toLong))) + // The scheduler delay includes the network delay to send the task to the worker + // machine and to send back the result (but not the time to fetch the task result, + // if it needed to be fetched from the block manager on the worker). + val schedulerDelays = validTasks.map{case (info, metrics, exception) => + val totalExecutionTime = { + if (info.gettingResultTime > 0) { + (info.gettingResultTime - info.launchTime).toDouble + } else { + (info.finishTime - info.launchTime).toDouble + } + } + totalExecutionTime - metrics.get.executorRunTime + } + val schedulerDelayQuantiles = ("Scheduler delay" +: + Distribution(schedulerDelays).get.getQuantiles().map( + millis => parent.formatDuration(millis.toLong))) + def getQuantileCols(data: Seq[Double]) = Distribution(data).get.getQuantiles().map(d => Utils.bytesToString(d.toLong)) @@ -119,7 +148,10 @@ private[spark] class StagePage(parent: JobProgressUI) { } val shuffleWriteQuantiles = "Shuffle Write" +: getQuantileCols(shuffleWriteSizes) - val listings: Seq[Seq[String]] = Seq(serviceQuantiles, + val listings: Seq[Seq[String]] = Seq( + serviceQuantiles, + gettingResultQuantiles, + schedulerDelayQuantiles, if (hasShuffleRead) shuffleReadQuantiles else Nil, if (hasShuffleWrite) shuffleWriteQuantiles else Nil) @@ -133,7 +165,7 @@ private[spark] class StagePage(parent: JobProgressUI) { summary ++ <h4>Summary Metrics for {numCompleted} Completed Tasks</h4> ++ <div>{summaryTable.getOrElse("No tasks have reported metrics yet.")}</div> ++ - <h4>Tasks</h4> ++ taskTable; + <h4>Tasks</h4> ++ taskTable headerSparkPage(content, parent.sc, "Details for Stage %d".format(stageId), Stages) } @@ -152,21 +184,18 @@ private[spark] class StagePage(parent: JobProgressUI) { else metrics.map(m => parent.formatDuration(m.executorRunTime)).getOrElse("") val gcTime = metrics.map(m => m.jvmGCTime).getOrElse(0L) - var shuffleReadSortable: String = "" - var shuffleReadReadable: String = "" - if (shuffleRead) { - shuffleReadSortable = metrics.flatMap{m => m.shuffleReadMetrics}.map{s => s.remoteBytesRead}.toString() - shuffleReadReadable = metrics.flatMap{m => m.shuffleReadMetrics}.map{s => - Utils.bytesToString(s.remoteBytesRead)}.getOrElse("") - } + val maybeShuffleRead = metrics.flatMap{m => m.shuffleReadMetrics}.map{s => s.remoteBytesRead} + val shuffleReadSortable = maybeShuffleRead.map(_.toString).getOrElse("") + val shuffleReadReadable = maybeShuffleRead.map{Utils.bytesToString(_)}.getOrElse("") - var shuffleWriteSortable: String = "" - var shuffleWriteReadable: String = "" - if (shuffleWrite) { - shuffleWriteSortable = metrics.flatMap{m => m.shuffleWriteMetrics}.map{s => s.shuffleBytesWritten}.toString() - shuffleWriteReadable = metrics.flatMap{m => m.shuffleWriteMetrics}.map{s => - Utils.bytesToString(s.shuffleBytesWritten)}.getOrElse("") - } + val maybeShuffleWrite = metrics.flatMap{m => m.shuffleWriteMetrics}.map{s => s.shuffleBytesWritten} + val shuffleWriteSortable = maybeShuffleWrite.map(_.toString).getOrElse("") + val shuffleWriteReadable = maybeShuffleWrite.map{Utils.bytesToString(_)}.getOrElse("") + + val maybeWriteTime = metrics.flatMap{m => m.shuffleWriteMetrics}.map{s => s.shuffleWriteTime} + val writeTimeSortable = maybeWriteTime.map(_.toString).getOrElse("") + val writeTimeReadable = maybeWriteTime.map{ t => t / (1000 * 1000)}.map{ ms => + if (ms == 0) "" else parent.formatDuration(ms)}.getOrElse("") <tr> <td>{info.index}</td> @@ -187,8 +216,8 @@ private[spark] class StagePage(parent: JobProgressUI) { </td> }} {if (shuffleWrite) { - <td>{metrics.flatMap{m => m.shuffleWriteMetrics}.map{s => - parent.formatDuration(s.shuffleWriteTime / (1000 * 1000))}.getOrElse("")} + <td sorttable_customkey={writeTimeSortable}> + {writeTimeReadable} </td> <td sorttable_customkey={shuffleWriteSortable}> {shuffleWriteReadable} diff --git a/core/src/main/scala/org/apache/spark/util/AppendOnlyMap.scala b/core/src/main/scala/org/apache/spark/util/AppendOnlyMap.scala index f60deafc6f..8bb4ee3bfa 100644 --- a/core/src/main/scala/org/apache/spark/util/AppendOnlyMap.scala +++ b/core/src/main/scala/org/apache/spark/util/AppendOnlyMap.scala @@ -35,6 +35,7 @@ class AppendOnlyMap[K, V](initialCapacity: Int = 64) extends Iterable[(K, V)] wi private var capacity = nextPowerOf2(initialCapacity) private var mask = capacity - 1 private var curSize = 0 + private var growThreshold = LOAD_FACTOR * capacity // Holds keys and values in the same array for memory locality; specifically, the order of // elements is key0, value0, key1, value1, key2, value2, etc. @@ -56,7 +57,7 @@ class AppendOnlyMap[K, V](initialCapacity: Int = 64) extends Iterable[(K, V)] wi var i = 1 while (true) { val curKey = data(2 * pos) - if (k.eq(curKey) || k == curKey) { + if (k.eq(curKey) || k.equals(curKey)) { return data(2 * pos + 1).asInstanceOf[V] } else if (curKey.eq(null)) { return null.asInstanceOf[V] @@ -80,9 +81,23 @@ class AppendOnlyMap[K, V](initialCapacity: Int = 64) extends Iterable[(K, V)] wi haveNullValue = true return } - val isNewEntry = putInto(data, k, value.asInstanceOf[AnyRef]) - if (isNewEntry) { - incrementSize() + var pos = rehash(key.hashCode) & mask + var i = 1 + while (true) { + val curKey = data(2 * pos) + if (curKey.eq(null)) { + data(2 * pos) = k + data(2 * pos + 1) = value.asInstanceOf[AnyRef] + incrementSize() // Since we added a new key + return + } else if (k.eq(curKey) || k.equals(curKey)) { + data(2 * pos + 1) = value.asInstanceOf[AnyRef] + return + } else { + val delta = i + pos = (pos + delta) & mask + i += 1 + } } } @@ -104,7 +119,7 @@ class AppendOnlyMap[K, V](initialCapacity: Int = 64) extends Iterable[(K, V)] wi var i = 1 while (true) { val curKey = data(2 * pos) - if (k.eq(curKey) || k == curKey) { + if (k.eq(curKey) || k.equals(curKey)) { val newValue = updateFunc(true, data(2 * pos + 1).asInstanceOf[V]) data(2 * pos + 1) = newValue.asInstanceOf[AnyRef] return newValue @@ -161,45 +176,17 @@ class AppendOnlyMap[K, V](initialCapacity: Int = 64) extends Iterable[(K, V)] wi /** Increase table size by 1, rehashing if necessary */ private def incrementSize() { curSize += 1 - if (curSize > LOAD_FACTOR * capacity) { + if (curSize > growThreshold) { growTable() } } /** - * Re-hash a value to deal better with hash functions that don't differ - * in the lower bits, similar to java.util.HashMap + * Re-hash a value to deal better with hash functions that don't differ in the lower bits. + * We use the Murmur Hash 3 finalization step that's also used in fastutil. */ private def rehash(h: Int): Int = { - val r = h ^ (h >>> 20) ^ (h >>> 12) - r ^ (r >>> 7) ^ (r >>> 4) - } - - /** - * Put an entry into a table represented by data, returning true if - * this increases the size of the table or false otherwise. Assumes - * that "data" has at least one empty slot. - */ - private def putInto(data: Array[AnyRef], key: AnyRef, value: AnyRef): Boolean = { - val mask = (data.length / 2) - 1 - var pos = rehash(key.hashCode) & mask - var i = 1 - while (true) { - val curKey = data(2 * pos) - if (curKey.eq(null)) { - data(2 * pos) = key - data(2 * pos + 1) = value.asInstanceOf[AnyRef] - return true - } else if (curKey.eq(key) || curKey == key) { - data(2 * pos + 1) = value.asInstanceOf[AnyRef] - return false - } else { - val delta = i - pos = (pos + delta) & mask - i += 1 - } - } - return false // Never reached but needed to keep compiler happy + it.unimi.dsi.fastutil.HashCommon.murmurHash3(h) } /** Double the table's size and re-hash everything */ @@ -211,16 +198,36 @@ class AppendOnlyMap[K, V](initialCapacity: Int = 64) extends Iterable[(K, V)] wi throw new Exception("Can't make capacity bigger than 2^29 elements") } val newData = new Array[AnyRef](2 * newCapacity) - var pos = 0 - while (pos < capacity) { - if (!data(2 * pos).eq(null)) { - putInto(newData, data(2 * pos), data(2 * pos + 1)) + val newMask = newCapacity - 1 + // Insert all our old values into the new array. Note that because our old keys are + // unique, there's no need to check for equality here when we insert. + var oldPos = 0 + while (oldPos < capacity) { + if (!data(2 * oldPos).eq(null)) { + val key = data(2 * oldPos) + val value = data(2 * oldPos + 1) + var newPos = rehash(key.hashCode) & newMask + var i = 1 + var keepGoing = true + while (keepGoing) { + val curKey = newData(2 * newPos) + if (curKey.eq(null)) { + newData(2 * newPos) = key + newData(2 * newPos + 1) = value + keepGoing = false + } else { + val delta = i + newPos = (newPos + delta) & newMask + i += 1 + } + } } - pos += 1 + oldPos += 1 } data = newData capacity = newCapacity - mask = newCapacity - 1 + mask = newMask + growThreshold = LOAD_FACTOR * newCapacity } private def nextPowerOf2(n: Int): Int = { diff --git a/core/src/main/scala/org/apache/spark/util/Utils.scala b/core/src/main/scala/org/apache/spark/util/Utils.scala index fe932d8ede..a79e64e810 100644 --- a/core/src/main/scala/org/apache/spark/util/Utils.scala +++ b/core/src/main/scala/org/apache/spark/util/Utils.scala @@ -823,4 +823,28 @@ private[spark] object Utils extends Logging { return System.getProperties().clone() .asInstanceOf[java.util.Properties].toMap[String, String] } + + /** + * Method executed for repeating a task for side effects. + * Unlike a for comprehension, it permits JVM JIT optimization + */ + def times(numIters: Int)(f: => Unit): Unit = { + var i = 0 + while (i < numIters) { + f + i += 1 + } + } + + /** + * Timing method based on iterations that permit JVM JIT optimization. + * @param numIters number of iterations + * @param f function to be executed + */ + def timeIt(numIters: Int)(f: => Unit): Long = { + val start = System.currentTimeMillis + times(numIters)(f) + System.currentTimeMillis - start + } + } diff --git a/core/src/main/scala/org/apache/spark/util/XORShiftRandom.scala b/core/src/main/scala/org/apache/spark/util/XORShiftRandom.scala new file mode 100644 index 0000000000..e9907e6c85 --- /dev/null +++ b/core/src/main/scala/org/apache/spark/util/XORShiftRandom.scala @@ -0,0 +1,94 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.util + +import java.util.{Random => JavaRandom} +import org.apache.spark.util.Utils.timeIt + +/** + * This class implements a XORShift random number generator algorithm + * Source: + * Marsaglia, G. (2003). Xorshift RNGs. Journal of Statistical Software, Vol. 8, Issue 14. + * @see <a href="http://www.jstatsoft.org/v08/i14/paper">Paper</a> + * This implementation is approximately 3.5 times faster than + * {@link java.util.Random java.util.Random}, partly because of the algorithm, but also due + * to renouncing thread safety. JDK's implementation uses an AtomicLong seed, this class + * uses a regular Long. We can forgo thread safety since we use a new instance of the RNG + * for each thread. + */ +private[spark] class XORShiftRandom(init: Long) extends JavaRandom(init) { + + def this() = this(System.nanoTime) + + private var seed = init + + // we need to just override next - this will be called by nextInt, nextDouble, + // nextGaussian, nextLong, etc. + override protected def next(bits: Int): Int = { + var nextSeed = seed ^ (seed << 21) + nextSeed ^= (nextSeed >>> 35) + nextSeed ^= (nextSeed << 4) + seed = nextSeed + (nextSeed & ((1L << bits) -1)).asInstanceOf[Int] + } +} + +/** Contains benchmark method and main method to run benchmark of the RNG */ +private[spark] object XORShiftRandom { + + /** + * Main method for running benchmark + * @param args takes one argument - the number of random numbers to generate + */ + def main(args: Array[String]): Unit = { + if (args.length != 1) { + println("Benchmark of XORShiftRandom vis-a-vis java.util.Random") + println("Usage: XORShiftRandom number_of_random_numbers_to_generate") + System.exit(1) + } + println(benchmark(args(0).toInt)) + } + + /** + * @param numIters Number of random numbers to generate while running the benchmark + * @return Map of execution times for {@link java.util.Random java.util.Random} + * and XORShift + */ + def benchmark(numIters: Int) = { + + val seed = 1L + val million = 1e6.toInt + val javaRand = new JavaRandom(seed) + val xorRand = new XORShiftRandom(seed) + + // this is just to warm up the JIT - we're not timing anything + timeIt(1e6.toInt) { + javaRand.nextInt() + xorRand.nextInt() + } + + val iters = timeIt(numIters)(_) + + /* Return results as a map instead of just printing to screen + in case the user wants to do something with them */ + Map("javaTime" -> iters {javaRand.nextInt()}, + "xorTime" -> iters {xorRand.nextInt()}) + + } + +}
\ No newline at end of file diff --git a/core/src/main/scala/org/apache/spark/util/collection/OpenHashSet.scala b/core/src/main/scala/org/apache/spark/util/collection/OpenHashSet.scala index 4592e4f939..40986e3731 100644 --- a/core/src/main/scala/org/apache/spark/util/collection/OpenHashSet.scala +++ b/core/src/main/scala/org/apache/spark/util/collection/OpenHashSet.scala @@ -79,6 +79,7 @@ class OpenHashSet[@specialized(Long, Int) T: ClassManifest]( protected var _capacity = nextPowerOf2(initialCapacity) protected var _mask = _capacity - 1 protected var _size = 0 + protected var _growThreshold = (loadFactor * _capacity).toInt protected var _bitset = new BitSet(_capacity) @@ -115,7 +116,29 @@ class OpenHashSet[@specialized(Long, Int) T: ClassManifest]( * @return The position where the key is placed, plus the highest order bit is set if the key * exists previously. */ - def addWithoutResize(k: T): Int = putInto(_bitset, _data, k) + def addWithoutResize(k: T): Int = { + var pos = hashcode(hasher.hash(k)) & _mask + var i = 1 + while (true) { + if (!_bitset.get(pos)) { + // This is a new key. + _data(pos) = k + _bitset.set(pos) + _size += 1 + return pos | NONEXISTENCE_MASK + } else if (_data(pos) == k) { + // Found an existing key. + return pos + } else { + val delta = i + pos = (pos + delta) & _mask + i += 1 + } + } + // Never reached here + assert(INVALID_POS != INVALID_POS) + INVALID_POS + } /** * Rehash the set if it is overloaded. @@ -126,7 +149,7 @@ class OpenHashSet[@specialized(Long, Int) T: ClassManifest]( * to a new position (in the new data array). */ def rehashIfNeeded(k: T, allocateFunc: (Int) => Unit, moveFunc: (Int, Int) => Unit) { - if (_size > loadFactor * _capacity) { + if (_size > _growThreshold) { rehash(k, allocateFunc, moveFunc) } } @@ -161,37 +184,6 @@ class OpenHashSet[@specialized(Long, Int) T: ClassManifest]( def nextPos(fromPos: Int): Int = _bitset.nextSetBit(fromPos) /** - * Put an entry into the set. Return the position where the key is placed. In addition, the - * highest bit in the returned position is set if the key exists prior to this put. - * - * This function assumes the data array has at least one empty slot. - */ - private def putInto(bitset: BitSet, data: Array[T], k: T): Int = { - val mask = data.length - 1 - var pos = hashcode(hasher.hash(k)) & mask - var i = 1 - while (true) { - if (!bitset.get(pos)) { - // This is a new key. - data(pos) = k - bitset.set(pos) - _size += 1 - return pos | NONEXISTENCE_MASK - } else if (data(pos) == k) { - // Found an existing key. - return pos - } else { - val delta = i - pos = (pos + delta) & mask - i += 1 - } - } - // Never reached here - assert(INVALID_POS != INVALID_POS) - INVALID_POS - } - - /** * Double the table's size and re-hash everything. We are not really using k, but it is declared * so Scala compiler can specialize this method (which leads to calling the specialized version * of putInto). @@ -204,34 +196,49 @@ class OpenHashSet[@specialized(Long, Int) T: ClassManifest]( */ private def rehash(k: T, allocateFunc: (Int) => Unit, moveFunc: (Int, Int) => Unit) { val newCapacity = _capacity * 2 - require(newCapacity <= (1 << 29), "Can't make capacity bigger than 2^29 elements") - allocateFunc(newCapacity) - val newData = new Array[T](newCapacity) val newBitset = new BitSet(newCapacity) - var pos = 0 - _size = 0 - while (pos < _capacity) { - if (_bitset.get(pos)) { - val newPos = putInto(newBitset, newData, _data(pos)) - moveFunc(pos, newPos & POSITION_MASK) + val newData = new Array[T](newCapacity) + val newMask = newCapacity - 1 + + var oldPos = 0 + while (oldPos < capacity) { + if (_bitset.get(oldPos)) { + val key = _data(oldPos) + var newPos = hashcode(hasher.hash(key)) & newMask + var i = 1 + var keepGoing = true + // No need to check for equality here when we insert so this has one less if branch than + // the similar code path in addWithoutResize. + while (keepGoing) { + if (!newBitset.get(newPos)) { + // Inserting the key at newPos + newData(newPos) = key + newBitset.set(newPos) + moveFunc(oldPos, newPos) + keepGoing = false + } else { + val delta = i + newPos = (newPos + delta) & newMask + i += 1 + } + } } - pos += 1 + oldPos += 1 } + _bitset = newBitset _data = newData _capacity = newCapacity - _mask = newCapacity - 1 + _mask = newMask + _growThreshold = (loadFactor * newCapacity).toInt } /** - * Re-hash a value to deal better with hash functions that don't differ - * in the lower bits, similar to java.util.HashMap + * Re-hash a value to deal better with hash functions that don't differ in the lower bits. + * We use the Murmur Hash 3 finalization step that's also used in fastutil. */ - private def hashcode(h: Int): Int = { - val r = h ^ (h >>> 20) ^ (h >>> 12) - r ^ (r >>> 7) ^ (r >>> 4) - } + private def hashcode(h: Int): Int = it.unimi.dsi.fastutil.HashCommon.murmurHash3(h) private def nextPowerOf2(n: Int): Int = { val highBit = Integer.highestOneBit(n) diff --git a/core/src/test/scala/org/apache/spark/CheckpointSuite.scala b/core/src/test/scala/org/apache/spark/CheckpointSuite.scala index f26c44d3e7..d2226aa5a5 100644 --- a/core/src/test/scala/org/apache/spark/CheckpointSuite.scala +++ b/core/src/test/scala/org/apache/spark/CheckpointSuite.scala @@ -62,8 +62,6 @@ class CheckpointSuite extends FunSuite with LocalSparkContext with Logging { testCheckpointing(_.sample(false, 0.5, 0)) testCheckpointing(_.glom()) testCheckpointing(_.mapPartitions(_.map(_.toString))) - testCheckpointing(r => new MapPartitionsWithContextRDD(r, - (context: TaskContext, iter: Iterator[Int]) => iter.map(_.toString), false )) testCheckpointing(_.map(x => (x % 2, 1)).reduceByKey(_ + _).mapValues(_.toString)) testCheckpointing(_.map(x => (x % 2, 1)).reduceByKey(_ + _).flatMapValues(x => 1 to x)) testCheckpointing(_.pipe(Seq("cat"))) diff --git a/core/src/test/scala/org/apache/spark/JavaAPISuite.java b/core/src/test/scala/org/apache/spark/JavaAPISuite.java index 352036f182..4234f6eac7 100644 --- a/core/src/test/scala/org/apache/spark/JavaAPISuite.java +++ b/core/src/test/scala/org/apache/spark/JavaAPISuite.java @@ -365,6 +365,20 @@ public class JavaAPISuite implements Serializable { } @Test + public void javaDoubleRDDHistoGram() { + JavaDoubleRDD rdd = sc.parallelizeDoubles(Arrays.asList(1.0, 2.0, 3.0, 4.0)); + // Test using generated buckets + Tuple2<double[], long[]> results = rdd.histogram(2); + double[] expected_buckets = {1.0, 2.5, 4.0}; + long[] expected_counts = {2, 2}; + Assert.assertArrayEquals(expected_buckets, results._1, 0.1); + Assert.assertArrayEquals(expected_counts, results._2); + // Test with provided buckets + long[] histogram = rdd.histogram(expected_buckets); + Assert.assertArrayEquals(expected_counts, histogram); + } + + @Test public void map() { JavaRDD<Integer> rdd = sc.parallelize(Arrays.asList(1, 2, 3, 4, 5)); JavaDoubleRDD doubles = rdd.map(new DoubleFunction<Integer>() { diff --git a/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala b/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala index d8a0e983b2..1121e06e2e 100644 --- a/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala +++ b/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala @@ -114,7 +114,7 @@ class JobCancellationSuite extends FunSuite with ShouldMatchers with BeforeAndAf // Once A is cancelled, job B should finish fairly quickly. assert(jobB.get() === 100) } - +/* test("two jobs sharing the same stage") { // sem1: make sure cancel is issued after some tasks are launched // sem2: make sure the first stage is not finished until cancel is issued @@ -148,7 +148,7 @@ class JobCancellationSuite extends FunSuite with ShouldMatchers with BeforeAndAf intercept[SparkException] { f1.get() } intercept[SparkException] { f2.get() } } - + */ def testCount() { // Cancel before launching any tasks { diff --git a/core/src/test/scala/org/apache/spark/LocalSparkContext.scala b/core/src/test/scala/org/apache/spark/LocalSparkContext.scala index 459e257d79..8dd5786da6 100644 --- a/core/src/test/scala/org/apache/spark/LocalSparkContext.scala +++ b/core/src/test/scala/org/apache/spark/LocalSparkContext.scala @@ -30,7 +30,7 @@ trait LocalSparkContext extends BeforeAndAfterEach with BeforeAndAfterAll { self @transient var sc: SparkContext = _ override def beforeAll() { - InternalLoggerFactory.setDefaultFactory(new Slf4JLoggerFactory()); + InternalLoggerFactory.setDefaultFactory(new Slf4JLoggerFactory()) super.beforeAll() } diff --git a/core/src/test/scala/org/apache/spark/PartitionPruningRDDSuite.scala b/core/src/test/scala/org/apache/spark/PartitionPruningRDDSuite.scala deleted file mode 100644 index 21f16ef2c6..0000000000 --- a/core/src/test/scala/org/apache/spark/PartitionPruningRDDSuite.scala +++ /dev/null @@ -1,45 +0,0 @@ -/* - * Licensed to the Apache Software Foundation (ASF) under one or more - * contributor license agreements. See the NOTICE file distributed with - * this work for additional information regarding copyright ownership. - * The ASF licenses this file to You under the Apache License, Version 2.0 - * (the "License"); you may not use this file except in compliance with - * the License. You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ - -package org.apache.spark - -import org.scalatest.FunSuite -import org.apache.spark.SparkContext._ -import org.apache.spark.rdd.{RDD, PartitionPruningRDD} - - -class PartitionPruningRDDSuite extends FunSuite with SharedSparkContext { - - test("Pruned Partitions inherit locality prefs correctly") { - class TestPartition(i: Int) extends Partition { - def index = i - } - val rdd = new RDD[Int](sc, Nil) { - override protected def getPartitions = { - Array[Partition]( - new TestPartition(1), - new TestPartition(2), - new TestPartition(3)) - } - def compute(split: Partition, context: TaskContext) = {Iterator()} - } - val prunedRDD = PartitionPruningRDD.create(rdd, {x => if (x==2) true else false}) - val p = prunedRDD.partitions(0) - assert(p.index == 2) - assert(prunedRDD.partitions.length == 1) - } -} diff --git a/core/src/test/scala/org/apache/spark/PartitioningSuite.scala b/core/src/test/scala/org/apache/spark/PartitioningSuite.scala index 7d938917f2..1374d01774 100644 --- a/core/src/test/scala/org/apache/spark/PartitioningSuite.scala +++ b/core/src/test/scala/org/apache/spark/PartitioningSuite.scala @@ -142,11 +142,11 @@ class PartitioningSuite extends FunSuite with SharedSparkContext { .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); + 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/org/apache/spark/SparkContextSchedulerCreationSuite.scala b/core/src/test/scala/org/apache/spark/SparkContextSchedulerCreationSuite.scala new file mode 100644 index 0000000000..151af0d213 --- /dev/null +++ b/core/src/test/scala/org/apache/spark/SparkContextSchedulerCreationSuite.scala @@ -0,0 +1,140 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark + +import org.scalatest.{FunSuite, PrivateMethodTester} + +import org.apache.spark.scheduler.TaskScheduler +import org.apache.spark.scheduler.cluster.{ClusterScheduler, SimrSchedulerBackend, SparkDeploySchedulerBackend} +import org.apache.spark.scheduler.cluster.mesos.{CoarseMesosSchedulerBackend, MesosSchedulerBackend} +import org.apache.spark.scheduler.local.LocalScheduler + +class SparkContextSchedulerCreationSuite + extends FunSuite with PrivateMethodTester with LocalSparkContext with Logging { + + def createTaskScheduler(master: String): TaskScheduler = { + // Create local SparkContext to setup a SparkEnv. We don't actually want to start() the + // real schedulers, so we don't want to create a full SparkContext with the desired scheduler. + sc = new SparkContext("local", "test") + val createTaskSchedulerMethod = PrivateMethod[TaskScheduler]('createTaskScheduler) + SparkContext invokePrivate createTaskSchedulerMethod(sc, master, "test") + } + + test("bad-master") { + val e = intercept[SparkException] { + createTaskScheduler("localhost:1234") + } + assert(e.getMessage.contains("Could not parse Master URL")) + } + + test("local") { + createTaskScheduler("local") match { + case s: LocalScheduler => + assert(s.threads === 1) + assert(s.maxFailures === 0) + case _ => fail() + } + } + + test("local-n") { + createTaskScheduler("local[5]") match { + case s: LocalScheduler => + assert(s.threads === 5) + assert(s.maxFailures === 0) + case _ => fail() + } + } + + test("local-n-failures") { + createTaskScheduler("local[4, 2]") match { + case s: LocalScheduler => + assert(s.threads === 4) + assert(s.maxFailures === 2) + case _ => fail() + } + } + + test("simr") { + createTaskScheduler("simr://uri") match { + case s: ClusterScheduler => + assert(s.backend.isInstanceOf[SimrSchedulerBackend]) + case _ => fail() + } + } + + test("local-cluster") { + createTaskScheduler("local-cluster[3, 14, 512]") match { + case s: ClusterScheduler => + assert(s.backend.isInstanceOf[SparkDeploySchedulerBackend]) + case _ => fail() + } + } + + def testYarn(master: String, expectedClassName: String) { + try { + createTaskScheduler(master) match { + case s: ClusterScheduler => + assert(s.getClass === Class.forName(expectedClassName)) + case _ => fail() + } + } catch { + case e: SparkException => + assert(e.getMessage.contains("YARN mode not available")) + logWarning("YARN not available, could not test actual YARN scheduler creation") + case e: Throwable => fail(e) + } + } + + test("yarn-standalone") { + testYarn("yarn-standalone", "org.apache.spark.scheduler.cluster.YarnClusterScheduler") + } + + test("yarn-client") { + testYarn("yarn-client", "org.apache.spark.scheduler.cluster.YarnClientClusterScheduler") + } + + def testMesos(master: String, expectedClass: Class[_]) { + try { + createTaskScheduler(master) match { + case s: ClusterScheduler => + assert(s.backend.getClass === expectedClass) + case _ => fail() + } + } catch { + case e: UnsatisfiedLinkError => + assert(e.getMessage.contains("no mesos in")) + logWarning("Mesos not available, could not test actual Mesos scheduler creation") + case e: Throwable => fail(e) + } + } + + test("mesos fine-grained") { + System.setProperty("spark.mesos.coarse", "false") + testMesos("mesos://localhost:1234", classOf[MesosSchedulerBackend]) + } + + test("mesos coarse-grained") { + System.setProperty("spark.mesos.coarse", "true") + testMesos("mesos://localhost:1234", classOf[CoarseMesosSchedulerBackend]) + } + + test("mesos with zookeeper") { + System.setProperty("spark.mesos.coarse", "false") + testMesos("zk://localhost:1234,localhost:2345", classOf[MesosSchedulerBackend]) + } +} diff --git a/core/src/test/scala/org/apache/spark/rdd/AsyncRDDActionsSuite.scala b/core/src/test/scala/org/apache/spark/rdd/AsyncRDDActionsSuite.scala index da032b17d9..0d4c10db8e 100644 --- a/core/src/test/scala/org/apache/spark/rdd/AsyncRDDActionsSuite.scala +++ b/core/src/test/scala/org/apache/spark/rdd/AsyncRDDActionsSuite.scala @@ -19,6 +19,8 @@ package org.apache.spark.rdd import java.util.concurrent.Semaphore +import scala.concurrent.{Await, TimeoutException} +import scala.concurrent.duration.Duration import scala.concurrent.ExecutionContext.Implicits.global import org.scalatest.{BeforeAndAfterAll, FunSuite} @@ -173,4 +175,28 @@ class AsyncRDDActionsSuite extends FunSuite with BeforeAndAfterAll with Timeouts sem.acquire(2) } } + + /** + * Awaiting FutureAction results + */ + test("FutureAction result, infinite wait") { + val f = sc.parallelize(1 to 100, 4) + .countAsync() + assert(Await.result(f, Duration.Inf) === 100) + } + + test("FutureAction result, finite wait") { + val f = sc.parallelize(1 to 100, 4) + .countAsync() + assert(Await.result(f, Duration(30, "seconds")) === 100) + } + + test("FutureAction result, timeout") { + val f = sc.parallelize(1 to 100, 4) + .mapPartitions(itr => { Thread.sleep(20); itr }) + .countAsync() + intercept[TimeoutException] { + Await.result(f, Duration(20, "milliseconds")) + } + } } diff --git a/core/src/test/scala/org/apache/spark/rdd/DoubleRDDSuite.scala b/core/src/test/scala/org/apache/spark/rdd/DoubleRDDSuite.scala new file mode 100644 index 0000000000..7f50a5a47c --- /dev/null +++ b/core/src/test/scala/org/apache/spark/rdd/DoubleRDDSuite.scala @@ -0,0 +1,271 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.rdd + +import scala.math.abs +import scala.collection.mutable.ArrayBuffer + +import org.scalatest.FunSuite + +import org.apache.spark.SparkContext._ +import org.apache.spark.rdd._ +import org.apache.spark._ + +class DoubleRDDSuite extends FunSuite with SharedSparkContext { + // Verify tests on the histogram functionality. We test with both evenly + // and non-evenly spaced buckets as the bucket lookup function changes. + test("WorksOnEmpty") { + // Make sure that it works on an empty input + val rdd: RDD[Double] = sc.parallelize(Seq()) + val buckets = Array(0.0, 10.0) + val histogramResults = rdd.histogram(buckets) + val histogramResults2 = rdd.histogram(buckets, true) + val expectedHistogramResults = Array(0) + assert(histogramResults === expectedHistogramResults) + assert(histogramResults2 === expectedHistogramResults) + } + + test("WorksWithOutOfRangeWithOneBucket") { + // Verify that if all of the elements are out of range the counts are zero + val rdd = sc.parallelize(Seq(10.01, -0.01)) + val buckets = Array(0.0, 10.0) + val histogramResults = rdd.histogram(buckets) + val histogramResults2 = rdd.histogram(buckets, true) + val expectedHistogramResults = Array(0) + assert(histogramResults === expectedHistogramResults) + assert(histogramResults2 === expectedHistogramResults) + } + + test("WorksInRangeWithOneBucket") { + // Verify the basic case of one bucket and all elements in that bucket works + val rdd = sc.parallelize(Seq(1, 2, 3, 4)) + val buckets = Array(0.0, 10.0) + val histogramResults = rdd.histogram(buckets) + val histogramResults2 = rdd.histogram(buckets, true) + val expectedHistogramResults = Array(4) + assert(histogramResults === expectedHistogramResults) + assert(histogramResults2 === expectedHistogramResults) + } + + test("WorksInRangeWithOneBucketExactMatch") { + // Verify the basic case of one bucket and all elements in that bucket works + val rdd = sc.parallelize(Seq(1, 2, 3, 4)) + val buckets = Array(1.0, 4.0) + val histogramResults = rdd.histogram(buckets) + val histogramResults2 = rdd.histogram(buckets, true) + val expectedHistogramResults = Array(4) + assert(histogramResults === expectedHistogramResults) + assert(histogramResults2 === expectedHistogramResults) + } + + test("WorksWithOutOfRangeWithTwoBuckets") { + // Verify that out of range works with two buckets + val rdd = sc.parallelize(Seq(10.01, -0.01)) + val buckets = Array(0.0, 5.0, 10.0) + val histogramResults = rdd.histogram(buckets) + val histogramResults2 = rdd.histogram(buckets, true) + val expectedHistogramResults = Array(0, 0) + assert(histogramResults === expectedHistogramResults) + assert(histogramResults2 === expectedHistogramResults) + } + + test("WorksWithOutOfRangeWithTwoUnEvenBuckets") { + // Verify that out of range works with two un even buckets + val rdd = sc.parallelize(Seq(10.01, -0.01)) + val buckets = Array(0.0, 4.0, 10.0) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(0, 0) + assert(histogramResults === expectedHistogramResults) + } + + test("WorksInRangeWithTwoBuckets") { + // Make sure that it works with two equally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(1, 2, 3, 5, 6)) + val buckets = Array(0.0, 5.0, 10.0) + val histogramResults = rdd.histogram(buckets) + val histogramResults2 = rdd.histogram(buckets, true) + val expectedHistogramResults = Array(3, 2) + assert(histogramResults === expectedHistogramResults) + assert(histogramResults2 === expectedHistogramResults) + } + + test("WorksInRangeWithTwoBucketsAndNaN") { + // Make sure that it works with two equally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(1, 2, 3, 5, 6, Double.NaN)) + val buckets = Array(0.0, 5.0, 10.0) + val histogramResults = rdd.histogram(buckets) + val histogramResults2 = rdd.histogram(buckets, true) + val expectedHistogramResults = Array(3, 2) + assert(histogramResults === expectedHistogramResults) + assert(histogramResults2 === expectedHistogramResults) + } + + test("WorksInRangeWithTwoUnevenBuckets") { + // Make sure that it works with two unequally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(1, 2, 3, 5, 6)) + val buckets = Array(0.0, 5.0, 11.0) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(3, 2) + assert(histogramResults === expectedHistogramResults) + } + + test("WorksMixedRangeWithTwoUnevenBuckets") { + // Make sure that it works with two unequally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(-0.01, 0.0, 1, 2, 3, 5, 6, 11.0, 11.01)) + val buckets = Array(0.0, 5.0, 11.0) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(4, 3) + assert(histogramResults === expectedHistogramResults) + } + + test("WorksMixedRangeWithFourUnevenBuckets") { + // Make sure that it works with two unequally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(-0.01, 0.0, 1, 2, 3, 5, 6, 11.01, 12.0, 199.0, + 200.0, 200.1)) + val buckets = Array(0.0, 5.0, 11.0, 12.0, 200.0) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(4, 2, 1, 3) + assert(histogramResults === expectedHistogramResults) + } + + test("WorksMixedRangeWithUnevenBucketsAndNaN") { + // Make sure that it works with two unequally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(-0.01, 0.0, 1, 2, 3, 5, 6, 11.01, 12.0, 199.0, + 200.0, 200.1, Double.NaN)) + val buckets = Array(0.0, 5.0, 11.0, 12.0, 200.0) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(4, 2, 1, 3) + assert(histogramResults === expectedHistogramResults) + } + // Make sure this works with a NaN end bucket + test("WorksMixedRangeWithUnevenBucketsAndNaNAndNaNRange") { + // Make sure that it works with two unequally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(-0.01, 0.0, 1, 2, 3, 5, 6, 11.01, 12.0, 199.0, + 200.0, 200.1, Double.NaN)) + val buckets = Array(0.0, 5.0, 11.0, 12.0, 200.0, Double.NaN) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(4, 2, 1, 2, 3) + assert(histogramResults === expectedHistogramResults) + } + // Make sure this works with a NaN end bucket and an inifity + test("WorksMixedRangeWithUnevenBucketsAndNaNAndNaNRangeAndInfity") { + // Make sure that it works with two unequally spaced buckets and elements in each + val rdd = sc.parallelize(Seq(-0.01, 0.0, 1, 2, 3, 5, 6, 11.01, 12.0, 199.0, + 200.0, 200.1, 1.0/0.0, -1.0/0.0, Double.NaN)) + val buckets = Array(0.0, 5.0, 11.0, 12.0, 200.0, Double.NaN) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(4, 2, 1, 2, 4) + assert(histogramResults === expectedHistogramResults) + } + + test("WorksWithOutOfRangeWithInfiniteBuckets") { + // Verify that out of range works with two buckets + val rdd = sc.parallelize(Seq(10.01, -0.01, Double.NaN)) + val buckets = Array(-1.0/0.0 , 0.0, 1.0/0.0) + val histogramResults = rdd.histogram(buckets) + val expectedHistogramResults = Array(1, 1) + assert(histogramResults === expectedHistogramResults) + } + // Test the failure mode with an invalid bucket array + test("ThrowsExceptionOnInvalidBucketArray") { + val rdd = sc.parallelize(Seq(1.0)) + // Empty array + intercept[IllegalArgumentException] { + val buckets = Array.empty[Double] + val result = rdd.histogram(buckets) + } + // Single element array + intercept[IllegalArgumentException] { + val buckets = Array(1.0) + val result = rdd.histogram(buckets) + } + } + + // Test automatic histogram function + test("WorksWithoutBucketsBasic") { + // Verify the basic case of one bucket and all elements in that bucket works + val rdd = sc.parallelize(Seq(1, 2, 3, 4)) + val (histogramBuckets, histogramResults) = rdd.histogram(1) + val expectedHistogramResults = Array(4) + val expectedHistogramBuckets = Array(1.0, 4.0) + assert(histogramResults === expectedHistogramResults) + assert(histogramBuckets === expectedHistogramBuckets) + } + // Test automatic histogram function with a single element + test("WorksWithoutBucketsBasicSingleElement") { + // Verify the basic case of one bucket and all elements in that bucket works + val rdd = sc.parallelize(Seq(1)) + val (histogramBuckets, histogramResults) = rdd.histogram(1) + val expectedHistogramResults = Array(1) + val expectedHistogramBuckets = Array(1.0, 1.0) + assert(histogramResults === expectedHistogramResults) + assert(histogramBuckets === expectedHistogramBuckets) + } + // Test automatic histogram function with a single element + test("WorksWithoutBucketsBasicNoRange") { + // Verify the basic case of one bucket and all elements in that bucket works + val rdd = sc.parallelize(Seq(1, 1, 1, 1)) + val (histogramBuckets, histogramResults) = rdd.histogram(1) + val expectedHistogramResults = Array(4) + val expectedHistogramBuckets = Array(1.0, 1.0) + assert(histogramResults === expectedHistogramResults) + assert(histogramBuckets === expectedHistogramBuckets) + } + + test("WorksWithoutBucketsBasicTwo") { + // Verify the basic case of one bucket and all elements in that bucket works + val rdd = sc.parallelize(Seq(1, 2, 3, 4)) + val (histogramBuckets, histogramResults) = rdd.histogram(2) + val expectedHistogramResults = Array(2, 2) + val expectedHistogramBuckets = Array(1.0, 2.5, 4.0) + assert(histogramResults === expectedHistogramResults) + assert(histogramBuckets === expectedHistogramBuckets) + } + + test("WorksWithoutBucketsWithMoreRequestedThanElements") { + // Verify the basic case of one bucket and all elements in that bucket works + val rdd = sc.parallelize(Seq(1, 2)) + val (histogramBuckets, histogramResults) = rdd.histogram(10) + val expectedHistogramResults = + Array(1, 0, 0, 0, 0, 0, 0, 0, 0, 1) + val expectedHistogramBuckets = + Array(1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0) + assert(histogramResults === expectedHistogramResults) + assert(histogramBuckets === expectedHistogramBuckets) + } + + // Test the failure mode with an invalid RDD + test("ThrowsExceptionOnInvalidRDDs") { + // infinity + intercept[UnsupportedOperationException] { + val rdd = sc.parallelize(Seq(1, 1.0/0.0)) + val result = rdd.histogram(1) + } + // NaN + intercept[UnsupportedOperationException] { + val rdd = sc.parallelize(Seq(1, Double.NaN)) + val result = rdd.histogram(1) + } + // Empty + intercept[UnsupportedOperationException] { + val rdd: RDD[Double] = sc.parallelize(Seq()) + val result = rdd.histogram(1) + } + } + +} diff --git a/core/src/test/scala/org/apache/spark/rdd/PartitionPruningRDDSuite.scala b/core/src/test/scala/org/apache/spark/rdd/PartitionPruningRDDSuite.scala new file mode 100644 index 0000000000..53a7b7c44d --- /dev/null +++ b/core/src/test/scala/org/apache/spark/rdd/PartitionPruningRDDSuite.scala @@ -0,0 +1,86 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.rdd + +import org.scalatest.FunSuite +import org.apache.spark.{TaskContext, Partition, SharedSparkContext} + + +class PartitionPruningRDDSuite extends FunSuite with SharedSparkContext { + + + test("Pruned Partitions inherit locality prefs correctly") { + + val rdd = new RDD[Int](sc, Nil) { + override protected def getPartitions = { + Array[Partition]( + new TestPartition(0, 1), + new TestPartition(1, 1), + new TestPartition(2, 1)) + } + + def compute(split: Partition, context: TaskContext) = { + Iterator() + } + } + val prunedRDD = PartitionPruningRDD.create(rdd, { + x => if (x == 2) true else false + }) + assert(prunedRDD.partitions.length == 1) + val p = prunedRDD.partitions(0) + assert(p.index == 0) + assert(p.asInstanceOf[PartitionPruningRDDPartition].parentSplit.index == 2) + } + + + test("Pruned Partitions can be unioned ") { + + val rdd = new RDD[Int](sc, Nil) { + override protected def getPartitions = { + Array[Partition]( + new TestPartition(0, 4), + new TestPartition(1, 5), + new TestPartition(2, 6)) + } + + def compute(split: Partition, context: TaskContext) = { + List(split.asInstanceOf[TestPartition].testValue).iterator + } + } + val prunedRDD1 = PartitionPruningRDD.create(rdd, { + x => if (x == 0) true else false + }) + + val prunedRDD2 = PartitionPruningRDD.create(rdd, { + x => if (x == 2) true else false + }) + + val merged = prunedRDD1 ++ prunedRDD2 + assert(merged.count() == 2) + val take = merged.take(2) + assert(take.apply(0) == 4) + assert(take.apply(1) == 6) + } +} + +class TestPartition(i: Int, value: Int) extends Partition with Serializable { + def index = i + + def testValue = this.value + +} diff --git a/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala b/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala index a4d41ebbff..706d84a58b 100644 --- a/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala +++ b/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala @@ -206,6 +206,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont submit(rdd, Array(0)) complete(taskSets(0), List((Success, 42))) assert(results === Map(0 -> 42)) + assertDataStructuresEmpty } test("local job") { @@ -219,6 +220,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont val jobId = scheduler.nextJobId.getAndIncrement() runEvent(JobSubmitted(jobId, rdd, jobComputeFunc, Array(0), true, null, listener)) assert(results === Map(0 -> 42)) + assertDataStructuresEmpty } test("run trivial job w/ dependency") { @@ -227,6 +229,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont submit(finalRdd, Array(0)) complete(taskSets(0), Seq((Success, 42))) assert(results === Map(0 -> 42)) + assertDataStructuresEmpty } test("cache location preferences w/ dependency") { @@ -239,12 +242,14 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont assertLocations(taskSet, Seq(Seq("hostA", "hostB"))) complete(taskSet, Seq((Success, 42))) assert(results === Map(0 -> 42)) + assertDataStructuresEmpty } test("trivial job failure") { submit(makeRdd(1, Nil), Array(0)) failed(taskSets(0), "some failure") assert(failure.getMessage === "Job aborted: some failure") + assertDataStructuresEmpty } test("run trivial shuffle") { @@ -260,6 +265,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont Array(makeBlockManagerId("hostA"), makeBlockManagerId("hostB"))) complete(taskSets(1), Seq((Success, 42))) assert(results === Map(0 -> 42)) + assertDataStructuresEmpty } test("run trivial shuffle with fetch failure") { @@ -285,6 +291,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1.host) === Array("hostA", "hostB")) complete(taskSets(3), Seq((Success, 43))) assert(results === Map(0 -> 42, 1 -> 43)) + assertDataStructuresEmpty } test("ignore late map task completions") { @@ -313,6 +320,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont Array(makeBlockManagerId("hostB"), makeBlockManagerId("hostA"))) complete(taskSets(1), Seq((Success, 42), (Success, 43))) assert(results === Map(0 -> 42, 1 -> 43)) + assertDataStructuresEmpty } test("run trivial shuffle with out-of-band failure and retry") { @@ -329,15 +337,16 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont complete(taskSets(0), Seq( (Success, makeMapStatus("hostA", 1)), (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") { + // 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)) + assertDataStructuresEmpty + } + + test("recursive shuffle failures") { val shuffleOneRdd = makeRdd(2, Nil) val shuffleDepOne = new ShuffleDependency(shuffleOneRdd, null) val shuffleTwoRdd = makeRdd(2, List(shuffleDepOne)) @@ -363,6 +372,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont complete(taskSets(4), Seq((Success, makeMapStatus("hostA", 1)))) complete(taskSets(5), Seq((Success, 42))) assert(results === Map(0 -> 42)) + assertDataStructuresEmpty } test("cached post-shuffle") { @@ -394,6 +404,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont complete(taskSets(3), Seq((Success, makeMapStatus("hostD", 1)))) complete(taskSets(4), Seq((Success, 42))) assert(results === Map(0 -> 42)) + assertDataStructuresEmpty } /** @@ -413,4 +424,18 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont private def makeBlockManagerId(host: String): BlockManagerId = BlockManagerId("exec-" + host, host, 12345, 0) + private def assertDataStructuresEmpty = { + assert(scheduler.pendingTasks.isEmpty) + assert(scheduler.activeJobs.isEmpty) + assert(scheduler.failed.isEmpty) + assert(scheduler.idToActiveJob.isEmpty) + assert(scheduler.jobIdToStageIds.isEmpty) + assert(scheduler.stageIdToJobIds.isEmpty) + assert(scheduler.stageIdToStage.isEmpty) + assert(scheduler.stageToInfos.isEmpty) + assert(scheduler.resultStageToJob.isEmpty) + assert(scheduler.running.isEmpty) + assert(scheduler.shuffleToMapStage.isEmpty) + assert(scheduler.waiting.isEmpty) + } } diff --git a/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala b/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala index 984881861c..002368ff55 100644 --- a/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala +++ b/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala @@ -31,6 +31,7 @@ import org.apache.spark.rdd.RDD class JobLoggerSuite extends FunSuite with LocalSparkContext with ShouldMatchers { + val WAIT_TIMEOUT_MILLIS = 10000 test("inner method") { sc = new SparkContext("local", "joblogger") @@ -92,6 +93,8 @@ class JobLoggerSuite extends FunSuite with LocalSparkContext with ShouldMatchers val rdd = sc.parallelize(1 to 1e2.toInt, 4).map{ i => (i % 12, 2 * i) } rdd.reduceByKey(_+_).collect() + assert(sc.dagScheduler.listenerBus.waitUntilEmpty(WAIT_TIMEOUT_MILLIS)) + val user = System.getProperty("user.name", SparkContext.SPARK_UNKNOWN_USER) joblogger.getLogDir should be ("/tmp/spark-%s".format(user)) @@ -120,7 +123,9 @@ class JobLoggerSuite extends FunSuite with LocalSparkContext with ShouldMatchers sc.addSparkListener(joblogger) val rdd = sc.parallelize(1 to 1e2.toInt, 4).map{ i => (i % 12, 2 * i) } rdd.reduceByKey(_+_).collect() - + + assert(sc.dagScheduler.listenerBus.waitUntilEmpty(WAIT_TIMEOUT_MILLIS)) + joblogger.onJobStartCount should be (1) joblogger.onJobEndCount should be (1) joblogger.onTaskEndCount should be (8) diff --git a/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala b/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala index 0b9056344c..ef4c4c0f14 100644 --- a/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala +++ b/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala @@ -5,9 +5,9 @@ import java.io.{FileWriter, File} import scala.collection.mutable import com.google.common.io.Files -import org.scalatest.{BeforeAndAfterEach, FunSuite} +import org.scalatest.{BeforeAndAfterAll, BeforeAndAfterEach, FunSuite} -class DiskBlockManagerSuite extends FunSuite with BeforeAndAfterEach { +class DiskBlockManagerSuite extends FunSuite with BeforeAndAfterEach with BeforeAndAfterAll { val rootDir0 = Files.createTempDir() rootDir0.deleteOnExit() @@ -16,6 +16,12 @@ class DiskBlockManagerSuite extends FunSuite with BeforeAndAfterEach { val rootDirs = rootDir0.getName + "," + rootDir1.getName println("Created root dirs: " + rootDirs) + // This suite focuses primarily on consolidation features, + // so we coerce consolidation if not already enabled. + val consolidateProp = "spark.shuffle.consolidateFiles" + val oldConsolidate = Option(System.getProperty(consolidateProp)) + System.setProperty(consolidateProp, "true") + val shuffleBlockManager = new ShuffleBlockManager(null) { var idToSegmentMap = mutable.Map[ShuffleBlockId, FileSegment]() override def getBlockLocation(id: ShuffleBlockId) = idToSegmentMap(id) @@ -23,6 +29,10 @@ class DiskBlockManagerSuite extends FunSuite with BeforeAndAfterEach { var diskBlockManager: DiskBlockManager = _ + override def afterAll() { + oldConsolidate.map(c => System.setProperty(consolidateProp, c)) + } + override def beforeEach() { diskBlockManager = new DiskBlockManager(shuffleBlockManager, rootDirs) shuffleBlockManager.idToSegmentMap.clear() diff --git a/core/src/test/scala/org/apache/spark/util/XORShiftRandomSuite.scala b/core/src/test/scala/org/apache/spark/util/XORShiftRandomSuite.scala new file mode 100644 index 0000000000..b78367b6ca --- /dev/null +++ b/core/src/test/scala/org/apache/spark/util/XORShiftRandomSuite.scala @@ -0,0 +1,76 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.util + +import java.util.Random +import org.scalatest.FlatSpec +import org.scalatest.FunSuite +import org.scalatest.matchers.ShouldMatchers +import org.apache.spark.util.Utils.times + +class XORShiftRandomSuite extends FunSuite with ShouldMatchers { + + def fixture = new { + val seed = 1L + val xorRand = new XORShiftRandom(seed) + val hundMil = 1e8.toInt + } + + /* + * This test is based on a chi-squared test for randomness. The values are hard-coded + * so as not to create Spark's dependency on apache.commons.math3 just to call one + * method for calculating the exact p-value for a given number of random numbers + * and bins. In case one would want to move to a full-fledged test based on + * apache.commons.math3, the relevant class is here: + * org.apache.commons.math3.stat.inference.ChiSquareTest + */ + test ("XORShift generates valid random numbers") { + + val f = fixture + + val numBins = 10 + // create 10 bins + val bins = Array.fill(numBins)(0) + + // populate bins based on modulus of the random number + times(f.hundMil) {bins(math.abs(f.xorRand.nextInt) % 10) += 1} + + /* since the seed is deterministic, until the algorithm is changed, we know the result will be + * exactly this: Array(10004908, 9993136, 9994600, 10000744, 10000091, 10002474, 10002272, + * 10000790, 10002286, 9998699), so the test will never fail at the prespecified (5%) + * significance level. However, should the RNG implementation change, the test should still + * pass at the same significance level. The chi-squared test done in R gave the following + * results: + * > chisq.test(c(10004908, 9993136, 9994600, 10000744, 10000091, 10002474, 10002272, + * 10000790, 10002286, 9998699)) + * Chi-squared test for given probabilities + * data: c(10004908, 9993136, 9994600, 10000744, 10000091, 10002474, 10002272, 10000790, + * 10002286, 9998699) + * X-squared = 11.975, df = 9, p-value = 0.2147 + * Note that the p-value was ~0.22. The test will fail if alpha < 0.05, which for 100 million + * random numbers + * and 10 bins will happen at X-squared of ~16.9196. So, the test will fail if X-squared + * is greater than or equal to that number. + */ + val binSize = f.hundMil/numBins + val xSquared = bins.map(x => math.pow((binSize - x), 2)/binSize).sum + xSquared should be < (16.9196) + + } + +}
\ No newline at end of file diff --git a/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala b/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala index ca3f684668..63e874fed3 100644 --- a/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala +++ b/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala @@ -2,8 +2,20 @@ package org.apache.spark.util.collection import scala.collection.mutable.HashSet import org.scalatest.FunSuite - -class OpenHashMapSuite extends FunSuite { +import org.scalatest.matchers.ShouldMatchers +import org.apache.spark.util.SizeEstimator + +class OpenHashMapSuite extends FunSuite with ShouldMatchers { + + test("size for specialized, primitive value (int)") { + val capacity = 1024 + val map = new OpenHashMap[String, Int](capacity) + val actualSize = SizeEstimator.estimate(map) + // 64 bit for pointers, 32 bit for ints, and 1 bit for the bitset. + val expectedSize = capacity * (64 + 32 + 1) / 8 + // Make sure we are not allocating a significant amount of memory beyond our expected. + actualSize should be <= (expectedSize * 1.1).toLong + } test("initialization") { val goodMap1 = new OpenHashMap[String, Int](1) diff --git a/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala b/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala index 4e11e8a628..4768a1e60b 100644 --- a/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala +++ b/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala @@ -1,9 +1,27 @@ package org.apache.spark.util.collection import org.scalatest.FunSuite +import org.scalatest.matchers.ShouldMatchers +import org.apache.spark.util.SizeEstimator -class OpenHashSetSuite extends FunSuite { + +class OpenHashSetSuite extends FunSuite with ShouldMatchers { + + test("size for specialized, primitive int") { + val loadFactor = 0.7 + val set = new OpenHashSet[Int](64, loadFactor) + for (i <- 0 until 1024) { + set.add(i) + } + assert(set.size === 1024) + assert(set.capacity > 1024) + val actualSize = SizeEstimator.estimate(set) + // 32 bits for the ints + 1 bit for the bitset + val expectedSize = set.capacity * (32 + 1) / 8 + // Make sure we are not allocating a significant amount of memory beyond our expected. + actualSize should be <= (expectedSize * 1.1).toLong + } test("primitive int") { val set = new OpenHashSet[Int] diff --git a/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashSetSuite.scala b/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashMapSuite.scala index dfd6aed2c4..2220b4f0d5 100644 --- a/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashSetSuite.scala +++ b/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashMapSuite.scala @@ -2,8 +2,20 @@ package org.apache.spark.util.collection import scala.collection.mutable.HashSet import org.scalatest.FunSuite +import org.scalatest.matchers.ShouldMatchers +import org.apache.spark.util.SizeEstimator -class PrimitiveKeyOpenHashSetSuite extends FunSuite { +class PrimitiveKeyOpenHashMapSuite extends FunSuite with ShouldMatchers { + + test("size for specialized, primitive key, value (int, int)") { + val capacity = 1024 + val map = new PrimitiveKeyOpenHashMap[Int, Int](capacity) + val actualSize = SizeEstimator.estimate(map) + // 32 bit for keys, 32 bit for values, and 1 bit for the bitset. + val expectedSize = capacity * (32 + 32 + 1) / 8 + // Make sure we are not allocating a significant amount of memory beyond our expected. + actualSize should be <= (expectedSize * 1.1).toLong + } test("initialization") { val goodMap1 = new PrimitiveKeyOpenHashMap[Int, Int](1) diff --git a/docs/_layouts/global.html b/docs/_layouts/global.html index 0c1d657cde..ad7969d012 100755 --- a/docs/_layouts/global.html +++ b/docs/_layouts/global.html @@ -74,12 +74,12 @@ <li class="dropdown"> <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 Core for Java/Scala</a></li> + <li><a href="api/core/index.html#org.apache.spark.package">Spark Core for Java/Scala</a></li> <li><a href="api/pyspark/index.html">Spark Core for Python</a></li> <li class="divider"></li> - <li><a href="api/streaming/index.html">Spark Streaming</a></li> - <li><a href="api/mllib/index.html">MLlib (Machine Learning)</a></li> - <li><a href="api/bagel/index.html">Bagel (Pregel on Spark)</a></li> + <li><a href="api/streaming/index.html#org.apache.spark.streaming.package">Spark Streaming</a></li> + <li><a href="api/mllib/index.html#org.apache.spark.mllib.package">MLlib (Machine Learning)</a></li> + <li><a href="api/bagel/index.html#org.apache.spark.bagel.package">Bagel (Pregel on Spark)</a></li> </ul> </li> diff --git a/docs/bagel-programming-guide.md b/docs/bagel-programming-guide.md index 140190a38c..de001e6c52 100644 --- a/docs/bagel-programming-guide.md +++ b/docs/bagel-programming-guide.md @@ -106,7 +106,7 @@ _Example_ ## Operations -Here are the actions and types in the Bagel API. See [Bagel.scala](https://github.com/apache/incubator-spark/blob/master/bagel/src/main/scala/spark/bagel/Bagel.scala) for details. +Here are the actions and types in the Bagel API. See [Bagel.scala](https://github.com/apache/incubator-spark/blob/master/bagel/src/main/scala/org/apache/spark/bagel/Bagel.scala) for details. ### Actions diff --git a/docs/building-with-maven.md b/docs/building-with-maven.md index 19c01e179f..c709001632 100644 --- a/docs/building-with-maven.md +++ b/docs/building-with-maven.md @@ -45,6 +45,12 @@ For Apache Hadoop 2.x, 0.23.x, Cloudera CDH MRv2, and other Hadoop versions with # Cloudera CDH 4.2.0 with MapReduce v2 $ mvn -Phadoop2-yarn -Dhadoop.version=2.0.0-cdh4.2.0 -Dyarn.version=2.0.0-chd4.2.0 -DskipTests clean package +Hadoop versions 2.2.x and newer can be built by setting the ```new-yarn``` and the ```yarn.version``` as follows: + + # Apache Hadoop 2.2.X and newer + $ mvn -Dyarn.version=2.2.0 -Dhadoop.version=2.2.0 -Pnew-yarn + +The build process handles Hadoop 2.2.x as a special case that uses the directory ```new-yarn```, which supports the new YARN API. Furthermore, for this version, the build depends on artifacts published by the spark-project to enable Akka 2.0.5 to work with protobuf 2.5. ## Spark Tests in Maven ## diff --git a/docs/cluster-overview.md b/docs/cluster-overview.md index 5927f736f3..e16703292c 100644 --- a/docs/cluster-overview.md +++ b/docs/cluster-overview.md @@ -45,7 +45,7 @@ The system currently supports three cluster managers: easy to set up a cluster. * [Apache Mesos](running-on-mesos.html) -- a general cluster manager that can also run Hadoop MapReduce and service applications. -* [Hadoop YARN](running-on-yarn.html) -- the resource manager in Hadoop 2.0. +* [Hadoop YARN](running-on-yarn.html) -- the resource manager in Hadoop 2. In addition, Spark's [EC2 launch scripts](ec2-scripts.html) make it easy to launch a standalone cluster on Amazon EC2. diff --git a/docs/configuration.md b/docs/configuration.md index 97183bafdb..62c4457a3e 100644 --- a/docs/configuration.md +++ b/docs/configuration.md @@ -327,7 +327,41 @@ Apart from these, the following properties are also available, and may be useful Too large a value decreases parallelism during broadcast (makes it slower); however, if it is too small, <code>BlockManager</code> might take a performance hit. </td> </tr> - +<tr> + <td>spark.shuffle.consolidateFiles</td> + <td>false</td> + <td> + If set to "true", consolidates intermediate files created during a shuffle. Creating fewer files can improve filesystem performance for shuffles with large numbers of reduce tasks. It is recommended to set this to "true" when using ext4 or xfs filesystems. On ext3, this option might degrade performance on machines with many (>8) cores due to filesystem limitations. + </td> +</tr> +<tr> + <td>spark.speculation</td> + <td>false</td> + <td> + If set to "true", performs speculative execution of tasks. This means if one or more tasks are running slowly in a stage, they will be re-launched. + </td> +</tr> +<tr> + <td>spark.speculation.interval</td> + <td>100</td> + <td> + How often Spark will check for tasks to speculate, in milliseconds. + </td> +</tr> +<tr> + <td>spark.speculation.quantile</td> + <td>0.75</td> + <td> + Percentage of tasks which must be complete before speculation is enabled for a particular stage. + </td> +</tr> +<tr> + <td>spark.speculation.multiplier</td> + <td>1.5</td> + <td> + How many times slower a task is than the median to be considered for speculation. + </td> +</tr> </table> # Environment Variables diff --git a/docs/hadoop-third-party-distributions.md b/docs/hadoop-third-party-distributions.md index b33af2cf24..de6a2b0a43 100644 --- a/docs/hadoop-third-party-distributions.md +++ b/docs/hadoop-third-party-distributions.md @@ -10,7 +10,7 @@ with these distributions: # Compile-time Hadoop Version When compiling Spark, you'll need to -[set the SPARK_HADOOP_VERSION flag](http://localhost:4000/index.html#a-note-about-hadoop-versions): +[set the SPARK_HADOOP_VERSION flag](index.html#a-note-about-hadoop-versions): SPARK_HADOOP_VERSION=1.0.4 sbt/sbt assembly @@ -40,6 +40,7 @@ the _exact_ Hadoop version you are running to avoid any compatibility errors. <tr><td>HDP 1.2</td><td>1.1.2</td></tr> <tr><td>HDP 1.1</td><td>1.0.3</td></tr> <tr><td>HDP 1.0</td><td>1.0.3</td></tr> + <tr><td>HDP 2.0</td><td>2.2.0</td></tr> </table> </td> </tr> diff --git a/docs/index.md b/docs/index.md index bd386a8a8f..d3ac696d1e 100644 --- a/docs/index.md +++ b/docs/index.md @@ -56,14 +56,16 @@ Hadoop, you must build Spark against the same version that your cluster uses. By default, Spark links to Hadoop 1.0.4. You can change this by setting the `SPARK_HADOOP_VERSION` variable when compiling: - SPARK_HADOOP_VERSION=1.2.1 sbt/sbt assembly + SPARK_HADOOP_VERSION=2.2.0 sbt/sbt assembly -In addition, if you wish to run Spark on [YARN](running-on-yarn.md), set +In addition, if you wish to run Spark on [YARN](running-on-yarn.html), set `SPARK_YARN` to `true`: SPARK_HADOOP_VERSION=2.0.5-alpha SPARK_YARN=true sbt/sbt assembly -(Note that on Windows, you need to set the environment variables on separate lines, e.g., `set SPARK_HADOOP_VERSION=1.2.1`.) +Note that on Windows, you need to set the environment variables on separate lines, e.g., `set SPARK_HADOOP_VERSION=1.2.1`. + +For this version of Spark (0.8.1) Hadoop 2.2.x (or newer) users will have to build Spark and publish it locally. See [Launching Spark on YARN](running-on-yarn.html). This is needed because Hadoop 2.2 has non backwards compatible API changes. # Where to Go from Here diff --git a/docs/job-scheduling.md b/docs/job-scheduling.md index d304c5497b..dbcb9ae343 100644 --- a/docs/job-scheduling.md +++ b/docs/job-scheduling.md @@ -91,7 +91,7 @@ The fair scheduler also supports grouping jobs into _pools_, and setting differe (e.g. weight) for each pool. This can be useful to create a "high-priority" pool for more important jobs, for example, or to group the jobs of each user together and give _users_ equal shares regardless of how many concurrent jobs they have instead of giving _jobs_ equal shares. This approach is modeled after the -[Hadoop Fair Scheduler](http://hadoop.apache.org/docs/stable/fair_scheduler.html). +[Hadoop Fair Scheduler](http://hadoop.apache.org/docs/current/hadoop-yarn/hadoop-yarn-site/FairScheduler.html). Without any intervention, newly submitted jobs go into a _default pool_, but jobs' pools can be set by adding the `spark.scheduler.pool` "local property" to the SparkContext in the thread that's submitting them. diff --git a/docs/monitoring.md b/docs/monitoring.md index 5f456b999b..5ed0474477 100644 --- a/docs/monitoring.md +++ b/docs/monitoring.md @@ -50,6 +50,7 @@ Each instance can report to zero or more _sinks_. Sinks are contained in the * `GangliaSink`: Sends metrics to a Ganglia node or multicast group. * `JmxSink`: Registers metrics for viewing in a JXM console. * `MetricsServlet`: Adds a servlet within the existing Spark UI to serve metrics data as JSON data. +* `GraphiteSink`: Sends metrics to a Graphite node. The syntax of the metrics configuration file is defined in an example configuration file, `$SPARK_HOME/conf/metrics.conf.template`. diff --git a/docs/running-on-yarn.md b/docs/running-on-yarn.md index 4056e9c15d..9e4c4e123f 100644 --- a/docs/running-on-yarn.md +++ b/docs/running-on-yarn.md @@ -17,6 +17,7 @@ This can be built by setting the Hadoop version and `SPARK_YARN` environment var The assembled JAR will be something like this: `./assembly/target/scala-{{site.SCALA_VERSION}}/spark-assembly_{{site.SPARK_VERSION}}-hadoop2.0.5.jar`. +The build process now also supports new YARN versions (2.2.x). See below. # Preparations @@ -45,6 +46,10 @@ System Properties: 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. +There are two scheduler mode that can be used to launch spark application on YARN. + +## Launch spark application by YARN Client with yarn-standalone mode. + The command to launch the YARN Client is as follows: SPARK_JAR=<SPARK_ASSEMBLY_JAR_FILE> ./spark-class org.apache.spark.deploy.yarn.Client \ @@ -52,6 +57,7 @@ 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-class <ApplicationMaster_CLASS> --master-memory <MEMORY_FOR_MASTER> \ --worker-memory <MEMORY_PER_WORKER> \ --worker-cores <CORES_PER_WORKER> \ @@ -85,12 +91,37 @@ For example: $ cat $YARN_APP_LOGS_DIR/$YARN_APP_ID/container*_000001/stdout Pi is roughly 3.13794 -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. +The above starts a YARN Client programs which start the default Application Master. Then SparkPi will be run as a child thread of Application Master, YARN Client will periodically polls the Application Master for status updates and displays them in the console. The client will exit once your application has finished running. + +With this mode, your application is actually run on the remote machine where the Application Master is run upon. Thus application that involve local interaction will not work well, e.g. spark-shell. + +## Launch spark application with yarn-client mode. + +With yarn-client mode, the application will be launched locally. Just like running application or spark-shell on Local / Mesos / Standalone mode. The launch method is also the similar with them, just make sure that when you need to specify a master url, use "yarn-client" instead. And you also need to export the env value for SPARK_JAR and SPARK_YARN_APP_JAR + +In order to tune worker core/number/memory etc. You need to export SPARK_WORKER_CORES, SPARK_WORKER_MEMORY, SPARK_WORKER_INSTANCES e.g. by ./conf/spark-env.sh + +For example: + + SPARK_JAR=./assembly/target/scala-{{site.SCALA_VERSION}}/spark-assembly-{{site.SPARK_VERSION}}-hadoop2.0.5-alpha.jar \ + SPARK_YARN_APP_JAR=examples/target/scala-{{site.SCALA_VERSION}}/spark-examples-assembly-{{site.SPARK_VERSION}}.jar \ + ./run-example org.apache.spark.examples.SparkPi yarn-client + + + SPARK_JAR=./assembly/target/scala-{{site.SCALA_VERSION}}/spark-assembly-{{site.SPARK_VERSION}}-hadoop2.0.5-alpha.jar \ + SPARK_YARN_APP_JAR=examples/target/scala-{{site.SCALA_VERSION}}/spark-examples-assembly-{{site.SPARK_VERSION}}.jar \ + MASTER=yarn-client ./spark-shell + +# Building Spark for Hadoop/YARN 2.2.x + +Hadoop 2.2.x users must build Spark and publish it locally. The SBT build process handles Hadoop 2.2.x as a special case. This version of Hadoop has new YARN API changes and depends on a Protobuf version (2.5) that is not compatible with the Akka version (2.0.5) that Spark uses. Therefore, if the Hadoop version (e.g. set through ```SPARK_HADOOP_VERSION```) starts with 2.2.0 or higher then the build process will depend on Akka artifacts distributed by the Spark project compatible with Protobuf 2.5. Furthermore, the build process then uses the directory ```new-yarn``` (instead of ```yarn```), which supports the new YARN API. The build process should seamlessly work out of the box. + +See [Building Spark with Maven](building-with-maven.html) for instructions on how to build Spark using the Maven process. # Important Notes -- 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. - The local directories used for spark will be the local directories configured for YARN (Hadoop Yarn config yarn.nodemanager.local-dirs). If the user specifies spark.local.dir, it will be ignored. - The --files and --archives options support specifying file names with the # similar to Hadoop. For example you can specify: --files localtest.txt#appSees.txt and this will upload the file you have locally named localtest.txt into HDFS but this will be linked to by the name appSees.txt and your application should use the name as appSees.txt to reference it when running on YARN. - The --addJars option allows the SparkContext.addJar function to work if you are using it with local files. It does not need to be used if you are using it with HDFS, HTTP, HTTPS, or FTP files. +- YARN 2.2.x users cannot simply depend on the Spark packages without building Spark, as the published Spark artifacts are compiled to work with the pre 2.2 API. Those users must build Spark and publish it locally. diff --git a/docs/spark-standalone.md b/docs/spark-standalone.md index 17066ef0dd..b822265b5a 100644 --- a/docs/spark-standalone.md +++ b/docs/spark-standalone.md @@ -51,11 +51,11 @@ Finally, the following configuration options can be passed to the master and wor </tr> <tr> <td><code>-c CORES</code>, <code>--cores CORES</code></td> - <td>Total CPU cores to allow Spark applicatons to use on the machine (default: all available); only on worker</td> + <td>Total CPU cores to allow Spark applications to use on the machine (default: all available); only on worker</td> </tr> <tr> <td><code>-m MEM</code>, <code>--memory MEM</code></td> - <td>Total amount of memory to allow Spark applicatons to use on the machine, in a format like 1000M or 2G (default: your machine's total RAM minus 1 GB); only on worker</td> + <td>Total amount of memory to allow Spark applications to use on the machine, in a format like 1000M or 2G (default: your machine's total RAM minus 1 GB); only on worker</td> </tr> <tr> <td><code>-d DIR</code>, <code>--work-dir DIR</code></td> diff --git a/docs/streaming-programming-guide.md b/docs/streaming-programming-guide.md index 851e30fe76..82f42e0b8d 100644 --- a/docs/streaming-programming-guide.md +++ b/docs/streaming-programming-guide.md @@ -214,7 +214,7 @@ ssc.stop() {% endhighlight %} # Example -A simple example to start off is the [NetworkWordCount](https://github.com/apache/incubator-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` . +A simple example to start off is the [NetworkWordCount](https://github.com/apache/incubator-spark/tree/master/examples/src/main/scala/org/apache/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/org/apache/spark/streaming/examples/NetworkWordCount.scala` . {% highlight scala %} import org.apache.spark.streaming.{Seconds, StreamingContext} @@ -283,7 +283,7 @@ 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-example org.apache.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. +You can find more examples in `<Spark repo>/streaming/src/main/scala/org/apache/spark/streaming/examples/`. They can be run in the similar manner using `./run-example org.apache.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()`. @@ -483,7 +483,7 @@ Similar to [Spark's Java API](java-programming-guide.html), we also provide a Ja 1. Functions for transformations must be implemented as subclasses of [Function](api/core/index.html#org.apache.spark.api.java.function.Function) and [Function2](api/core/index.html#org.apache.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#org.apache.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` +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/org/apache/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. @@ -527,5 +527,5 @@ JavaPairDStream<String, Integer> wordCounts = words.map( # Where to Go from Here * API docs - [Scala](api/streaming/index.html#org.apache.spark.streaming.package) and [Java](api/streaming/index.html#org.apache.spark.streaming.api.java.package) -* More examples - [Scala](https://github.com/apache/incubator-spark/tree/master/examples/src/main/scala/spark/streaming/examples) and [Java](https://github.com/apache/incubator-spark/tree/master/examples/src/main/java/spark/streaming/examples) +* More examples - [Scala](https://github.com/apache/incubator-spark/tree/master/examples/src/main/scala/org/apache/spark/streaming/examples) and [Java](https://github.com/apache/incubator-spark/tree/master/examples/src/main/java/org/apache/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 f491ae9b95..a4be188169 100644 --- a/docs/tuning.md +++ b/docs/tuning.md @@ -39,7 +39,8 @@ in your operations) and performance. It provides two serialization libraries: for best performance. You can switch to using Kryo by calling `System.setProperty("spark.serializer", "org.apache.spark.serializer.KryoSerializer")` -*before* creating your SparkContext. The only reason it is not the default is because of the custom +*before* creating your SparkContext. This setting configures the serializer used for not only shuffling data between worker +nodes but also when serializing RDDs to disk. The only reason Kryo is not the default is because of the custom registration requirement, but we recommend trying it in any network-intensive application. Finally, to register your classes with Kryo, create a public class that extends @@ -67,7 +68,7 @@ The [Kryo documentation](http://code.google.com/p/kryo/) describes more advanced registration options, such as adding custom serialization code. If your objects are large, you may also need to increase the `spark.kryoserializer.buffer.mb` -system property. The default is 32, but this value needs to be large enough to hold the *largest* +system property. The default is 2, but this value needs to be large enough to hold the *largest* object you will serialize. Finally, if you don't register your classes, Kryo will still work, but it will have to store the diff --git a/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala b/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala index 4af45b2b4a..83db8b9e26 100644 --- a/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala +++ b/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala @@ -120,7 +120,7 @@ object LocalALS { System.exit(1) } } - printf("Running with M=%d, U=%d, F=%d, iters=%d\n", M, U, F, ITERATIONS); + printf("Running with M=%d, U=%d, F=%d, iters=%d\n", M, U, F, ITERATIONS) val R = generateR() diff --git a/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala b/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala index 5a7a9d1bd8..8543ce0e32 100644 --- a/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala +++ b/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala @@ -65,7 +65,7 @@ object SparkTC { oldCount = nextCount // Perform the join, obtaining an RDD of (y, (z, x)) pairs, // then project the result to obtain the new (x, z) paths. - tc = tc.union(tc.join(edges).map(x => (x._2._2, x._2._1))).distinct().cache(); + tc = tc.union(tc.join(edges).map(x => (x._2._2, x._2._1))).distinct().cache() nextCount = tc.count() } while (nextCount != oldCount) diff --git a/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala b/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala index cd3423a07b..af52b7e9a1 100644 --- a/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala +++ b/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala @@ -120,7 +120,7 @@ object FeederActor { println("Feeder started as:" + feeder) - actorSystem.awaitTermination(); + actorSystem.awaitTermination() } } diff --git a/examples/src/main/scala/org/apache/spark/streaming/examples/MQTTWordCount.scala b/examples/src/main/scala/org/apache/spark/streaming/examples/MQTTWordCount.scala index af698a01d5..ff332a0282 100644 --- a/examples/src/main/scala/org/apache/spark/streaming/examples/MQTTWordCount.scala +++ b/examples/src/main/scala/org/apache/spark/streaming/examples/MQTTWordCount.scala @@ -54,12 +54,12 @@ object MQTTPublisher { client.connect() - val msgtopic: MqttTopic = client.getTopic(topic); + val msgtopic: MqttTopic = client.getTopic(topic) val msg: String = "hello mqtt demo for spark streaming" while (true) { val message: MqttMessage = new MqttMessage(String.valueOf(msg).getBytes()) - msgtopic.publish(message); + msgtopic.publish(message) println("Published data. topic: " + msgtopic.getName() + " Message: " + message) } client.disconnect() diff --git a/mllib/src/main/scala/org/apache/spark/mllib/clustering/KMeans.scala b/mllib/src/main/scala/org/apache/spark/mllib/clustering/KMeans.scala index edbf77dbcc..0dee9399a8 100644 --- a/mllib/src/main/scala/org/apache/spark/mllib/clustering/KMeans.scala +++ b/mllib/src/main/scala/org/apache/spark/mllib/clustering/KMeans.scala @@ -18,15 +18,16 @@ package org.apache.spark.mllib.clustering import scala.collection.mutable.ArrayBuffer -import scala.util.Random + +import org.jblas.DoubleMatrix import org.apache.spark.SparkContext import org.apache.spark.SparkContext._ import org.apache.spark.rdd.RDD import org.apache.spark.Logging import org.apache.spark.mllib.util.MLUtils +import org.apache.spark.util.XORShiftRandom -import org.jblas.DoubleMatrix /** @@ -195,7 +196,7 @@ class KMeans private ( */ private def initRandom(data: RDD[Array[Double]]): Array[ClusterCenters] = { // Sample all the cluster centers in one pass to avoid repeated scans - val sample = data.takeSample(true, runs * k, new Random().nextInt()).toSeq + val sample = data.takeSample(true, runs * k, new XORShiftRandom().nextInt()).toSeq Array.tabulate(runs)(r => sample.slice(r * k, (r + 1) * k).toArray) } @@ -210,7 +211,7 @@ class KMeans private ( */ private def initKMeansParallel(data: RDD[Array[Double]]): Array[ClusterCenters] = { // Initialize each run's center to a random point - val seed = new Random().nextInt() + val seed = new XORShiftRandom().nextInt() val sample = data.takeSample(true, runs, seed).toSeq val centers = Array.tabulate(runs)(r => ArrayBuffer(sample(r))) @@ -222,7 +223,7 @@ class KMeans private ( for (r <- 0 until runs) yield (r, KMeans.pointCost(centerArrays(r), point)) }.reduceByKey(_ + _).collectAsMap() val chosen = data.mapPartitionsWithIndex { (index, points) => - val rand = new Random(seed ^ (step << 16) ^ index) + val rand = new XORShiftRandom(seed ^ (step << 16) ^ index) for { p <- points r <- 0 until runs diff --git a/new-yarn/pom.xml b/new-yarn/pom.xml new file mode 100644 index 0000000000..8a065c6d7d --- /dev/null +++ b/new-yarn/pom.xml @@ -0,0 +1,161 @@ +<?xml version="1.0" encoding="UTF-8"?> +<!-- + ~ Licensed to the Apache Software Foundation (ASF) under one or more + ~ contributor license agreements. See the NOTICE file distributed with + ~ this work for additional information regarding copyright ownership. + ~ The ASF licenses this file to You under the Apache License, Version 2.0 + ~ (the "License"); you may not use this file except in compliance with + ~ the License. You may obtain a copy of the License at + ~ + ~ http://www.apache.org/licenses/LICENSE-2.0 + ~ + ~ Unless required by applicable law or agreed to in writing, software + ~ distributed under the License is distributed on an "AS IS" BASIS, + ~ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + ~ See the License for the specific language governing permissions and + ~ limitations under the License. + --> +<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> + <parent> + <groupId>org.apache.spark</groupId> + <artifactId>spark-parent</artifactId> + <version>0.9.0-incubating-SNAPSHOT</version> + <relativePath>../pom.xml</relativePath> + </parent> + + <groupId>org.apache.spark</groupId> + <artifactId>spark-yarn_2.9.3</artifactId> + <packaging>jar</packaging> + <name>Spark Project YARN Support</name> + <url>http://spark.incubator.apache.org/</url> + + <dependencies> + <dependency> + <groupId>org.apache.spark</groupId> + <artifactId>spark-core_2.9.3</artifactId> + <version>${project.version}</version> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-api</artifactId> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-common</artifactId> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-yarn-client</artifactId> + </dependency> + <dependency> + <groupId>org.apache.hadoop</groupId> + <artifactId>hadoop-client</artifactId> + <version>${yarn.version}</version> + </dependency> + <dependency> + <groupId>org.apache.avro</groupId> + <artifactId>avro</artifactId> + </dependency> + <dependency> + <groupId>org.apache.avro</groupId> + <artifactId>avro-ipc</artifactId> + </dependency> + <dependency> + <groupId>org.scalatest</groupId> + <artifactId>scalatest_2.9.3</artifactId> + <scope>test</scope> + </dependency> + <dependency> + <groupId>org.mockito</groupId> + <artifactId>mockito-all</artifactId> + <scope>test</scope> + </dependency> + </dependencies> + + <build> + <outputDirectory>target/scala-${scala.version}/classes</outputDirectory> + <testOutputDirectory>target/scala-${scala.version}/test-classes</testOutputDirectory> + <plugins> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-shade-plugin</artifactId> + <configuration> + <shadedArtifactAttached>false</shadedArtifactAttached> + <outputFile>${project.build.directory}/${project.artifactId}-${project.version}-shaded.jar</outputFile> + <artifactSet> + <includes> + <include>*:*</include> + </includes> + </artifactSet> + <filters> + <filter> + <artifact>*:*</artifact> + <excludes> + <exclude>META-INF/*.SF</exclude> + <exclude>META-INF/*.DSA</exclude> + <exclude>META-INF/*.RSA</exclude> + </excludes> + </filter> + </filters> + </configuration> + <executions> + <execution> + <phase>package</phase> + <goals> + <goal>shade</goal> + </goals> + <configuration> + <transformers> + <transformer implementation="org.apache.maven.plugins.shade.resource.ServicesResourceTransformer" /> + <transformer implementation="org.apache.maven.plugins.shade.resource.AppendingTransformer"> + <resource>reference.conf</resource> + </transformer> + </transformers> + </configuration> + </execution> + </executions> + </plugin> + <plugin> + <groupId>org.apache.maven.plugins</groupId> + <artifactId>maven-antrun-plugin</artifactId> + <executions> + <execution> + <phase>test</phase> + <goals> + <goal>run</goal> + </goals> + <configuration> + <exportAntProperties>true</exportAntProperties> + <tasks> + <property name="spark.classpath" refid="maven.test.classpath" /> + <property environment="env" /> + <fail message="Please set the SCALA_HOME (or SCALA_LIBRARY_PATH if scala is on the path) environment variables and retry."> + <condition> + <not> + <or> + <isset property="env.SCALA_HOME" /> + <isset property="env.SCALA_LIBRARY_PATH" /> + </or> + </not> + </condition> + </fail> + </tasks> + </configuration> + </execution> + </executions> + </plugin> + <plugin> + <groupId>org.scalatest</groupId> + <artifactId>scalatest-maven-plugin</artifactId> + <configuration> + <environmentVariables> + <SPARK_HOME>${basedir}/..</SPARK_HOME> + <SPARK_TESTING>1</SPARK_TESTING> + <SPARK_CLASSPATH>${spark.classpath}</SPARK_CLASSPATH> + </environmentVariables> + </configuration> + </plugin> + </plugins> + </build> +</project> diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala new file mode 100644 index 0000000000..eeeca3ea8a --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala @@ -0,0 +1,446 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.io.IOException +import java.net.Socket +import java.util.concurrent.CopyOnWriteArrayList +import java.util.concurrent.atomic.{AtomicInteger, AtomicReference} + +import scala.collection.JavaConversions._ + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.{FileSystem, Path} +import org.apache.hadoop.net.NetUtils +import org.apache.hadoop.security.UserGroupInformation +import org.apache.hadoop.util.ShutdownHookManager +import org.apache.hadoop.yarn.api._ +import org.apache.hadoop.yarn.api.protocolrecords._ +import org.apache.hadoop.yarn.api.records._ +import org.apache.hadoop.yarn.client.api.AMRMClient +import org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest +import org.apache.hadoop.yarn.conf.YarnConfiguration +import org.apache.hadoop.yarn.ipc.YarnRPC +import org.apache.hadoop.yarn.util.{ConverterUtils, Records} + +import org.apache.spark.{SparkContext, Logging} +import org.apache.spark.util.Utils + + +class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) extends Logging { + + def this(args: ApplicationMasterArguments) = this(args, new Configuration()) + + private var rpc: YarnRPC = YarnRPC.create(conf) + private val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + private var appAttemptId: ApplicationAttemptId = _ + private var userThread: Thread = _ + private val fs = FileSystem.get(yarnConf) + + private var yarnAllocator: YarnAllocationHandler = _ + private var isFinished: Boolean = false + private var uiAddress: String = _ + private val maxAppAttempts: Int = conf.getInt( + YarnConfiguration.RM_AM_MAX_ATTEMPTS, YarnConfiguration.DEFAULT_RM_AM_MAX_ATTEMPTS) + private var isLastAMRetry: Boolean = true + private var amClient: AMRMClient[ContainerRequest] = _ + + // Default to numWorkers * 2, with minimum of 3 + private val maxNumWorkerFailures = System.getProperty("spark.yarn.max.worker.failures", + math.max(args.numWorkers * 2, 3).toString()).toInt + + def run() { + // Setup the directories so things go to YARN approved directories rather + // than user specified and /tmp. + System.setProperty("spark.local.dir", getLocalDirs()) + + // Use priority 30 as it's higher then HDFS. It's same priority as MapReduce is using. + ShutdownHookManager.get().addShutdownHook(new AppMasterShutdownHook(this), 30) + + appAttemptId = getApplicationAttemptId() + isLastAMRetry = appAttemptId.getAttemptId() >= maxAppAttempts + amClient = AMRMClient.createAMRMClient() + amClient.init(yarnConf) + amClient.start() + + // Workaround until hadoop moves to something which has + // https://issues.apache.org/jira/browse/HADOOP-8406 - fixed in (2.0.2-alpha but no 0.23 line) + // 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() + + waitForSparkContextInitialized() + + // Do this after Spark master is up and SparkContext is created so that we can register UI Url. + val appMasterResponse: RegisterApplicationMasterResponse = registerApplicationMaster() + + // Allocate all containers + allocateWorkers() + + // Wait for the user class to Finish + userThread.join() + + System.exit(0) + } + + /** Get the Yarn approved local directories. */ + private def getLocalDirs(): String = { + // Hadoop 0.23 and 2.x have different Environment variable names for the + // local dirs, so lets check both. We assume one of the 2 is set. + // LOCAL_DIRS => 2.X, YARN_LOCAL_DIRS => 0.23.X + val localDirs = Option(System.getenv("YARN_LOCAL_DIRS")) + .getOrElse(Option(System.getenv("LOCAL_DIRS")) + .getOrElse("")) + + if (localDirs.isEmpty()) { + throw new Exception("Yarn Local dirs can't be empty") + } + localDirs + } + + private def getApplicationAttemptId(): ApplicationAttemptId = { + val envs = System.getenv() + val containerIdString = envs.get(ApplicationConstants.Environment.CONTAINER_ID.name()) + val containerId = ConverterUtils.toContainerId(containerIdString) + val appAttemptId = containerId.getApplicationAttemptId() + logInfo("ApplicationAttemptId: " + appAttemptId) + appAttemptId + } + + private def registerApplicationMaster(): RegisterApplicationMasterResponse = { + logInfo("Registering the ApplicationMaster") + amClient.registerApplicationMaster(Utils.localHostName(), 0, uiAddress) + } + + private def waitForSparkMaster() { + logInfo("Waiting for Spark driver to be reachable.") + var driverUp = false + var tries = 0 + val numTries = System.getProperty("spark.yarn.applicationMaster.waitTries", "10").toInt + while (!driverUp && tries < numTries) { + 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("Driver now available: %s:%s".format(driverHost, driverPort)) + driverUp = true + } catch { + case e: Exception => { + logWarning("Failed to connect to driver at %s:%s, retrying ...". + format(driverHost, driverPort)) + Thread.sleep(100) + tries = tries + 1 + } + } + } + } + + private def startUserClass(): Thread = { + logInfo("Starting the user JAR in a separate Thread") + val mainMethod = Class.forName( + args.userClass, + false /* initialize */, + Thread.currentThread.getContextClassLoader).getMethod("main", classOf[Array[String]]) + val t = new Thread { + override def run() { + var successed = false + try { + // Copy + var mainArgs: Array[String] = new Array[String](args.userArgs.size) + args.userArgs.copyToArray(mainArgs, 0, args.userArgs.size) + mainMethod.invoke(null, mainArgs) + // some job script has "System.exit(0)" at the end, for example SparkPi, SparkLR + // userThread will stop here unless it has uncaught exception thrown out + // It need shutdown hook to set SUCCEEDED + successed = true + } finally { + logDebug("finishing main") + isLastAMRetry = true + if (successed) { + ApplicationMaster.this.finishApplicationMaster(FinalApplicationStatus.SUCCEEDED) + } else { + ApplicationMaster.this.finishApplicationMaster(FinalApplicationStatus.FAILED) + } + } + } + } + t.start() + t + } + + // This need to happen before allocateWorkers() + private def waitForSparkContextInitialized() { + logInfo("Waiting for Spark context initialization") + try { + var sparkContext: SparkContext = null + ApplicationMaster.sparkContextRef.synchronized { + var numTries = 0 + val waitTime = 10000L + val maxNumTries = System.getProperty("spark.yarn.ApplicationMaster.waitTries", "10").toInt + while (ApplicationMaster.sparkContextRef.get() == null && numTries < maxNumTries) { + logInfo("Waiting for Spark context initialization ... " + numTries) + numTries = numTries + 1 + ApplicationMaster.sparkContextRef.wait(waitTime) + } + sparkContext = ApplicationMaster.sparkContextRef.get() + assert(sparkContext != null || numTries >= maxNumTries) + + if (sparkContext != null) { + uiAddress = sparkContext.ui.appUIAddress + this.yarnAllocator = YarnAllocationHandler.newAllocator( + yarnConf, + amClient, + appAttemptId, + args, + sparkContext.preferredNodeLocationData) + } else { + logWarning("Unable to retrieve SparkContext inspite of waiting for %d, maxNumTries = %d". + format(numTries * waitTime, maxNumTries)) + this.yarnAllocator = YarnAllocationHandler.newAllocator( + yarnConf, + amClient, + appAttemptId, + args) + } + } + } finally { + // In case of exceptions, etc - ensure that count is at least ALLOCATOR_LOOP_WAIT_COUNT : + // so that the loop (in ApplicationMaster.sparkContextInitialized) breaks. + ApplicationMaster.incrementAllocatorLoop(ApplicationMaster.ALLOCATOR_LOOP_WAIT_COUNT) + } + } + + private def allocateWorkers() { + try { + 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 + yarnAllocator.addResourceRequests(args.numWorkers) + // Exits the loop if the user thread exits. + while (yarnAllocator.getNumWorkersRunning < args.numWorkers && userThread.isAlive) { + if (yarnAllocator.getNumWorkersFailed >= maxNumWorkerFailures) { + finishApplicationMaster(FinalApplicationStatus.FAILED, + "max number of worker failures reached") + } + yarnAllocator.allocateResources() + ApplicationMaster.incrementAllocatorLoop(1) + Thread.sleep(100) + } + } finally { + // In case of exceptions, etc - ensure that count is at least 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 the app will get killed after expiration + // (def: 10mins) timeout. + if (userThread.isAlive) { + // Ensure that progress is sent before YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS elapses. + val timeoutInterval = yarnConf.getInt(YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS, 120000) + + // we want to be reasonably responsive without causing too many requests to RM. + val schedulerInterval = + System.getProperty("spark.yarn.scheduler.heartbeat.interval-ms", "5000").toLong + + // must be <= timeoutInterval / 2. + val interval = math.min(timeoutInterval / 2, schedulerInterval) + + launchReporterThread(interval) + } + } + + private def launchReporterThread(_sleepTime: Long): Thread = { + val sleepTime = if (_sleepTime <= 0 ) 0 else _sleepTime + + val t = new Thread { + override def run() { + while (userThread.isAlive) { + if (yarnAllocator.getNumWorkersFailed >= maxNumWorkerFailures) { + finishApplicationMaster(FinalApplicationStatus.FAILED, + "max number of worker failures reached") + } + val missingWorkerCount = args.numWorkers - yarnAllocator.getNumWorkersRunning - + yarnAllocator.getNumPendingAllocate + if (missingWorkerCount > 0) { + logInfo("Allocating %d containers to make up for (potentially) lost containers". + format(missingWorkerCount)) + yarnAllocator.addResourceRequests(missingWorkerCount) + } + 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) + t + } + + private def sendProgress() { + logDebug("Sending progress") + // Simulated with an allocate request with no nodes requested. + yarnAllocator.allocateResources() + } + + /* + 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(status: FinalApplicationStatus, diagnostics: String = "") { + synchronized { + if (isFinished) { + return + } + isFinished = true + } + + logInfo("finishApplicationMaster with " + status) + // Set tracking URL to empty since we don't have a history server. + amClient.unregisterApplicationMaster(status, "" /* appMessage */, "" /* appTrackingUrl */) + } + + /** + * Clean up the staging directory. + */ + private def cleanupStagingDir() { + var stagingDirPath: Path = null + try { + val preserveFiles = System.getProperty("spark.yarn.preserve.staging.files", "false").toBoolean + if (!preserveFiles) { + stagingDirPath = new Path(System.getenv("SPARK_YARN_STAGING_DIR")) + if (stagingDirPath == null) { + logError("Staging directory is null") + return + } + logInfo("Deleting staging directory " + stagingDirPath) + fs.delete(stagingDirPath, true) + } + } catch { + case ioe: IOException => + logError("Failed to cleanup staging dir " + stagingDirPath, ioe) + } + } + + // The shutdown hook that runs when a signal is received AND during normal close of the JVM. + class AppMasterShutdownHook(appMaster: ApplicationMaster) extends Runnable { + + def run() { + logInfo("AppMaster received a signal.") + // we need to clean up staging dir before HDFS is shut down + // make sure we don't delete it until this is the last AM + if (appMaster.isLastAMRetry) appMaster.cleanupStagingDir() + } + } +} + +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 + + private val applicationMasters = new CopyOnWriteArrayList[ApplicationMaster]() + + val sparkContextRef: AtomicReference[SparkContext] = + new AtomicReference[SparkContext](null /* initialValue */) + + val yarnAllocatorLoop: AtomicInteger = new AtomicInteger(0) + + def incrementAllocatorLoop(by: Int) { + val count = yarnAllocatorLoop.getAndAdd(by) + if (count >= ALLOCATOR_LOOP_WAIT_COUNT) { + yarnAllocatorLoop.synchronized { + // to wake threads off wait ... + yarnAllocatorLoop.notifyAll() + } + } + } + + def register(master: ApplicationMaster) { + applicationMasters.add(master) + } + + // TODO(harvey): See whether this should be discarded - it isn't used anywhere atm... + 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 the Client from declaring failure even though we exited properly. + // Note that this will unfortunately not properly clean up the staging files because it gets + // called too late, after the filesystem is already shutdown. + if (modified) { + Runtime.getRuntime().addShutdownHook(new Thread with Logging { + // This is not only logs, but also ensures that log system is initialized for this instance + // when we are actually 'run'-ing. + 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(FinalApplicationStatus.SUCCEEDED) + } + } + } ) + } + + // 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/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMasterArguments.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMasterArguments.scala new file mode 100644 index 0000000000..f76a5ddd39 --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMasterArguments.scala @@ -0,0 +1,94 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import org.apache.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: org.apache.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/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala new file mode 100644 index 0000000000..94678815e8 --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala @@ -0,0 +1,519 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.net.{InetAddress, UnknownHostException, URI} +import java.nio.ByteBuffer + +import scala.collection.JavaConversions._ +import scala.collection.mutable.HashMap +import scala.collection.mutable.Map + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.{FileContext, FileStatus, FileSystem, Path, FileUtil} +import org.apache.hadoop.fs.permission.FsPermission; +import org.apache.hadoop.io.DataOutputBuffer +import org.apache.hadoop.mapred.Master +import org.apache.hadoop.net.NetUtils +import org.apache.hadoop.security.UserGroupInformation +import org.apache.hadoop.yarn.api._ +import org.apache.hadoop.yarn.api.ApplicationConstants.Environment +import org.apache.hadoop.yarn.api.protocolrecords._ +import org.apache.hadoop.yarn.api.records._ +import org.apache.hadoop.yarn.client.api.impl.YarnClientImpl +import org.apache.hadoop.yarn.conf.YarnConfiguration +import org.apache.hadoop.yarn.ipc.YarnRPC +import org.apache.hadoop.yarn.util.{Apps, Records} + +import org.apache.spark.Logging +import org.apache.spark.util.Utils +import org.apache.spark.deploy.SparkHadoopUtil + + +/** + * The entry point (starting in Client#main() and Client#run()) for launching Spark on YARN. The + * Client submits an application to the global ResourceManager to launch Spark's ApplicationMaster, + * which will launch a Spark master process and negotiate resources throughout its duration. + */ +class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl with Logging { + + var rpc: YarnRPC = YarnRPC.create(conf) + val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + val credentials = UserGroupInformation.getCurrentUser().getCredentials() + private val SPARK_STAGING: String = ".sparkStaging" + private val distCacheMgr = new ClientDistributedCacheManager() + + // Staging directory is private! -> rwx-------- + val STAGING_DIR_PERMISSION: FsPermission = FsPermission.createImmutable(0700: Short) + // App files are world-wide readable and owner writable -> rw-r--r-- + val APP_FILE_PERMISSION: FsPermission = FsPermission.createImmutable(0644: Short) + + def this(args: ClientArguments) = this(new Configuration(), args) + + def runApp(): ApplicationId = { + validateArgs() + // Initialize and start the client service. + init(yarnConf) + start() + + // Log details about this YARN cluster (e.g, the number of slave machines/NodeManagers). + logClusterResourceDetails() + + // Prepare to submit a request to the ResourcManager (specifically its ApplicationsManager (ASM) + // interface). + + // Get a new client application. + val newApp = super.createApplication() + val newAppResponse = newApp.getNewApplicationResponse() + val appId = newAppResponse.getApplicationId() + + verifyClusterResources(newAppResponse) + + // Set up resource and environment variables. + val appStagingDir = getAppStagingDir(appId) + val localResources = prepareLocalResources(appStagingDir) + val launchEnv = setupLaunchEnv(localResources, appStagingDir) + val amContainer = createContainerLaunchContext(newAppResponse, localResources, launchEnv) + + // Set up an application submission context. + val appContext = newApp.getApplicationSubmissionContext() + appContext.setApplicationName(args.appName) + appContext.setQueue(args.amQueue) + appContext.setAMContainerSpec(amContainer) + + // Memory for the ApplicationMaster. + val memoryResource = Records.newRecord(classOf[Resource]).asInstanceOf[Resource] + memoryResource.setMemory(args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + appContext.setResource(memoryResource) + + // Finally, submit and monitor the application. + submitApp(appContext) + appId + } + + def run() { + val appId = runApp() + monitorApplication(appId) + System.exit(0) + } + + // TODO(harvey): This could just go in ClientArguments. + def validateArgs() = { + Map( + (System.getenv("SPARK_JAR") == null) -> "Error: You must set SPARK_JAR environment variable!", + (args.userJar == null) -> "Error: You must specify a user jar!", + (args.userClass == null) -> "Error: You must specify a user class!", + (args.numWorkers <= 0) -> "Error: You must specify atleast 1 worker!", + (args.amMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> ("Error: AM memory size must be" + + "greater than: " + YarnAllocationHandler.MEMORY_OVERHEAD), + (args.workerMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> ("Error: Worker memory size" + + "must be greater than: " + YarnAllocationHandler.MEMORY_OVERHEAD.toString) + ).foreach { case(cond, errStr) => + if (cond) { + logError(errStr) + args.printUsageAndExit(1) + } + } + } + + def getAppStagingDir(appId: ApplicationId): String = { + SPARK_STAGING + Path.SEPARATOR + appId.toString() + Path.SEPARATOR + } + + def logClusterResourceDetails() { + val clusterMetrics: YarnClusterMetrics = super.getYarnClusterMetrics + logInfo("Got Cluster metric info from ApplicationsManager (ASM), number of NodeManagers: " + + clusterMetrics.getNumNodeManagers) + + val queueInfo: QueueInfo = super.getQueueInfo(args.amQueue) + logInfo("""Queue info ... queueName: %s, queueCurrentCapacity: %s, queueMaxCapacity: %s, + queueApplicationCount = %s, queueChildQueueCount = %s""".format( + queueInfo.getQueueName, + queueInfo.getCurrentCapacity, + queueInfo.getMaximumCapacity, + queueInfo.getApplications.size, + queueInfo.getChildQueues.size)) + } + + def verifyClusterResources(app: GetNewApplicationResponse) = { + val maxMem = app.getMaximumResourceCapability().getMemory() + logInfo("Max mem capabililty of a single resource in this cluster " + maxMem) + + // If we have requested more then the clusters max for a single resource then exit. + if (args.workerMemory > maxMem) { + logError("Required worker memory (%d MB), is above the max threshold (%d MB) of this cluster.". + format(args.workerMemory, maxMem)) + System.exit(1) + } + val amMem = args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD + if (amMem > maxMem) { + logError("Required AM memory (%d) is above the max threshold (%d) of this cluster". + format(args.amMemory, maxMem)) + System.exit(1) + } + + // We could add checks to make sure the entire cluster has enough resources but that involves + // getting all the node reports and computing ourselves. + } + + /** See if two file systems are the same or not. */ + private def compareFs(srcFs: FileSystem, destFs: FileSystem): Boolean = { + val srcUri = srcFs.getUri() + val dstUri = destFs.getUri() + if (srcUri.getScheme() == null) { + return false + } + if (!srcUri.getScheme().equals(dstUri.getScheme())) { + return false + } + var srcHost = srcUri.getHost() + var dstHost = dstUri.getHost() + if ((srcHost != null) && (dstHost != null)) { + try { + srcHost = InetAddress.getByName(srcHost).getCanonicalHostName() + dstHost = InetAddress.getByName(dstHost).getCanonicalHostName() + } catch { + case e: UnknownHostException => + return false + } + if (!srcHost.equals(dstHost)) { + return false + } + } else if (srcHost == null && dstHost != null) { + return false + } else if (srcHost != null && dstHost == null) { + return false + } + //check for ports + if (srcUri.getPort() != dstUri.getPort()) { + return false + } + return true + } + + /** Copy the file into HDFS if needed. */ + private def copyRemoteFile( + dstDir: Path, + originalPath: Path, + replication: Short, + setPerms: Boolean = false): Path = { + val fs = FileSystem.get(conf) + val remoteFs = originalPath.getFileSystem(conf) + var newPath = originalPath + if (! compareFs(remoteFs, fs)) { + newPath = new Path(dstDir, originalPath.getName()) + logInfo("Uploading " + originalPath + " to " + newPath) + FileUtil.copy(remoteFs, originalPath, fs, newPath, false, conf) + fs.setReplication(newPath, replication) + if (setPerms) fs.setPermission(newPath, new FsPermission(APP_FILE_PERMISSION)) + } + // Resolve any symlinks in the URI path so using a "current" symlink to point to a specific + // version shows the specific version in the distributed cache configuration + val qualPath = fs.makeQualified(newPath) + val fc = FileContext.getFileContext(qualPath.toUri(), conf) + val destPath = fc.resolvePath(qualPath) + destPath + } + + def prepareLocalResources(appStagingDir: String): HashMap[String, LocalResource] = { + logInfo("Preparing Local resources") + // Upload Spark and the application JAR to the remote file system if necessary. Add them as + // local resources to the application master. + val fs = FileSystem.get(conf) + + val delegTokenRenewer = Master.getMasterPrincipal(conf) + if (UserGroupInformation.isSecurityEnabled()) { + if (delegTokenRenewer == null || delegTokenRenewer.length() == 0) { + logError("Can't get Master Kerberos principal for use as renewer") + System.exit(1) + } + } + val dst = new Path(fs.getHomeDirectory(), appStagingDir) + val replication = System.getProperty("spark.yarn.submit.file.replication", "3").toShort + + if (UserGroupInformation.isSecurityEnabled()) { + val dstFs = dst.getFileSystem(conf) + dstFs.addDelegationTokens(delegTokenRenewer, credentials) + } + + val localResources = HashMap[String, LocalResource]() + FileSystem.mkdirs(fs, dst, new FsPermission(STAGING_DIR_PERMISSION)) + + val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]() + + Map( + Client.SPARK_JAR -> System.getenv("SPARK_JAR"), Client.APP_JAR -> args.userJar, + Client.LOG4J_PROP -> System.getenv("SPARK_LOG4J_CONF") + ).foreach { case(destName, _localPath) => + val localPath: String = if (_localPath != null) _localPath.trim() else "" + if (! localPath.isEmpty()) { + var localURI = new URI(localPath) + // If not specified assume these are in the local filesystem to keep behavior like Hadoop + if (localURI.getScheme() == null) { + localURI = new URI(FileSystem.getLocal(conf).makeQualified(new Path(localPath)).toString) + } + val setPermissions = if (destName.equals(Client.APP_JAR)) true else false + val destPath = copyRemoteFile(dst, new Path(localURI), replication, setPermissions) + distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE, + destName, statCache) + } + } + + // Handle jars local to the ApplicationMaster. + if ((args.addJars != null) && (!args.addJars.isEmpty())){ + args.addJars.split(',').foreach { case file: String => + val localURI = new URI(file.trim()) + val localPath = new Path(localURI) + val linkname = Option(localURI.getFragment()).getOrElse(localPath.getName()) + val destPath = copyRemoteFile(dst, localPath, replication) + // Only add the resource to the Spark ApplicationMaster. + val appMasterOnly = true + distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE, + linkname, statCache, appMasterOnly) + } + } + + // Handle any distributed cache files + if ((args.files != null) && (!args.files.isEmpty())){ + args.files.split(',').foreach { case file: String => + val localURI = new URI(file.trim()) + val localPath = new Path(localURI) + val linkname = Option(localURI.getFragment()).getOrElse(localPath.getName()) + val destPath = copyRemoteFile(dst, localPath, replication) + distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE, + linkname, statCache) + } + } + + // Handle any distributed cache archives + if ((args.archives != null) && (!args.archives.isEmpty())) { + args.archives.split(',').foreach { case file:String => + val localURI = new URI(file.trim()) + val localPath = new Path(localURI) + val linkname = Option(localURI.getFragment()).getOrElse(localPath.getName()) + val destPath = copyRemoteFile(dst, localPath, replication) + distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.ARCHIVE, + linkname, statCache) + } + } + + UserGroupInformation.getCurrentUser().addCredentials(credentials) + localResources + } + + def setupLaunchEnv( + localResources: HashMap[String, LocalResource], + stagingDir: String): HashMap[String, String] = { + logInfo("Setting up the launch environment") + val log4jConfLocalRes = localResources.getOrElse(Client.LOG4J_PROP, null) + + val env = new HashMap[String, String]() + + Client.populateClasspath(yarnConf, log4jConfLocalRes != null, env) + env("SPARK_YARN_MODE") = "true" + env("SPARK_YARN_STAGING_DIR") = stagingDir + + // Set the environment variables to be passed on to the Workers. + distCacheMgr.setDistFilesEnv(env) + distCacheMgr.setDistArchivesEnv(env) + + // Allow users to specify some environment variables. + Apps.setEnvFromInputString(env, System.getenv("SPARK_YARN_USER_ENV")) + + // Add each SPARK_* key to the environment. + System.getenv().filterKeys(_.startsWith("SPARK")).foreach { case (k,v) => env(k) = v } + + 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) + + // TODO: Need a replacement for the following code to fix -Xmx? + // val minResMemory: Int = newApp.getMinimumResourceCapability().getMemory() + // var amMemory = ((args.amMemory / minResMemory) * minResMemory) + + // ((if ((args.amMemory % minResMemory) == 0) 0 else minResMemory) - + // YarnAllocationHandler.MEMORY_OVERHEAD) + + // Extra options for the JVM + var JAVA_OPTS = "" + + // Add Xmx for AM memory + JAVA_OPTS += "-Xmx" + args.amMemory + "m" + + val tmpDir = new Path(Environment.PWD.$(), YarnConfiguration.DEFAULT_CONTAINER_TEMP_DIR) + JAVA_OPTS += " -Djava.io.tmpdir=" + tmpDir + + // TODO: Remove once cpuset version is pushed out. + // The context is, default gc for server class machines ends up using all cores to do gc - + // hence if there are multiple containers in same node, Spark GC affects all other containers' + // performance (which can be that of other Spark containers) + // Instead of using this, rely on cpusets by YARN to enforce "proper" Spark behavior in + // multi-tenant environments. Not sure how default Java GC behaves if it is limited to subset + // of cores on a node. + val useConcurrentAndIncrementalGC = env.isDefinedAt("SPARK_USE_CONC_INCR_GC") && + java.lang.Boolean.parseBoolean(env("SPARK_USE_CONC_INCR_GC")) + if (useConcurrentAndIncrementalGC) { + // In our expts, using (default) throughput collector has severe perf ramifications 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 + var javaCommand = "java" + val javaHome = System.getenv("JAVA_HOME") + if ((javaHome != null && !javaHome.isEmpty()) || env.isDefinedAt("JAVA_HOME")) { + javaCommand = Environment.JAVA_HOME.$() + "/bin/java" + } + + val commands = List[String]( + javaCommand + + " -server " + + JAVA_OPTS + + " " + args.amClass + + " --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 starting the Spark ApplicationMaster: " + commands(0)) + amContainer.setCommands(commands) + + // Setup security tokens. + val dob = new DataOutputBuffer() + credentials.writeTokenStorageToStream(dob) + amContainer.setTokens(ByteBuffer.wrap(dob.getData())) + + 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 clientToAMToken: " + report.getClientToAMToken() + "\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 + } + } + true + } +} + +object Client { + val SPARK_JAR: String = "spark.jar" + val APP_JAR: String = "app.jar" + val LOG4J_PROP: String = "log4j.properties" + + def main(argStrings: Array[String]) { + // Set an env variable indicating we are running in YARN mode. + // Note: anything env variable with SPARK_ prefix gets propagated to all (remote) processes - + // see Client#setupLaunchEnv(). + System.setProperty("SPARK_YARN_MODE", "true") + + val args = new ClientArguments(argStrings) + + (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) + } + } + + def populateClasspath(conf: Configuration, addLog4j: Boolean, env: HashMap[String, String]) { + Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$()) + // If log4j present, ensure ours overrides all others + if (addLog4j) { + Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() + + Path.SEPARATOR + LOG4J_PROP) + } + // Normally the users app.jar is last in case conflicts with spark jars + val userClasspathFirst = System.getProperty("spark.yarn.user.classpath.first", "false") + .toBoolean + if (userClasspathFirst) { + Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() + + Path.SEPARATOR + APP_JAR) + } + Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() + + Path.SEPARATOR + SPARK_JAR) + Client.populateHadoopClasspath(conf, env) + + if (!userClasspathFirst) { + Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() + + Path.SEPARATOR + APP_JAR) + } + Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() + + Path.SEPARATOR + "*") + } +} diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala new file mode 100644 index 0000000000..9efb28a942 --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala @@ -0,0 +1,148 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import scala.collection.mutable.{ArrayBuffer, HashMap} + +import org.apache.spark.scheduler.{InputFormatInfo, SplitInfo} +import org.apache.spark.util.IntParam +import org.apache.spark.util.MemoryParam + + +// TODO: Add code and support for ensuring that yarn resource 'tasks' are location aware ! +class ClientArguments(val args: Array[String]) { + var addJars: String = null + var files: String = null + var archives: String = null + var userJar: String = null + var userClass: String = null + var userArgs: Seq[String] = Seq[String]() + var workerMemory = 1024 // MB + var workerCores = 1 + var numWorkers = 2 + var amQueue = System.getProperty("QUEUE", "default") + var amMemory: Int = 512 // MB + var amClass: String = "org.apache.spark.deploy.yarn.ApplicationMaster" + var appName: String = "Spark" + // TODO + var inputFormatInfo: List[InputFormatInfo] = null + // TODO(harvey) + var priority = 0 + + 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-class") :: value :: tail => + amClass = 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 ("--queue") :: value :: tail => + amQueue = value + args = tail + + case ("--name") :: value :: tail => + appName = value + + case ("--addJars") :: value :: tail => + addJars = value + args = tail + + case ("--files") :: value :: tail => + files = value + args = tail + + case ("--archives") :: value :: tail => + archives = 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: org.apache.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-class CLASS_NAME Class Name for Master (Default: spark.deploy.yarn.ApplicationMaster)\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" + + " --name NAME The name of your application (Default: Spark)\n" + + " --queue QUEUE The hadoop queue to use for allocation requests (Default: 'default')\n" + + " --addJars jars Comma separated list of local jars that want SparkContext.addJar to work with.\n" + + " --files files Comma separated list of files to be distributed with the job.\n" + + " --archives archives Comma separated list of archives to be distributed with the job." + ) + System.exit(exitCode) + } + +} diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala new file mode 100644 index 0000000000..5f159b073f --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala @@ -0,0 +1,228 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.net.URI + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.FileStatus +import org.apache.hadoop.fs.FileSystem +import org.apache.hadoop.fs.Path +import org.apache.hadoop.fs.permission.FsAction +import org.apache.hadoop.yarn.api.records.LocalResource +import org.apache.hadoop.yarn.api.records.LocalResourceVisibility +import org.apache.hadoop.yarn.api.records.LocalResourceType +import org.apache.hadoop.yarn.util.{Records, ConverterUtils} + +import org.apache.spark.Logging + +import scala.collection.mutable.HashMap +import scala.collection.mutable.LinkedHashMap +import scala.collection.mutable.Map + + +/** Client side methods to setup the Hadoop distributed cache */ +class ClientDistributedCacheManager() extends Logging { + private val distCacheFiles: Map[String, Tuple3[String, String, String]] = + LinkedHashMap[String, Tuple3[String, String, String]]() + private val distCacheArchives: Map[String, Tuple3[String, String, String]] = + LinkedHashMap[String, Tuple3[String, String, String]]() + + + /** + * Add a resource to the list of distributed cache resources. This list can + * be sent to the ApplicationMaster and possibly the workers so that it can + * be downloaded into the Hadoop distributed cache for use by this application. + * Adds the LocalResource to the localResources HashMap passed in and saves + * the stats of the resources to they can be sent to the workers and verified. + * + * @param fs FileSystem + * @param conf Configuration + * @param destPath path to the resource + * @param localResources localResource hashMap to insert the resource into + * @param resourceType LocalResourceType + * @param link link presented in the distributed cache to the destination + * @param statCache cache to store the file/directory stats + * @param appMasterOnly Whether to only add the resource to the app master + */ + def addResource( + fs: FileSystem, + conf: Configuration, + destPath: Path, + localResources: HashMap[String, LocalResource], + resourceType: LocalResourceType, + link: String, + statCache: Map[URI, FileStatus], + appMasterOnly: Boolean = false) = { + val destStatus = fs.getFileStatus(destPath) + val amJarRsrc = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource] + amJarRsrc.setType(resourceType) + val visibility = getVisibility(conf, destPath.toUri(), statCache) + amJarRsrc.setVisibility(visibility) + amJarRsrc.setResource(ConverterUtils.getYarnUrlFromPath(destPath)) + amJarRsrc.setTimestamp(destStatus.getModificationTime()) + amJarRsrc.setSize(destStatus.getLen()) + if (link == null || link.isEmpty()) throw new Exception("You must specify a valid link name") + localResources(link) = amJarRsrc + + if (appMasterOnly == false) { + val uri = destPath.toUri() + val pathURI = new URI(uri.getScheme(), uri.getAuthority(), uri.getPath(), null, link) + if (resourceType == LocalResourceType.FILE) { + distCacheFiles(pathURI.toString()) = (destStatus.getLen().toString(), + destStatus.getModificationTime().toString(), visibility.name()) + } else { + distCacheArchives(pathURI.toString()) = (destStatus.getLen().toString(), + destStatus.getModificationTime().toString(), visibility.name()) + } + } + } + + /** + * Adds the necessary cache file env variables to the env passed in + * @param env + */ + def setDistFilesEnv(env: Map[String, String]) = { + val (keys, tupleValues) = distCacheFiles.unzip + val (sizes, timeStamps, visibilities) = tupleValues.unzip3 + + if (keys.size > 0) { + env("SPARK_YARN_CACHE_FILES") = keys.reduceLeft[String] { (acc,n) => acc + "," + n } + env("SPARK_YARN_CACHE_FILES_TIME_STAMPS") = + timeStamps.reduceLeft[String] { (acc,n) => acc + "," + n } + env("SPARK_YARN_CACHE_FILES_FILE_SIZES") = + sizes.reduceLeft[String] { (acc,n) => acc + "," + n } + env("SPARK_YARN_CACHE_FILES_VISIBILITIES") = + visibilities.reduceLeft[String] { (acc,n) => acc + "," + n } + } + } + + /** + * Adds the necessary cache archive env variables to the env passed in + * @param env + */ + def setDistArchivesEnv(env: Map[String, String]) = { + val (keys, tupleValues) = distCacheArchives.unzip + val (sizes, timeStamps, visibilities) = tupleValues.unzip3 + + if (keys.size > 0) { + env("SPARK_YARN_CACHE_ARCHIVES") = keys.reduceLeft[String] { (acc,n) => acc + "," + n } + env("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") = + timeStamps.reduceLeft[String] { (acc,n) => acc + "," + n } + env("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") = + sizes.reduceLeft[String] { (acc,n) => acc + "," + n } + env("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") = + visibilities.reduceLeft[String] { (acc,n) => acc + "," + n } + } + } + + /** + * Returns the local resource visibility depending on the cache file permissions + * @param conf + * @param uri + * @param statCache + * @return LocalResourceVisibility + */ + def getVisibility(conf: Configuration, uri: URI, statCache: Map[URI, FileStatus]): + LocalResourceVisibility = { + if (isPublic(conf, uri, statCache)) { + return LocalResourceVisibility.PUBLIC + } + return LocalResourceVisibility.PRIVATE + } + + /** + * Returns a boolean to denote whether a cache file is visible to all(public) + * or not + * @param conf + * @param uri + * @param statCache + * @return true if the path in the uri is visible to all, false otherwise + */ + def isPublic(conf: Configuration, uri: URI, statCache: Map[URI, FileStatus]): Boolean = { + val fs = FileSystem.get(uri, conf) + val current = new Path(uri.getPath()) + //the leaf level file should be readable by others + if (!checkPermissionOfOther(fs, current, FsAction.READ, statCache)) { + return false + } + return ancestorsHaveExecutePermissions(fs, current.getParent(), statCache) + } + + /** + * Returns true if all ancestors of the specified path have the 'execute' + * permission set for all users (i.e. that other users can traverse + * the directory heirarchy to the given path) + * @param fs + * @param path + * @param statCache + * @return true if all ancestors have the 'execute' permission set for all users + */ + def ancestorsHaveExecutePermissions(fs: FileSystem, path: Path, + statCache: Map[URI, FileStatus]): Boolean = { + var current = path + while (current != null) { + //the subdirs in the path should have execute permissions for others + if (!checkPermissionOfOther(fs, current, FsAction.EXECUTE, statCache)) { + return false + } + current = current.getParent() + } + return true + } + + /** + * Checks for a given path whether the Other permissions on it + * imply the permission in the passed FsAction + * @param fs + * @param path + * @param action + * @param statCache + * @return true if the path in the uri is visible to all, false otherwise + */ + def checkPermissionOfOther(fs: FileSystem, path: Path, + action: FsAction, statCache: Map[URI, FileStatus]): Boolean = { + val status = getFileStatus(fs, path.toUri(), statCache) + val perms = status.getPermission() + val otherAction = perms.getOtherAction() + if (otherAction.implies(action)) { + return true + } + return false + } + + /** + * Checks to see if the given uri exists in the cache, if it does it + * returns the existing FileStatus, otherwise it stats the uri, stores + * it in the cache, and returns the FileStatus. + * @param fs + * @param uri + * @param statCache + * @return FileStatus + */ + def getFileStatus(fs: FileSystem, uri: URI, statCache: Map[URI, FileStatus]): FileStatus = { + val stat = statCache.get(uri) match { + case Some(existstat) => existstat + case None => + val newStat = fs.getFileStatus(new Path(uri)) + statCache.put(uri, newStat) + newStat + } + return stat + } +} diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala new file mode 100644 index 0000000000..c38f33e212 --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala @@ -0,0 +1,223 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.net.Socket +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.util.{ConverterUtils, Records} +import akka.actor._ +import akka.remote.{RemoteClientShutdown, RemoteClientDisconnected, RemoteClientLifeCycleEvent} +import akka.actor.Terminated +import org.apache.spark.{SparkContext, Logging} +import org.apache.spark.util.{Utils, AkkaUtils} +import org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend +import org.apache.spark.scheduler.SplitInfo +import org.apache.hadoop.yarn.client.api.AMRMClient +import org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest + +class WorkerLauncher(args: ApplicationMasterArguments, conf: Configuration) extends Logging { + + def this(args: ApplicationMasterArguments) = this(args, new Configuration()) + + private var appAttemptId: ApplicationAttemptId = _ + private var reporterThread: Thread = _ + private val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + + private var yarnAllocator: YarnAllocationHandler = _ + private var driverClosed:Boolean = false + + private var amClient: AMRMClient[ContainerRequest] = _ + + val actorSystem : ActorSystem = AkkaUtils.createActorSystem("sparkYarnAM", Utils.localHostName, 0)._1 + var actor: ActorRef = _ + + // This actor just working as a monitor to watch on Driver Actor. + class MonitorActor(driverUrl: String) extends Actor { + + var driver: ActorRef = null + + override def preStart() { + logInfo("Listen to driver: " + driverUrl) + driver = context.actorFor(driverUrl) + context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent]) + context.watch(driver) // Doesn't work with remote actors, but useful for testing + } + + override def receive = { + case Terminated(_) | RemoteClientDisconnected(_, _) | RemoteClientShutdown(_, _) => + logInfo("Driver terminated or disconnected! Shutting down.") + driverClosed = true + } + } + + def run() { + + amClient = AMRMClient.createAMRMClient() + amClient.init(yarnConf) + amClient.start() + + appAttemptId = getApplicationAttemptId() + registerApplicationMaster() + + waitForSparkMaster() + + // Allocate all containers + allocateWorkers() + + // Launch a progress reporter thread, else app will get killed after expiration (def: 10mins) timeout + // 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)) + reporterThread = launchReporterThread(interval) + + // Wait for the reporter thread to Finish. + reporterThread.join() + + finishApplicationMaster(FinalApplicationStatus.SUCCEEDED) + actorSystem.shutdown() + + logInfo("Exited") + System.exit(0) + } + + private def getApplicationAttemptId(): ApplicationAttemptId = { + val envs = System.getenv() + val containerIdString = envs.get(ApplicationConstants.Environment.CONTAINER_ID.name()) + val containerId = ConverterUtils.toContainerId(containerIdString) + val appAttemptId = containerId.getApplicationAttemptId() + logInfo("ApplicationAttemptId: " + appAttemptId) + appAttemptId + } + + private def registerApplicationMaster(): RegisterApplicationMasterResponse = { + logInfo("Registering the ApplicationMaster") + // TODO:(Raymond) Find out Spark UI address and fill in here? + amClient.registerApplicationMaster(Utils.localHostName(), 0, "") + } + + private def waitForSparkMaster() { + logInfo("Waiting for Spark driver to be reachable.") + var driverUp = false + val hostport = args.userArgs(0) + val (driverHost, driverPort) = Utils.parseHostPort(hostport) + while(!driverUp) { + try { + val socket = new Socket(driverHost, driverPort) + socket.close() + logInfo("Driver now available: %s:%s".format(driverHost, driverPort)) + driverUp = true + } catch { + case e: Exception => + logError("Failed to connect to driver at %s:%s, retrying ...". + format(driverHost, driverPort)) + Thread.sleep(100) + } + } + System.setProperty("spark.driver.host", driverHost) + System.setProperty("spark.driver.port", driverPort.toString) + + val driverUrl = "akka://spark@%s:%s/user/%s".format( + driverHost, driverPort.toString, CoarseGrainedSchedulerBackend.ACTOR_NAME) + + actor = actorSystem.actorOf(Props(new MonitorActor(driverUrl)), name = "YarnAM") + } + + + private def allocateWorkers() { + + // Fixme: should get preferredNodeLocationData from SparkContext, just fake a empty one for now. + val preferredNodeLocationData: scala.collection.Map[String, scala.collection.Set[SplitInfo]] = + scala.collection.immutable.Map() + + yarnAllocator = YarnAllocationHandler.newAllocator( + yarnConf, + amClient, + appAttemptId, + args, + 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 + + yarnAllocator.addResourceRequests(args.numWorkers) + while(yarnAllocator.getNumWorkersRunning < args.numWorkers) { + yarnAllocator.allocateResources() + Thread.sleep(100) + } + + logInfo("All workers have launched.") + + } + + // 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 (!driverClosed) { + val missingWorkerCount = args.numWorkers - yarnAllocator.getNumWorkersRunning - + yarnAllocator.getNumPendingAllocate + if (missingWorkerCount > 0) { + logInfo("Allocating %d containers to make up for (potentially) lost containers". + format(missingWorkerCount)) + yarnAllocator.addResourceRequests(missingWorkerCount) + } + 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) + t + } + + private def sendProgress() { + logDebug("Sending progress") + // simulated with an allocate request with no nodes requested ... + yarnAllocator.allocateResources() + } + + def finishApplicationMaster(status: FinalApplicationStatus) { + logInfo("finish ApplicationMaster with " + status) + amClient.unregisterApplicationMaster(status, "" /* appMessage */, "" /* appTrackingUrl */) + } + +} + + +object WorkerLauncher { + def main(argStrings: Array[String]) { + val args = new ApplicationMasterArguments(argStrings) + new WorkerLauncher(args).run() + } +} diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala new file mode 100644 index 0000000000..9f5523c4b9 --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala @@ -0,0 +1,209 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.net.URI +import java.nio.ByteBuffer +import java.security.PrivilegedExceptionAction + +import scala.collection.JavaConversions._ +import scala.collection.mutable.HashMap + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.Path +import org.apache.hadoop.io.DataOutputBuffer +import org.apache.hadoop.net.NetUtils +import org.apache.hadoop.security.UserGroupInformation +import org.apache.hadoop.yarn.api._ +import org.apache.hadoop.yarn.api.ApplicationConstants.Environment +import org.apache.hadoop.yarn.api.records._ +import org.apache.hadoop.yarn.api.records.impl.pb.ProtoUtils +import org.apache.hadoop.yarn.api.protocolrecords._ +import org.apache.hadoop.yarn.client.api.NMClient +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.spark.Logging + + +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 nmClient: NMClient = _ + val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + + def run = { + logInfo("Starting Worker Container") + nmClient = NMClient.createNMClient() + nmClient.init(yarnConf) + nmClient.start() + startContainer + } + + def startContainer = { + logInfo("Setting up ContainerLaunchContext") + + val ctx = Records.newRecord(classOf[ContainerLaunchContext]) + .asInstanceOf[ContainerLaunchContext] + + 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") + " " + } + + JAVA_OPTS += " -Djava.io.tmpdir=" + + new Path(Environment.PWD.$(), YarnConfiguration.DEFAULT_CONTAINER_TEMP_DIR) + " " + + // 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 " + } +*/ + + val credentials = UserGroupInformation.getCurrentUser().getCredentials() + val dob = new DataOutputBuffer() + credentials.writeTokenStorageToStream(dob) + ctx.setTokens(ByteBuffer.wrap(dob.getData())) + + var javaCommand = "java" + val javaHome = System.getenv("JAVA_HOME") + if ((javaHome != null && !javaHome.isEmpty()) || env.isDefinedAt("JAVA_HOME")) { + javaCommand = Environment.JAVA_HOME.$() + "/bin/java" + } + + val commands = List[String](javaCommand + + " -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 + + " org.apache.spark.executor.CoarseGrainedExecutorBackend " + + 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 + nmClient.startContainer(container, ctx) + } + + private def setupDistributedCache( + file: String, + rtype: LocalResourceType, + localResources: HashMap[String, LocalResource], + timestamp: String, + size: String, + vis: String) = { + val uri = new URI(file) + val amJarRsrc = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource] + amJarRsrc.setType(rtype) + amJarRsrc.setVisibility(LocalResourceVisibility.valueOf(vis)) + amJarRsrc.setResource(ConverterUtils.getYarnUrlFromURI(uri)) + amJarRsrc.setTimestamp(timestamp.toLong) + amJarRsrc.setSize(size.toLong) + localResources(uri.getFragment()) = amJarRsrc + } + + def prepareLocalResources: HashMap[String, LocalResource] = { + logInfo("Preparing Local resources") + val localResources = HashMap[String, LocalResource]() + + if (System.getenv("SPARK_YARN_CACHE_FILES") != null) { + val timeStamps = System.getenv("SPARK_YARN_CACHE_FILES_TIME_STAMPS").split(',') + val fileSizes = System.getenv("SPARK_YARN_CACHE_FILES_FILE_SIZES").split(',') + val distFiles = System.getenv("SPARK_YARN_CACHE_FILES").split(',') + val visibilities = System.getenv("SPARK_YARN_CACHE_FILES_VISIBILITIES").split(',') + for( i <- 0 to distFiles.length - 1) { + setupDistributedCache(distFiles(i), LocalResourceType.FILE, localResources, timeStamps(i), + fileSizes(i), visibilities(i)) + } + } + + if (System.getenv("SPARK_YARN_CACHE_ARCHIVES") != null) { + val timeStamps = System.getenv("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS").split(',') + val fileSizes = System.getenv("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES").split(',') + val distArchives = System.getenv("SPARK_YARN_CACHE_ARCHIVES").split(',') + val visibilities = System.getenv("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES").split(',') + for( i <- 0 to distArchives.length - 1) { + setupDistributedCache(distArchives(i), LocalResourceType.ARCHIVE, localResources, + timeStamps(i), fileSizes(i), visibilities(i)) + } + } + + logInfo("Prepared Local resources " + localResources) + localResources + } + + def prepareEnvironment: HashMap[String, String] = { + val env = new HashMap[String, String]() + + Client.populateClasspath(yarnConf, System.getenv("SPARK_YARN_LOG4J_PATH") != null, env) + + // Allow users to specify some environment variables + Apps.setEnvFromInputString(env, System.getenv("SPARK_YARN_USER_ENV")) + + System.getenv().filterKeys(_.startsWith("SPARK")).foreach { case (k,v) => env(k) = v } + env + } + +} diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala new file mode 100644 index 0000000000..dba0f7640e --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala @@ -0,0 +1,687 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.lang.{Boolean => JBoolean} +import java.util.{Collections, Set => JSet} +import java.util.concurrent.{CopyOnWriteArrayList, ConcurrentHashMap} +import java.util.concurrent.atomic.AtomicInteger + +import scala.collection +import scala.collection.JavaConversions._ +import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet} + +import org.apache.spark.Logging +import org.apache.spark.scheduler.SplitInfo +import org.apache.spark.scheduler.cluster.{ClusterScheduler, CoarseGrainedSchedulerBackend} +import org.apache.spark.util.Utils + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.yarn.api.ApplicationMasterProtocol +import org.apache.hadoop.yarn.api.records.ApplicationAttemptId +import org.apache.hadoop.yarn.api.records.{Container, ContainerId, ContainerStatus} +import org.apache.hadoop.yarn.api.records.{Priority, Resource, ResourceRequest} +import org.apache.hadoop.yarn.api.protocolrecords.{AllocateRequest, AllocateResponse} +import org.apache.hadoop.yarn.client.api.AMRMClient +import org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest +import org.apache.hadoop.yarn.util.{RackResolver, Records} + + +object AllocationType extends Enumeration ("HOST", "RACK", "ANY") { + type AllocationType = Value + val HOST, RACK, ANY = Value +} + +// TODO: +// Too many params. +// 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 amClient: AMRMClient[ContainerRequest], + 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] + + // Number of container requests that have been sent to, but not yet allocated by the + // ApplicationMaster. + private val numPendingAllocate = new AtomicInteger() + private val numWorkersRunning = new AtomicInteger() + // Used to generate a unique id per worker + private val workerIdCounter = new AtomicInteger() + private val lastResponseId = new AtomicInteger() + private val numWorkersFailed = new AtomicInteger() + + def getNumPendingAllocate: Int = numPendingAllocate.intValue + + def getNumWorkersRunning: Int = numWorkersRunning.intValue + + def getNumWorkersFailed: Int = numWorkersFailed.intValue + + def isResourceConstraintSatisfied(container: Container): Boolean = { + container.getResource.getMemory >= (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + } + + def releaseContainer(container: Container) { + val containerId = container.getId + pendingReleaseContainers.put(containerId, true) + amClient.releaseAssignedContainer(containerId) + } + + def allocateResources() { + // We have already set the container request. Poll the ResourceManager for a response. + // This doubles as a heartbeat if there are no pending container requests. + val progressIndicator = 0.1f + val allocateResponse = amClient.allocate(progressIndicator) + + val allocatedContainers = allocateResponse.getAllocatedContainers() + if (allocatedContainers.size > 0) { + var numPendingAllocateNow = numPendingAllocate.addAndGet(-1 * allocatedContainers.size) + + if (numPendingAllocateNow < 0) { + numPendingAllocateNow = numPendingAllocate.addAndGet(-1 * numPendingAllocateNow) + } + + logDebug(""" + Allocated containers: %d + Current worker count: %d + Containers released: %s + Containers to-be-released: %s + Cluster resources: %s + """.format( + allocatedContainers.size, + numWorkersRunning.get(), + releasedContainerList, + pendingReleaseContainers, + allocateResponse.getAvailableResources)) + + val hostToContainers = new HashMap[String, ArrayBuffer[Container]]() + + for (container <- allocatedContainers) { + if (isResourceConstraintSatisfied(container)) { + // Add the accepted `container` to the host's list of already accepted, + // allocated containers + val host = container.getNodeId.getHost + val containersForHost = hostToContainers.getOrElseUpdate(host, + new ArrayBuffer[Container]()) + containersForHost += container + } else { + // Release container, since it doesn't satisfy resource constraints. + releaseContainer(container) + } + } + + // Find the appropriate containers to use. + // TODO: Cleanup this group-by... + 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) + + val remainingContainersOpt = hostToContainers.get(candidateHost) + assert(remainingContainersOpt.isDefined) + var remainingContainers = remainingContainersOpt.get + + if (requiredHostCount >= remainingContainers.size) { + // Since we have <= required containers, add all remaining containers to + // `dataLocalContainers`. + dataLocalContainers.put(candidateHost, remainingContainers) + // There are no more free containers remaining. + remainingContainers = null + } else if (requiredHostCount > 0) { + // Container list has more containers than we need for data locality. + // Split the list into two: one based on the data local container count, + // (`remainingContainers.size` - `requiredHostCount`), and the other to hold remaining + // containers. + val (dataLocal, remaining) = remainingContainers.splitAt( + remainingContainers.size - requiredHostCount) + dataLocalContainers.put(candidateHost, dataLocal) + + // Invariant: remainingContainers == remaining + + // YARN has a nasty habit of allocating a ton of containers on a host - discourage this. + // Add each container in `remaining` to list of containers to release. If we have an + // insufficient number of containers, then the next allocation cycle will reallocate + // (but won't treat it as data local). + // TODO(harvey): Rephrase this comment some more. + for (container <- remaining) releaseContainer(container) + remainingContainers = null + } + + // For rack local containers + if (remainingContainers != null) { + val rack = YarnAllocationHandler.lookupRack(conf, candidateHost) + if (rack != null) { + val maxExpectedRackCount = preferredRackToCount.getOrElse(rack, 0) + val requiredRackCount = maxExpectedRackCount - allocatedContainersOnRack(rack) - + rackLocalContainers.getOrElse(rack, List()).size + + if (requiredRackCount >= remainingContainers.size) { + // Add all remaining containers to to `dataLocalContainers`. + dataLocalContainers.put(rack, remainingContainers) + remainingContainers = null + } else if (requiredRackCount > 0) { + // Container list has more containers that we need for data locality. + // Split the list into two: one based on the data local container count, + // (`remainingContainers.size` - `requiredHostCount`), and the other to hold remaining + // containers. + val (rackLocal, remaining) = remainingContainers.splitAt( + remainingContainers.size - requiredRackCount) + val existingRackLocal = rackLocalContainers.getOrElseUpdate(rack, + new ArrayBuffer[Container]()) + + existingRackLocal ++= rackLocal + + remainingContainers = remaining + } + } + } + + if (remainingContainers != null) { + // Not all containers have been consumed - add them to the list of off-rack containers. + 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 finally off-rack. + // Note that the list we create below tries to ensure that not all containers end up within + // a host if there is a sufficiently large number of hosts/containers. + val allocatedContainersToProcess = new ArrayBuffer[Container](allocatedContainers.size) + allocatedContainersToProcess ++= ClusterScheduler.prioritizeContainers(dataLocalContainers) + allocatedContainersToProcess ++= ClusterScheduler.prioritizeContainers(rackLocalContainers) + allocatedContainersToProcess ++= ClusterScheduler.prioritizeContainers(offRackContainers) + + // Run each of the allocated containers. + for (container <- allocatedContainersToProcess) { + val numWorkersRunningNow = numWorkersRunning.incrementAndGet() + val workerHostname = container.getNodeId.getHost + val containerId = container.getId + + val workerMemoryOverhead = (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + assert(container.getResource.getMemory >= workerMemoryOverhead) + + if (numWorkersRunningNow > maxWorkers) { + logInfo("""Ignoring container %s at host %s, since we already have the required number of + containers for it.""".format(containerId, workerHostname)) + releaseContainer(container) + numWorkersRunning.decrementAndGet() + } else { + val workerId = workerIdCounter.incrementAndGet().toString + val driverUrl = "akka://spark@%s:%s/user/%s".format( + System.getProperty("spark.driver.host"), + System.getProperty("spark.driver.port"), + CoarseGrainedSchedulerBackend.ACTOR_NAME) + + logInfo("Launching container %s for on host %s".format(containerId, workerHostname)) + + // To be safe, remove the container from `pendingReleaseContainers`. + 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) + } + } + logInfo("Launching WorkerRunnable. driverUrl: %s, workerHostname: %s".format(driverUrl, workerHostname)) + val workerRunnable = new WorkerRunnable( + container, + conf, + driverUrl, + workerId, + workerHostname, + workerMemory, + workerCores) + new Thread(workerRunnable).start() + } + } + logDebug(""" + Finished allocating %s containers (from %s originally). + Current number of workers running: %d, + releasedContainerList: %s, + pendingReleaseContainers: %s + """.format( + allocatedContainersToProcess, + allocatedContainers, + numWorkersRunning.get(), + releasedContainerList, + pendingReleaseContainers)) + } + + val completedContainers = allocateResponse.getCompletedContainersStatuses() + if (completedContainers.size > 0) { + logDebug("Completed %d containers".format(completedContainers.size)) + + for (completedContainer <- completedContainers) { + val containerId = completedContainer.getContainerId + + if (pendingReleaseContainers.containsKey(containerId)) { + // YarnAllocationHandler already marked the container for release, so remove it from + // `pendingReleaseContainers`. + pendingReleaseContainers.remove(containerId) + } else { + // Decrement the number of workers running. The next iteration of the ApplicationMaster's + // reporting thread will take care of allocating. + numWorkersRunning.decrementAndGet() + logInfo("Completed container %s (state: %s, exit status: %s)".format( + containerId, + completedContainer.getState, + completedContainer.getExitStatus())) + // Hadoop 2.2.X added a ContainerExitStatus we should switch to use + // there are some exit status' we shouldn't necessarily count against us, but for + // now I think its ok as none of the containers are expected to exit + if (completedContainer.getExitStatus() != 0) { + logInfo("Container marked as failed: " + containerId) + numWorkersFailed.incrementAndGet() + } + } + + allocatedHostToContainersMap.synchronized { + if (allocatedContainerToHostMap.containsKey(containerId)) { + val hostOpt = allocatedContainerToHostMap.get(containerId) + assert(hostOpt.isDefined) + val host = hostOpt.get + + val containerSetOpt = allocatedHostToContainersMap.get(host) + assert(containerSetOpt.isDefined) + val containerSet = containerSetOpt.get + + containerSet.remove(containerId) + if (containerSet.isEmpty) { + allocatedHostToContainersMap.remove(host) + } else { + allocatedHostToContainersMap.update(host, containerSet) + } + + allocatedContainerToHostMap.remove(containerId) + + // TODO: Move this part outside the synchronized block? + 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(""" + Finished processing %d completed containers. + Current number of workers running: %d, + releasedContainerList: %s, + pendingReleaseContainers: %s + """.format( + completedContainers.size, + numWorkersRunning.get(), + releasedContainerList, + pendingReleaseContainers)) + } + } + + def createRackResourceRequests( + hostContainers: ArrayBuffer[ContainerRequest] + ): ArrayBuffer[ContainerRequest] = { + // Generate modified racks and new set of hosts under it before issuing requests. + val rackToCounts = new HashMap[String, Int]() + + for (container <- hostContainers) { + val candidateHost = container.getNodes.last + assert(YarnAllocationHandler.ANY_HOST != candidateHost) + + val rack = YarnAllocationHandler.lookupRack(conf, candidateHost) + if (rack != null) { + var count = rackToCounts.getOrElse(rack, 0) + count += 1 + rackToCounts.put(rack, count) + } + } + + val requestedContainers = new ArrayBuffer[ContainerRequest](rackToCounts.size) + for ((rack, count) <- rackToCounts) { + requestedContainers ++= createResourceRequests( + AllocationType.RACK, + rack, + count, + YarnAllocationHandler.PRIORITY) + } + + requestedContainers + } + + 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 + } + + def addResourceRequests(numWorkers: Int) { + val containerRequests: List[ContainerRequest] = + if (numWorkers <= 0 || preferredHostToCount.isEmpty) { + logDebug("numWorkers: " + numWorkers + ", host preferences: " + + preferredHostToCount.isEmpty) + createResourceRequests( + AllocationType.ANY, + resource = null, + numWorkers, + YarnAllocationHandler.PRIORITY).toList + } else { + // Request for all hosts in preferred nodes and for numWorkers - + // candidates.size, request by default allocation policy. + val hostContainerRequests = new ArrayBuffer[ContainerRequest](preferredHostToCount.size) + for ((candidateHost, candidateCount) <- preferredHostToCount) { + val requiredCount = candidateCount - allocatedContainersOnHost(candidateHost) + + if (requiredCount > 0) { + hostContainerRequests ++= createResourceRequests( + AllocationType.HOST, + candidateHost, + requiredCount, + YarnAllocationHandler.PRIORITY) + } + } + val rackContainerRequests: List[ContainerRequest] = createRackResourceRequests( + hostContainerRequests).toList + + val anyContainerRequests = createResourceRequests( + AllocationType.ANY, + resource = null, + numWorkers, + YarnAllocationHandler.PRIORITY) + + val containerRequestBuffer = new ArrayBuffer[ContainerRequest]( + hostContainerRequests.size + rackContainerRequests.size() + anyContainerRequests.size) + + containerRequestBuffer ++= hostContainerRequests + containerRequestBuffer ++= rackContainerRequests + containerRequestBuffer ++= anyContainerRequests + containerRequestBuffer.toList + } + + for (request <- containerRequests) { + amClient.addContainerRequest(request) + } + + if (numWorkers > 0) { + numPendingAllocate.addAndGet(numWorkers) + logInfo("Will Allocate %d worker containers, each with %d memory".format( + numWorkers, + (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD))) + } else { + logDebug("Empty allocation request ...") + } + + for (request <- containerRequests) { + val nodes = request.getNodes + var hostStr = if (nodes == null || nodes.isEmpty) { + "Any" + } else { + nodes.last + } + logInfo("Container request (host: %s, priority: %s, capability: %s".format( + hostStr, + request.getPriority().getPriority, + request.getCapability)) + } + } + + private def createResourceRequests( + requestType: AllocationType.AllocationType, + resource: String, + numWorkers: Int, + priority: Int + ): ArrayBuffer[ContainerRequest] = { + + // If hostname is specified, then we need at least 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 = constructContainerRequests( + Array(hostname), + racks = null, + numWorkers, + priority) + + // Add `hostname` to the global (singleton) host->rack mapping in YarnAllocationHandler. + YarnAllocationHandler.populateRackInfo(conf, hostname) + nodeLocal + } + case AllocationType.RACK => { + val rack = resource + constructContainerRequests(hosts = null, Array(rack), numWorkers, priority) + } + case AllocationType.ANY => constructContainerRequests( + hosts = null, racks = null, numWorkers, priority) + case _ => throw new IllegalArgumentException( + "Unexpected/unsupported request type: " + requestType) + } + } + + private def constructContainerRequests( + hosts: Array[String], + racks: Array[String], + numWorkers: Int, + priority: Int + ): ArrayBuffer[ContainerRequest] = { + + val memoryResource = Records.newRecord(classOf[Resource]) + memoryResource.setMemory(workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + + val prioritySetting = Records.newRecord(classOf[Priority]) + prioritySetting.setPriority(priority) + + val requests = new ArrayBuffer[ContainerRequest]() + for (i <- 0 until numWorkers) { + requests += new ContainerRequest(memoryResource, hosts, racks, prioritySetting) + } + requests + } +} + +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 ResurceManager 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, + amClient: AMRMClient[ContainerRequest], + appAttemptId: ApplicationAttemptId, + args: ApplicationMasterArguments + ): YarnAllocationHandler = { + new YarnAllocationHandler( + conf, + amClient, + appAttemptId, + args.numWorkers, + args.workerMemory, + args.workerCores, + Map[String, Int](), + Map[String, Int]()) + } + + def newAllocator( + conf: Configuration, + amClient: AMRMClient[ContainerRequest], + appAttemptId: ApplicationAttemptId, + args: ApplicationMasterArguments, + map: collection.Map[String, + collection.Set[SplitInfo]] + ): YarnAllocationHandler = { + val (hostToSplitCount, rackToSplitCount) = generateNodeToWeight(conf, map) + new YarnAllocationHandler( + conf, + amClient, + appAttemptId, + args.numWorkers, + args.workerMemory, + args.workerCores, + hostToSplitCount, + rackToSplitCount) + } + + def newAllocator( + conf: Configuration, + amClient: AMRMClient[ContainerRequest], + 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, + amClient, + 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]] + ): (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]] = { + Option(rackToHostSet.get(rack)).map { set => + val convertedSet: collection.mutable.Set[String] = set + // TODO: Better way to get a Set[String] from JSet. + 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) + + // TODO(harvey): Figure out what this comment means... + // 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/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala new file mode 100644 index 0000000000..2ba2366ead --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala @@ -0,0 +1,43 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import org.apache.spark.deploy.SparkHadoopUtil +import org.apache.hadoop.mapred.JobConf +import org.apache.hadoop.security.UserGroupInformation +import org.apache.hadoop.yarn.conf.YarnConfiguration +import org.apache.hadoop.conf.Configuration + +/** + * Contains util methods to interact with Hadoop from spark. + */ +class YarnSparkHadoopUtil extends SparkHadoopUtil { + + // 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. + override def isYarnMode(): Boolean = { true } + + // Return an appropriate (subclass) of Configuration. Creating config can initializes some hadoop subsystems + // Always create a new config, dont reuse yarnConf. + override def newConfiguration(): Configuration = new YarnConfiguration(new Configuration()) + + // add any user credentials to the job conf which are necessary for running on a secure Hadoop cluster + override def addCredentials(conf: JobConf) { + val jobCreds = conf.getCredentials() + jobCreds.mergeAll(UserGroupInformation.getCurrentUser().getCredentials()) + } +} diff --git a/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala new file mode 100644 index 0000000000..63a0449e5a --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala @@ -0,0 +1,47 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.scheduler.cluster + +import org.apache.spark._ +import org.apache.hadoop.conf.Configuration +import org.apache.spark.deploy.yarn.YarnAllocationHandler +import org.apache.spark.util.Utils + +/** + * + * This scheduler launch worker through Yarn - by call into Client to launch WorkerLauncher as AM. + */ +private[spark] class YarnClientClusterScheduler(sc: SparkContext, conf: Configuration) extends ClusterScheduler(sc) { + + def this(sc: SparkContext) = this(sc, new Configuration()) + + // 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 + } + + override def postStartHook() { + + // The yarn application is running, but the worker might not yet ready + // Wait for a few seconds for the slaves to bootstrap and register with master - best case attempt + Thread.sleep(2000L) + logInfo("YarnClientClusterScheduler.postStartHook done") + } +} diff --git a/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala new file mode 100644 index 0000000000..b206780c78 --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala @@ -0,0 +1,109 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.scheduler.cluster + +import org.apache.hadoop.yarn.api.records.{ApplicationId, YarnApplicationState} +import org.apache.spark.{SparkException, Logging, SparkContext} +import org.apache.spark.deploy.yarn.{Client, ClientArguments} + +private[spark] class YarnClientSchedulerBackend( + scheduler: ClusterScheduler, + sc: SparkContext) + extends CoarseGrainedSchedulerBackend(scheduler, sc.env.actorSystem) + with Logging { + + var client: Client = null + var appId: ApplicationId = null + + override def start() { + super.start() + + val defalutWorkerCores = "2" + val defalutWorkerMemory = "512m" + val defaultWorkerNumber = "1" + + val userJar = System.getenv("SPARK_YARN_APP_JAR") + var workerCores = System.getenv("SPARK_WORKER_CORES") + var workerMemory = System.getenv("SPARK_WORKER_MEMORY") + var workerNumber = System.getenv("SPARK_WORKER_INSTANCES") + + if (userJar == null) + throw new SparkException("env SPARK_YARN_APP_JAR is not set") + + if (workerCores == null) + workerCores = defalutWorkerCores + if (workerMemory == null) + workerMemory = defalutWorkerMemory + if (workerNumber == null) + workerNumber = defaultWorkerNumber + + val driverHost = System.getProperty("spark.driver.host") + val driverPort = System.getProperty("spark.driver.port") + val hostport = driverHost + ":" + driverPort + + val argsArray = Array[String]( + "--class", "notused", + "--jar", userJar, + "--args", hostport, + "--worker-memory", workerMemory, + "--worker-cores", workerCores, + "--num-workers", workerNumber, + "--master-class", "org.apache.spark.deploy.yarn.WorkerLauncher" + ) + + val args = new ClientArguments(argsArray) + client = new Client(args) + appId = client.runApp() + waitForApp() + } + + def waitForApp() { + + // TODO : need a better way to find out whether the workers are ready or not + // maybe by resource usage report? + while(true) { + val report = client.getApplicationReport(appId) + + logInfo("Application report from ASM: \n" + + "\t appMasterRpcPort: " + report.getRpcPort() + "\n" + + "\t appStartTime: " + report.getStartTime() + "\n" + + "\t yarnAppState: " + report.getYarnApplicationState() + "\n" + ) + + // Ready to go, or already gone. + val state = report.getYarnApplicationState() + if (state == YarnApplicationState.RUNNING) { + return + } else if (state == YarnApplicationState.FINISHED || + state == YarnApplicationState.FAILED || + state == YarnApplicationState.KILLED) { + throw new SparkException("Yarn application already ended," + + "might be killed or not able to launch application master.") + } + + Thread.sleep(1000) + } + } + + override def stop() { + super.stop() + client.stop() + logInfo("Stoped") + } + +} diff --git a/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClusterScheduler.scala b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClusterScheduler.scala new file mode 100644 index 0000000000..29b3f22e13 --- /dev/null +++ b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClusterScheduler.scala @@ -0,0 +1,55 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.scheduler.cluster + +import org.apache.spark._ +import org.apache.spark.deploy.yarn.{ApplicationMaster, YarnAllocationHandler} +import org.apache.spark.util.Utils +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) { + + logInfo("Created YarnClusterScheduler") + + 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 + } + + 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/new-yarn/src/test/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManagerSuite.scala b/new-yarn/src/test/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManagerSuite.scala new file mode 100644 index 0000000000..2941356bc5 --- /dev/null +++ b/new-yarn/src/test/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManagerSuite.scala @@ -0,0 +1,220 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.net.URI + +import org.scalatest.FunSuite +import org.scalatest.mock.MockitoSugar +import org.mockito.Mockito.when + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.FileStatus +import org.apache.hadoop.fs.FileSystem +import org.apache.hadoop.fs.Path +import org.apache.hadoop.fs.permission.FsAction +import org.apache.hadoop.yarn.api.records.LocalResource +import org.apache.hadoop.yarn.api.records.LocalResourceVisibility +import org.apache.hadoop.yarn.api.records.LocalResourceType +import org.apache.hadoop.yarn.util.{Records, ConverterUtils} + +import scala.collection.mutable.HashMap +import scala.collection.mutable.Map + + +class ClientDistributedCacheManagerSuite extends FunSuite with MockitoSugar { + + class MockClientDistributedCacheManager extends ClientDistributedCacheManager { + override def getVisibility(conf: Configuration, uri: URI, statCache: Map[URI, FileStatus]): + LocalResourceVisibility = { + return LocalResourceVisibility.PRIVATE + } + } + + test("test getFileStatus empty") { + val distMgr = new ClientDistributedCacheManager() + val fs = mock[FileSystem] + val uri = new URI("/tmp/testing") + when(fs.getFileStatus(new Path(uri))).thenReturn(new FileStatus()) + val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]() + val stat = distMgr.getFileStatus(fs, uri, statCache) + assert(stat.getPath() === null) + } + + test("test getFileStatus cached") { + val distMgr = new ClientDistributedCacheManager() + val fs = mock[FileSystem] + val uri = new URI("/tmp/testing") + val realFileStatus = new FileStatus(10, false, 1, 1024, 10, 10, null, "testOwner", + null, new Path("/tmp/testing")) + when(fs.getFileStatus(new Path(uri))).thenReturn(new FileStatus()) + val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus](uri -> realFileStatus) + val stat = distMgr.getFileStatus(fs, uri, statCache) + assert(stat.getPath().toString() === "/tmp/testing") + } + + test("test addResource") { + val distMgr = new MockClientDistributedCacheManager() + val fs = mock[FileSystem] + val conf = new Configuration() + val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing") + val localResources = HashMap[String, LocalResource]() + val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]() + when(fs.getFileStatus(destPath)).thenReturn(new FileStatus()) + + distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE, "link", + statCache, false) + val resource = localResources("link") + assert(resource.getVisibility() === LocalResourceVisibility.PRIVATE) + assert(ConverterUtils.getPathFromYarnURL(resource.getResource()) === destPath) + assert(resource.getTimestamp() === 0) + assert(resource.getSize() === 0) + assert(resource.getType() === LocalResourceType.FILE) + + val env = new HashMap[String, String]() + distMgr.setDistFilesEnv(env) + assert(env("SPARK_YARN_CACHE_FILES") === "file:/foo.invalid.com:8080/tmp/testing#link") + assert(env("SPARK_YARN_CACHE_FILES_TIME_STAMPS") === "0") + assert(env("SPARK_YARN_CACHE_FILES_FILE_SIZES") === "0") + assert(env("SPARK_YARN_CACHE_FILES_VISIBILITIES") === LocalResourceVisibility.PRIVATE.name()) + + distMgr.setDistArchivesEnv(env) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES") === None) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") === None) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") === None) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") === None) + + //add another one and verify both there and order correct + val realFileStatus = new FileStatus(20, false, 1, 1024, 10, 30, null, "testOwner", + null, new Path("/tmp/testing2")) + val destPath2 = new Path("file:///foo.invalid.com:8080/tmp/testing2") + when(fs.getFileStatus(destPath2)).thenReturn(realFileStatus) + distMgr.addResource(fs, conf, destPath2, localResources, LocalResourceType.FILE, "link2", + statCache, false) + val resource2 = localResources("link2") + assert(resource2.getVisibility() === LocalResourceVisibility.PRIVATE) + assert(ConverterUtils.getPathFromYarnURL(resource2.getResource()) === destPath2) + assert(resource2.getTimestamp() === 10) + assert(resource2.getSize() === 20) + assert(resource2.getType() === LocalResourceType.FILE) + + val env2 = new HashMap[String, String]() + distMgr.setDistFilesEnv(env2) + val timestamps = env2("SPARK_YARN_CACHE_FILES_TIME_STAMPS").split(',') + val files = env2("SPARK_YARN_CACHE_FILES").split(',') + val sizes = env2("SPARK_YARN_CACHE_FILES_FILE_SIZES").split(',') + val visibilities = env2("SPARK_YARN_CACHE_FILES_VISIBILITIES") .split(',') + assert(files(0) === "file:/foo.invalid.com:8080/tmp/testing#link") + assert(timestamps(0) === "0") + assert(sizes(0) === "0") + assert(visibilities(0) === LocalResourceVisibility.PRIVATE.name()) + + assert(files(1) === "file:/foo.invalid.com:8080/tmp/testing2#link2") + assert(timestamps(1) === "10") + assert(sizes(1) === "20") + assert(visibilities(1) === LocalResourceVisibility.PRIVATE.name()) + } + + test("test addResource link null") { + val distMgr = new MockClientDistributedCacheManager() + val fs = mock[FileSystem] + val conf = new Configuration() + val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing") + val localResources = HashMap[String, LocalResource]() + val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]() + when(fs.getFileStatus(destPath)).thenReturn(new FileStatus()) + + intercept[Exception] { + distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE, null, + statCache, false) + } + assert(localResources.get("link") === None) + assert(localResources.size === 0) + } + + test("test addResource appmaster only") { + val distMgr = new MockClientDistributedCacheManager() + val fs = mock[FileSystem] + val conf = new Configuration() + val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing") + val localResources = HashMap[String, LocalResource]() + val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]() + val realFileStatus = new FileStatus(20, false, 1, 1024, 10, 30, null, "testOwner", + null, new Path("/tmp/testing")) + when(fs.getFileStatus(destPath)).thenReturn(realFileStatus) + + distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.ARCHIVE, "link", + statCache, true) + val resource = localResources("link") + assert(resource.getVisibility() === LocalResourceVisibility.PRIVATE) + assert(ConverterUtils.getPathFromYarnURL(resource.getResource()) === destPath) + assert(resource.getTimestamp() === 10) + assert(resource.getSize() === 20) + assert(resource.getType() === LocalResourceType.ARCHIVE) + + val env = new HashMap[String, String]() + distMgr.setDistFilesEnv(env) + assert(env.get("SPARK_YARN_CACHE_FILES") === None) + assert(env.get("SPARK_YARN_CACHE_FILES_TIME_STAMPS") === None) + assert(env.get("SPARK_YARN_CACHE_FILES_FILE_SIZES") === None) + assert(env.get("SPARK_YARN_CACHE_FILES_VISIBILITIES") === None) + + distMgr.setDistArchivesEnv(env) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES") === None) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") === None) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") === None) + assert(env.get("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") === None) + } + + test("test addResource archive") { + val distMgr = new MockClientDistributedCacheManager() + val fs = mock[FileSystem] + val conf = new Configuration() + val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing") + val localResources = HashMap[String, LocalResource]() + val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]() + val realFileStatus = new FileStatus(20, false, 1, 1024, 10, 30, null, "testOwner", + null, new Path("/tmp/testing")) + when(fs.getFileStatus(destPath)).thenReturn(realFileStatus) + + distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.ARCHIVE, "link", + statCache, false) + val resource = localResources("link") + assert(resource.getVisibility() === LocalResourceVisibility.PRIVATE) + assert(ConverterUtils.getPathFromYarnURL(resource.getResource()) === destPath) + assert(resource.getTimestamp() === 10) + assert(resource.getSize() === 20) + assert(resource.getType() === LocalResourceType.ARCHIVE) + + val env = new HashMap[String, String]() + + distMgr.setDistArchivesEnv(env) + assert(env("SPARK_YARN_CACHE_ARCHIVES") === "file:/foo.invalid.com:8080/tmp/testing#link") + assert(env("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") === "10") + assert(env("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") === "20") + assert(env("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") === LocalResourceVisibility.PRIVATE.name()) + + distMgr.setDistFilesEnv(env) + assert(env.get("SPARK_YARN_CACHE_FILES") === None) + assert(env.get("SPARK_YARN_CACHE_FILES_TIME_STAMPS") === None) + assert(env.get("SPARK_YARN_CACHE_FILES_FILE_SIZES") === None) + assert(env.get("SPARK_YARN_CACHE_FILES_VISIBILITIES") === None) + } + + +} @@ -102,7 +102,9 @@ <java.version>1.5</java.version> <scala.version>2.9.3</scala.version> <mesos.version>0.13.0</mesos.version> + <akka.group>com.typesafe.akka</akka.group> <akka.version>2.0.5</akka.version> + <protobuf.version>2.4.1</protobuf.version> <slf4j.version>1.7.2</slf4j.version> <log4j.version>1.2.17</log4j.version> <hadoop.version>1.0.4</hadoop.version> @@ -245,7 +247,7 @@ <dependency> <groupId>com.google.protobuf</groupId> <artifactId>protobuf-java</artifactId> - <version>2.4.1</version> + <version>${protobuf.version}</version> </dependency> <dependency> <groupId>com.twitter</groupId> @@ -258,7 +260,7 @@ <version>0.3.1</version> </dependency> <dependency> - <groupId>com.typesafe.akka</groupId> + <groupId>${akka.group}</groupId> <artifactId>akka-actor</artifactId> <version>${akka.version}</version> <exclusions> @@ -269,7 +271,7 @@ </exclusions> </dependency> <dependency> - <groupId>com.typesafe.akka</groupId> + <groupId>${akka.group}</groupId> <artifactId>akka-remote</artifactId> <version>${akka.version}</version> <exclusions> @@ -280,7 +282,7 @@ </exclusions> </dependency> <dependency> - <groupId>com.typesafe.akka</groupId> + <groupId>${akka.group}</groupId> <artifactId>akka-slf4j</artifactId> <version>${akka.version}</version> <exclusions> @@ -291,6 +293,17 @@ </exclusions> </dependency> <dependency> + <groupId>${akka.group}</groupId> + <artifactId>akka-zeromq</artifactId> + <version>${akka.version}</version> + <exclusions> + <exclusion> + <groupId>org.jboss.netty</groupId> + <artifactId>netty</artifactId> + </exclusion> + </exclusions> + </dependency> + <dependency> <groupId>it.unimi.dsi</groupId> <artifactId>fastutil</artifactId> <version>6.4.4</version> @@ -301,11 +314,6 @@ <version>1.2.0</version> </dependency> <dependency> - <groupId>com.github.scala-incubator.io</groupId> - <artifactId>scala-io-file_2.9.2</artifactId> - <version>0.4.1</version> - </dependency> - <dependency> <groupId>org.apache.mesos</groupId> <artifactId>mesos</artifactId> <version>${mesos.version}</version> @@ -347,6 +355,11 @@ <version>3.0.0</version> </dependency> <dependency> + <groupId>com.codahale.metrics</groupId> + <artifactId>metrics-graphite</artifactId> + <version>3.0.0</version> + </dependency> + <dependency> <groupId>org.scala-lang</groupId> <artifactId>scala-compiler</artifactId> <version>${scala.version}</version> @@ -765,6 +778,41 @@ </dependencies> </dependencyManagement> </profile> + + <profile> + <id>new-yarn</id> + <properties> + <akka.group>org.spark-project</akka.group> + <akka.version>2.0.5-protobuf-2.5-java-1.5</akka.version> + <hadoop.major.version>2</hadoop.major.version> + <hadoop.version>2.2.0</hadoop.version> + <protobuf.version>2.5.0</protobuf.version> + </properties> + + <modules> + <module>new-yarn</module> + </modules> + + <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> + </dependencies> + </dependencyManagement> + </profile> + <profile> <id>repl-bin</id> <activation> diff --git a/project/SparkBuild.scala b/project/SparkBuild.scala index bccf36c00f..ac87cffd9f 100644 --- a/project/SparkBuild.scala +++ b/project/SparkBuild.scala @@ -28,6 +28,11 @@ object SparkBuild extends Build { // "2.0.0-mr1-cdh4.2.0" for Cloudera Hadoop. Note that these variables can be set // through the environment variables SPARK_HADOOP_VERSION and SPARK_YARN. val DEFAULT_HADOOP_VERSION = "1.0.4" + + // Whether the Hadoop version to build against is 2.2.x, or a variant of it. This can be set + // through the SPARK_IS_NEW_HADOOP environment variable. + val DEFAULT_IS_NEW_HADOOP = false + val DEFAULT_YARN = false // HBase version; set as appropriate. @@ -55,8 +60,6 @@ object SparkBuild extends Build { lazy val mllib = Project("mllib", file("mllib"), settings = mllibSettings) dependsOn(core) - lazy val yarn = Project("yarn", file("yarn"), settings = yarnSettings) dependsOn(core) - lazy val assemblyProj = Project("assembly", file("assembly"), settings = assemblyProjSettings) .dependsOn(core, bagel, mllib, repl, streaming) dependsOn(maybeYarn: _*) @@ -68,14 +71,27 @@ object SparkBuild extends Build { // Allows build configuration to be set through environment variables lazy val hadoopVersion = scala.util.Properties.envOrElse("SPARK_HADOOP_VERSION", DEFAULT_HADOOP_VERSION) + lazy val isNewHadoop = scala.util.Properties.envOrNone("SPARK_IS_NEW_HADOOP") match { + case None => { + val isNewHadoopVersion = "2.[2-9]+".r.findFirstIn(hadoopVersion).isDefined + (isNewHadoopVersion|| DEFAULT_IS_NEW_HADOOP) + } + case Some(v) => v.toBoolean + } lazy val isYarnEnabled = scala.util.Properties.envOrNone("SPARK_YARN") match { case None => DEFAULT_YARN case Some(v) => v.toBoolean } + // Build against a protobuf-2.5 compatible Akka if Hadoop 2 is used. + lazy val protobufVersion = if (isNewHadoop) "2.5.0" else "2.4.1" + lazy val akkaVersion = if (isNewHadoop) "2.0.5-protobuf-2.5-java-1.5" else "2.0.5" + lazy val akkaGroup = if (isNewHadoop) "org.spark-project" else "com.typesafe.akka" + // Conditionally include the yarn sub-project - lazy val maybeYarn = if(isYarnEnabled) Seq[ClasspathDependency](yarn) else Seq[ClasspathDependency]() - lazy val maybeYarnRef = if(isYarnEnabled) Seq[ProjectReference](yarn) else Seq[ProjectReference]() + lazy val yarn = Project("yarn", file(if (isNewHadoop) "new-yarn" else "yarn"), settings = yarnSettings) dependsOn(core) + lazy val maybeYarn = if (isYarnEnabled) Seq[ClasspathDependency](yarn) else Seq[ClasspathDependency]() + lazy val maybeYarnRef = if (isYarnEnabled) Seq[ProjectReference](yarn) else Seq[ProjectReference]() // Everything except assembly, tools and examples belong to packageProjects lazy val packageProjects = Seq[ProjectReference](core, repl, bagel, streaming, mllib) ++ maybeYarnRef @@ -211,10 +227,10 @@ object SparkBuild extends Build { "com.ning" % "compress-lzf" % "0.8.4", "org.xerial.snappy" % "snappy-java" % "1.0.5", "org.ow2.asm" % "asm" % "4.0", - "com.google.protobuf" % "protobuf-java" % "2.4.1", - "com.typesafe.akka" % "akka-actor" % "2.0.5" excludeAll(excludeNetty), - "com.typesafe.akka" % "akka-remote" % "2.0.5" excludeAll(excludeNetty), - "com.typesafe.akka" % "akka-slf4j" % "2.0.5" excludeAll(excludeNetty), + "com.google.protobuf" % "protobuf-java" % protobufVersion, + akkaGroup % "akka-actor" % akkaVersion excludeAll(excludeNetty), + akkaGroup % "akka-remote" % akkaVersion excludeAll(excludeNetty), + akkaGroup % "akka-slf4j" % akkaVersion excludeAll(excludeNetty), "it.unimi.dsi" % "fastutil" % "6.4.4", "colt" % "colt" % "1.2.0", "net.liftweb" % "lift-json_2.9.2" % "2.5", @@ -230,6 +246,7 @@ object SparkBuild extends Build { "com.codahale.metrics" % "metrics-jvm" % "3.0.0", "com.codahale.metrics" % "metrics-json" % "3.0.0", "com.codahale.metrics" % "metrics-ganglia" % "3.0.0", + "com.codahale.metrics" % "metrics-graphite" % "3.0.0", "com.twitter" % "chill_2.9.3" % "0.3.1", "com.twitter" % "chill-java" % "0.3.1" ) @@ -289,7 +306,7 @@ object SparkBuild extends Build { "org.eclipse.paho" % "mqtt-client" % "0.4.0", "org.apache.flume" % "flume-ng-sdk" % "1.2.0" % "compile" excludeAll(excludeNetty, excludeSnappy), "org.twitter4j" % "twitter4j-stream" % "3.0.3" excludeAll(excludeNetty), - "com.typesafe.akka" % "akka-zeromq" % "2.0.5" excludeAll(excludeNetty), + akkaGroup % "akka-zeromq" % akkaVersion excludeAll(excludeNetty), "org.apache.kafka" % "kafka_2.9.2" % "0.8.0-beta1" exclude("com.sun.jdmk", "jmxtools") exclude("com.sun.jmx", "jmxri") diff --git a/python/epydoc.conf b/python/epydoc.conf index 1d0d002d36..0b42e729f8 100644 --- a/python/epydoc.conf +++ b/python/epydoc.conf @@ -32,6 +32,6 @@ target: docs/ private: no -exclude: pyspark.cloudpickle pyspark.worker pyspark.join pyspark.serializers +exclude: pyspark.cloudpickle pyspark.worker pyspark.join pyspark.java_gateway pyspark.examples pyspark.shell pyspark.test pyspark.rddsampler pyspark.daemon diff --git a/python/pyspark/accumulators.py b/python/pyspark/accumulators.py index da3d96689a..2204e9c9ca 100644 --- a/python/pyspark/accumulators.py +++ b/python/pyspark/accumulators.py @@ -90,9 +90,11 @@ import struct import SocketServer import threading from pyspark.cloudpickle import CloudPickler -from pyspark.serializers import read_int, read_with_length, load_pickle +from pyspark.serializers import read_int, PickleSerializer +pickleSer = PickleSerializer() + # Holds accumulators registered on the current machine, keyed by ID. This is then used to send # the local accumulator updates back to the driver program at the end of a task. _accumulatorRegistry = {} @@ -211,7 +213,7 @@ class _UpdateRequestHandler(SocketServer.StreamRequestHandler): from pyspark.accumulators import _accumulatorRegistry num_updates = read_int(self.rfile) for _ in range(num_updates): - (aid, update) = load_pickle(read_with_length(self.rfile)) + (aid, update) = pickleSer._read_with_length(self.rfile) _accumulatorRegistry[aid] += update # Write a byte in acknowledgement self.wfile.write(struct.pack("!b", 1)) diff --git a/python/pyspark/context.py b/python/pyspark/context.py index a7ca8bc888..cbd41e58c4 100644 --- a/python/pyspark/context.py +++ b/python/pyspark/context.py @@ -26,7 +26,7 @@ from pyspark.accumulators import Accumulator from pyspark.broadcast import Broadcast from pyspark.files import SparkFiles from pyspark.java_gateway import launch_gateway -from pyspark.serializers import dump_pickle, write_with_length, batched +from pyspark.serializers import PickleSerializer, BatchedSerializer, MUTF8Deserializer from pyspark.storagelevel import StorageLevel from pyspark.rdd import RDD @@ -42,7 +42,7 @@ class SparkContext(object): _gateway = None _jvm = None - _writeIteratorToPickleFile = None + _writeToFile = None _takePartition = None _next_accum_id = 0 _active_spark_context = None @@ -51,7 +51,7 @@ class SparkContext(object): def __init__(self, master, jobName, sparkHome=None, pyFiles=None, - environment=None, batchSize=1024): + environment=None, batchSize=1024, serializer=PickleSerializer()): """ Create a new SparkContext. @@ -67,6 +67,7 @@ class SparkContext(object): @param batchSize: The number of Python objects represented as a single Java object. Set 1 to disable batching or -1 to use an unlimited batch size. + @param serializer: The serializer for RDDs. >>> from pyspark.context import SparkContext @@ -83,7 +84,13 @@ class SparkContext(object): self.jobName = jobName self.sparkHome = sparkHome or None # None becomes null in Py4J self.environment = environment or {} - self.batchSize = batchSize # -1 represents a unlimited batch size + self._batchSize = batchSize # -1 represents an unlimited batch size + self._unbatched_serializer = serializer + if batchSize == 1: + self.serializer = self._unbatched_serializer + else: + self.serializer = BatchedSerializer(self._unbatched_serializer, + batchSize) # Create the Java SparkContext through Py4J empty_string_array = self._gateway.new_array(self._jvm.String, 0) @@ -125,8 +132,8 @@ class SparkContext(object): if not SparkContext._gateway: SparkContext._gateway = launch_gateway() SparkContext._jvm = SparkContext._gateway.jvm - SparkContext._writeIteratorToPickleFile = \ - SparkContext._jvm.PythonRDD.writeIteratorToPickleFile + SparkContext._writeToFile = \ + SparkContext._jvm.PythonRDD.writeToFile SparkContext._takePartition = \ SparkContext._jvm.PythonRDD.takePartition @@ -184,15 +191,17 @@ class SparkContext(object): # Make sure we distribute data evenly if it's smaller than self.batchSize if "__len__" not in dir(c): c = list(c) # Make it a list so we can compute its length - batchSize = min(len(c) // numSlices, self.batchSize) + batchSize = min(len(c) // numSlices, self._batchSize) if batchSize > 1: - c = batched(c, batchSize) - for x in c: - write_with_length(dump_pickle(x), tempFile) + serializer = BatchedSerializer(self._unbatched_serializer, + batchSize) + else: + serializer = self._unbatched_serializer + serializer.dump_stream(c, tempFile) tempFile.close() - readRDDFromPickleFile = self._jvm.PythonRDD.readRDDFromPickleFile - jrdd = readRDDFromPickleFile(self._jsc, tempFile.name, numSlices) - return RDD(jrdd, self) + readRDDFromFile = self._jvm.PythonRDD.readRDDFromFile + jrdd = readRDDFromFile(self._jsc, tempFile.name, numSlices) + return RDD(jrdd, self, serializer) def textFile(self, name, minSplits=None): """ @@ -201,21 +210,39 @@ class SparkContext(object): RDD of Strings. """ minSplits = minSplits or min(self.defaultParallelism, 2) - jrdd = self._jsc.textFile(name, minSplits) - return RDD(jrdd, self) + return RDD(self._jsc.textFile(name, minSplits), self, + MUTF8Deserializer()) - def _checkpointFile(self, name): + def _checkpointFile(self, name, input_deserializer): jrdd = self._jsc.checkpointFile(name) - return RDD(jrdd, self) + return RDD(jrdd, self, input_deserializer) def union(self, rdds): """ Build the union of a list of RDDs. + + This supports unions() of RDDs with different serialized formats, + although this forces them to be reserialized using the default + serializer: + + >>> path = os.path.join(tempdir, "union-text.txt") + >>> with open(path, "w") as testFile: + ... testFile.write("Hello") + >>> textFile = sc.textFile(path) + >>> textFile.collect() + [u'Hello'] + >>> parallelized = sc.parallelize(["World!"]) + >>> sorted(sc.union([textFile, parallelized]).collect()) + [u'Hello', 'World!'] """ + first_jrdd_deserializer = rdds[0]._jrdd_deserializer + if any(x._jrdd_deserializer != first_jrdd_deserializer for x in rdds): + rdds = [x._reserialize() for x in rdds] first = rdds[0]._jrdd rest = [x._jrdd for x in rdds[1:]] - rest = ListConverter().convert(rest, self.gateway._gateway_client) - return RDD(self._jsc.union(first, rest), self) + rest = ListConverter().convert(rest, self._gateway._gateway_client) + return RDD(self._jsc.union(first, rest), self, + rdds[0]._jrdd_deserializer) def broadcast(self, value): """ @@ -223,7 +250,9 @@ class SparkContext(object): object for reading it in distributed functions. The variable will be sent to each cluster only once. """ - jbroadcast = self._jsc.broadcast(bytearray(dump_pickle(value))) + pickleSer = PickleSerializer() + pickled = pickleSer.dumps(value) + jbroadcast = self._jsc.broadcast(bytearray(pickled)) return Broadcast(jbroadcast.id(), value, jbroadcast, self._pickled_broadcast_vars) @@ -235,7 +264,7 @@ class SparkContext(object): and floating-point numbers if you do not provide one. For other types, a custom AccumulatorParam can be used. """ - if accum_param == None: + if accum_param is None: if isinstance(value, int): accum_param = accumulators.INT_ACCUMULATOR_PARAM elif isinstance(value, float): diff --git a/python/pyspark/rdd.py b/python/pyspark/rdd.py index 7019fb8bee..d8da02072c 100644 --- a/python/pyspark/rdd.py +++ b/python/pyspark/rdd.py @@ -18,7 +18,7 @@ from base64 import standard_b64encode as b64enc import copy from collections import defaultdict -from itertools import chain, ifilter, imap, product +from itertools import chain, ifilter, imap import operator import os import sys @@ -27,9 +27,8 @@ from subprocess import Popen, PIPE from tempfile import NamedTemporaryFile from threading import Thread -from pyspark import cloudpickle -from pyspark.serializers import batched, Batch, dump_pickle, load_pickle, \ - read_from_pickle_file, pack_long +from pyspark.serializers import NoOpSerializer, CartesianDeserializer, \ + BatchedSerializer, CloudPickleSerializer, pack_long from pyspark.join import python_join, python_left_outer_join, \ python_right_outer_join, python_cogroup from pyspark.statcounter import StatCounter @@ -48,12 +47,12 @@ class RDD(object): operated on in parallel. """ - def __init__(self, jrdd, ctx): + def __init__(self, jrdd, ctx, jrdd_deserializer): self._jrdd = jrdd self.is_cached = False self.is_checkpointed = False self.ctx = ctx - self._partitionFunc = None + self._jrdd_deserializer = jrdd_deserializer @property def context(self): @@ -247,7 +246,23 @@ class RDD(object): >>> rdd.union(rdd).collect() [1, 1, 2, 3, 1, 1, 2, 3] """ - return RDD(self._jrdd.union(other._jrdd), self.ctx) + if self._jrdd_deserializer == other._jrdd_deserializer: + rdd = RDD(self._jrdd.union(other._jrdd), self.ctx, + self._jrdd_deserializer) + return rdd + else: + # These RDDs contain data in different serialized formats, so we + # must normalize them to the default serializer. + self_copy = self._reserialize() + other_copy = other._reserialize() + return RDD(self_copy._jrdd.union(other_copy._jrdd), self.ctx, + self.ctx.serializer) + + def _reserialize(self): + if self._jrdd_deserializer == self.ctx.serializer: + return self + else: + return self.map(lambda x: x, preservesPartitioning=True) def __add__(self, other): """ @@ -334,17 +349,9 @@ class RDD(object): [(1, 1), (1, 2), (2, 1), (2, 2)] """ # Due to batching, we can't use the Java cartesian method. - java_cartesian = RDD(self._jrdd.cartesian(other._jrdd), self.ctx) - def unpack_batches(pair): - (x, y) = pair - if type(x) == Batch or type(y) == Batch: - xs = x.items if type(x) == Batch else [x] - ys = y.items if type(y) == Batch else [y] - for pair in product(xs, ys): - yield pair - else: - yield pair - return java_cartesian.flatMap(unpack_batches) + deserializer = CartesianDeserializer(self._jrdd_deserializer, + other._jrdd_deserializer) + return RDD(self._jrdd.cartesian(other._jrdd), self.ctx, deserializer) def groupBy(self, f, numPartitions=None): """ @@ -391,8 +398,8 @@ class RDD(object): """ Return a list that contains all of the elements in this RDD. """ - picklesInJava = self._jrdd.collect().iterator() - return list(self._collect_iterator_through_file(picklesInJava)) + bytesInJava = self._jrdd.collect().iterator() + return list(self._collect_iterator_through_file(bytesInJava)) def _collect_iterator_through_file(self, iterator): # Transferring lots of data through Py4J can be slow because @@ -400,10 +407,10 @@ class RDD(object): # file and read it back. tempFile = NamedTemporaryFile(delete=False, dir=self.ctx._temp_dir) tempFile.close() - self.ctx._writeIteratorToPickleFile(iterator, tempFile.name) + self.ctx._writeToFile(iterator, tempFile.name) # Read the data into Python and deserialize it: with open(tempFile.name, 'rb') as tempFile: - for item in read_from_pickle_file(tempFile): + for item in self._jrdd_deserializer.load_stream(tempFile): yield item os.unlink(tempFile.name) @@ -571,7 +578,7 @@ class RDD(object): items = [] for partition in range(mapped._jrdd.splits().size()): iterator = self.ctx._takePartition(mapped._jrdd.rdd(), partition) - items.extend(self._collect_iterator_through_file(iterator)) + items.extend(mapped._collect_iterator_through_file(iterator)) if len(items) >= num: break return items[:num] @@ -598,7 +605,10 @@ class RDD(object): '0\\n1\\n2\\n3\\n4\\n5\\n6\\n7\\n8\\n9\\n' """ def func(split, iterator): - return (str(x).encode("utf-8") for x in iterator) + for x in iterator: + if not isinstance(x, basestring): + x = unicode(x) + yield x.encode("utf-8") keyed = PipelinedRDD(self, func) keyed._bypass_serializer = True keyed._jrdd.map(self.ctx._jvm.BytesToString()).saveAsTextFile(path) @@ -735,6 +745,7 @@ class RDD(object): # Transferring O(n) objects to Java is too expensive. Instead, we'll # form the hash buckets in Python, transferring O(numPartitions) objects # to Java. Each object is a (splitNumber, [objects]) pair. + outputSerializer = self.ctx._unbatched_serializer def add_shuffle_key(split, iterator): buckets = defaultdict(list) @@ -743,14 +754,14 @@ class RDD(object): buckets[partitionFunc(k) % numPartitions].append((k, v)) for (split, items) in buckets.iteritems(): yield pack_long(split) - yield dump_pickle(Batch(items)) + yield outputSerializer.dumps(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(numPartitions, id(partitionFunc)) jrdd = pairRDD.partitionBy(partitioner).values() - rdd = RDD(jrdd, self.ctx) + rdd = RDD(jrdd, self.ctx, BatchedSerializer(outputSerializer)) # This is required so that id(partitionFunc) remains unique, even if # partitionFunc is a lambda: rdd._partitionFunc = partitionFunc @@ -787,7 +798,8 @@ class RDD(object): numPartitions = self.ctx.defaultParallelism def combineLocally(iterator): combiners = {} - for (k, v) in iterator: + for x in iterator: + (k, v) = x if k not in combiners: combiners[k] = createCombiner(v) else: @@ -929,38 +941,39 @@ class PipelinedRDD(RDD): 20 """ def __init__(self, prev, func, preservesPartitioning=False): - if isinstance(prev, PipelinedRDD) and prev._is_pipelinable(): + if not isinstance(prev, PipelinedRDD) or not prev._is_pipelinable(): + # This transformation is the first in its stage: + self.func = func + self.preservesPartitioning = preservesPartitioning + self._prev_jrdd = prev._jrdd + self._prev_jrdd_deserializer = prev._jrdd_deserializer + else: prev_func = prev.func def pipeline_func(split, iterator): return func(split, prev_func(split, iterator)) self.func = pipeline_func self.preservesPartitioning = \ prev.preservesPartitioning and preservesPartitioning - self._prev_jrdd = prev._prev_jrdd - else: - self.func = func - self.preservesPartitioning = preservesPartitioning - self._prev_jrdd = prev._jrdd + self._prev_jrdd = prev._prev_jrdd # maintain the pipeline + self._prev_jrdd_deserializer = prev._prev_jrdd_deserializer self.is_cached = False self.is_checkpointed = False self.ctx = prev.ctx self.prev = prev self._jrdd_val = None + self._jrdd_deserializer = self.ctx.serializer self._bypass_serializer = False @property def _jrdd(self): if self._jrdd_val: return self._jrdd_val - func = self.func - if not self._bypass_serializer and self.ctx.batchSize != 1: - oldfunc = self.func - batchSize = self.ctx.batchSize - def batched_func(split, iterator): - return batched(oldfunc(split, iterator), batchSize) - func = batched_func - cmds = [func, self._bypass_serializer] - pipe_command = ' '.join(b64enc(cloudpickle.dumps(f)) for f in cmds) + if self._bypass_serializer: + serializer = NoOpSerializer() + else: + serializer = self.ctx.serializer + command = (self.func, self._prev_jrdd_deserializer, serializer) + pickled_command = CloudPickleSerializer().dumps(command) broadcast_vars = ListConverter().convert( [x._jbroadcast for x in self.ctx._pickled_broadcast_vars], self.ctx._gateway._gateway_client) @@ -971,8 +984,9 @@ class PipelinedRDD(RDD): includes = ListConverter().convert(self.ctx._python_includes, self.ctx._gateway._gateway_client) python_rdd = self.ctx._jvm.PythonRDD(self._prev_jrdd.rdd(), - pipe_command, env, includes, self.preservesPartitioning, self.ctx.pythonExec, - broadcast_vars, self.ctx._javaAccumulator, class_manifest) + bytearray(pickled_command), env, includes, self.preservesPartitioning, + self.ctx.pythonExec, broadcast_vars, self.ctx._javaAccumulator, + class_manifest) self._jrdd_val = python_rdd.asJavaRDD() return self._jrdd_val diff --git a/python/pyspark/serializers.py b/python/pyspark/serializers.py index 54fed1c9c7..811fa6f018 100644 --- a/python/pyspark/serializers.py +++ b/python/pyspark/serializers.py @@ -15,45 +15,269 @@ # limitations under the License. # -import struct +""" +PySpark supports custom serializers for transferring data; this can improve +performance. + +By default, PySpark uses L{PickleSerializer} to serialize objects using Python's +C{cPickle} serializer, which can serialize nearly any Python object. +Other serializers, like L{MarshalSerializer}, support fewer datatypes but can be +faster. + +The serializer is chosen when creating L{SparkContext}: + +>>> from pyspark.context import SparkContext +>>> from pyspark.serializers import MarshalSerializer +>>> sc = SparkContext('local', 'test', serializer=MarshalSerializer()) +>>> sc.parallelize(list(range(1000))).map(lambda x: 2 * x).take(10) +[0, 2, 4, 6, 8, 10, 12, 14, 16, 18] +>>> sc.stop() + +By default, PySpark serialize objects in batches; the batch size can be +controlled through SparkContext's C{batchSize} parameter +(the default size is 1024 objects): + +>>> sc = SparkContext('local', 'test', batchSize=2) +>>> rdd = sc.parallelize(range(16), 4).map(lambda x: x) + +Behind the scenes, this creates a JavaRDD with four partitions, each of +which contains two batches of two objects: + +>>> rdd.glom().collect() +[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15]] +>>> rdd._jrdd.count() +8L +>>> sc.stop() + +A batch size of -1 uses an unlimited batch size, and a size of 1 disables +batching: + +>>> sc = SparkContext('local', 'test', batchSize=1) +>>> rdd = sc.parallelize(range(16), 4).map(lambda x: x) +>>> rdd.glom().collect() +[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15]] +>>> rdd._jrdd.count() +16L +""" + import cPickle +from itertools import chain, izip, product +import marshal +import struct +from pyspark import cloudpickle + + +__all__ = ["PickleSerializer", "MarshalSerializer"] + + +class SpecialLengths(object): + END_OF_DATA_SECTION = -1 + PYTHON_EXCEPTION_THROWN = -2 + TIMING_DATA = -3 + + +class Serializer(object): + + def dump_stream(self, iterator, stream): + """ + Serialize an iterator of objects to the output stream. + """ + raise NotImplementedError + + def load_stream(self, stream): + """ + Return an iterator of deserialized objects from the input stream. + """ + raise NotImplementedError + + + def _load_stream_without_unbatching(self, stream): + return self.load_stream(stream) + + # Note: our notion of "equality" is that output generated by + # equal serializers can be deserialized using the same serializer. + + # This default implementation handles the simple cases; + # subclasses should override __eq__ as appropriate. + + def __eq__(self, other): + return isinstance(other, self.__class__) + + def __ne__(self, other): + return not self.__eq__(other) + + +class FramedSerializer(Serializer): + """ + Serializer that writes objects as a stream of (length, data) pairs, + where C{length} is a 32-bit integer and data is C{length} bytes. + """ + + def dump_stream(self, iterator, stream): + for obj in iterator: + self._write_with_length(obj, stream) + + def load_stream(self, stream): + while True: + try: + yield self._read_with_length(stream) + except EOFError: + return + + def _write_with_length(self, obj, stream): + serialized = self.dumps(obj) + write_int(len(serialized), stream) + stream.write(serialized) + + def _read_with_length(self, stream): + length = read_int(stream) + obj = stream.read(length) + if obj == "": + raise EOFError + return self.loads(obj) + + def dumps(self, obj): + """ + Serialize an object into a byte array. + When batching is used, this will be called with an array of objects. + """ + raise NotImplementedError + + def loads(self, obj): + """ + Deserialize an object from a byte array. + """ + raise NotImplementedError + + +class BatchedSerializer(Serializer): + """ + Serializes a stream of objects in batches by calling its wrapped + Serializer with streams of objects. + """ + + UNLIMITED_BATCH_SIZE = -1 + + def __init__(self, serializer, batchSize=UNLIMITED_BATCH_SIZE): + self.serializer = serializer + self.batchSize = batchSize + + def _batched(self, iterator): + if self.batchSize == self.UNLIMITED_BATCH_SIZE: + yield list(iterator) + else: + items = [] + count = 0 + for item in iterator: + items.append(item) + count += 1 + if count == self.batchSize: + yield items + items = [] + count = 0 + if items: + yield items + + def dump_stream(self, iterator, stream): + self.serializer.dump_stream(self._batched(iterator), stream) + + def load_stream(self, stream): + return chain.from_iterable(self._load_stream_without_unbatching(stream)) + + def _load_stream_without_unbatching(self, stream): + return self.serializer.load_stream(stream) + + def __eq__(self, other): + return isinstance(other, BatchedSerializer) and \ + other.serializer == self.serializer + + def __str__(self): + return "BatchedSerializer<%s>" % str(self.serializer) -class Batch(object): +class CartesianDeserializer(FramedSerializer): """ - Used to store multiple RDD entries as a single Java object. + Deserializes the JavaRDD cartesian() of two PythonRDDs. + """ + + def __init__(self, key_ser, val_ser): + self.key_ser = key_ser + self.val_ser = val_ser + + def load_stream(self, stream): + key_stream = self.key_ser._load_stream_without_unbatching(stream) + val_stream = self.val_ser._load_stream_without_unbatching(stream) + key_is_batched = isinstance(self.key_ser, BatchedSerializer) + val_is_batched = isinstance(self.val_ser, BatchedSerializer) + for (keys, vals) in izip(key_stream, val_stream): + keys = keys if key_is_batched else [keys] + vals = vals if val_is_batched else [vals] + for pair in product(keys, vals): + yield pair + + def __eq__(self, other): + return isinstance(other, CartesianDeserializer) and \ + self.key_ser == other.key_ser and self.val_ser == other.val_ser + + def __str__(self): + return "CartesianDeserializer<%s, %s>" % \ + (str(self.key_ser), str(self.val_ser)) - This relieves us from having to explicitly track whether an RDD - is stored as batches of objects and avoids problems when processing - the union() of batched and unbatched RDDs (e.g. the union() of textFile() - with another RDD). + +class NoOpSerializer(FramedSerializer): + + def loads(self, obj): return obj + def dumps(self, obj): return obj + + +class PickleSerializer(FramedSerializer): """ - def __init__(self, items): - self.items = items + Serializes objects using Python's cPickle serializer: + http://docs.python.org/2/library/pickle.html -def batched(iterator, batchSize): - if batchSize == -1: # unlimited batch size - yield Batch(list(iterator)) - else: - items = [] - count = 0 - for item in iterator: - items.append(item) - count += 1 - if count == batchSize: - yield Batch(items) - items = [] - count = 0 - if items: - yield Batch(items) + This serializer supports nearly any Python object, but may + not be as fast as more specialized serializers. + """ + def dumps(self, obj): return cPickle.dumps(obj, 2) + loads = cPickle.loads -def dump_pickle(obj): - return cPickle.dumps(obj, 2) +class CloudPickleSerializer(PickleSerializer): + def dumps(self, obj): return cloudpickle.dumps(obj, 2) -load_pickle = cPickle.loads + +class MarshalSerializer(FramedSerializer): + """ + Serializes objects using Python's Marshal serializer: + + http://docs.python.org/2/library/marshal.html + + This serializer is faster than PickleSerializer but supports fewer datatypes. + """ + + dumps = marshal.dumps + loads = marshal.loads + + +class MUTF8Deserializer(Serializer): + """ + Deserializes streams written by Java's DataOutputStream.writeUTF(). + """ + + def loads(self, stream): + length = struct.unpack('>H', stream.read(2))[0] + return stream.read(length).decode('utf8') + + def load_stream(self, stream): + while True: + try: + yield self.loads(stream) + except struct.error: + return + except EOFError: + return def read_long(stream): @@ -84,25 +308,4 @@ def write_int(value, stream): def write_with_length(obj, stream): write_int(len(obj), stream) - stream.write(obj) - - -def read_with_length(stream): - length = read_int(stream) - obj = stream.read(length) - if obj == "": - raise EOFError - return obj - - -def read_from_pickle_file(stream): - try: - while True: - obj = load_pickle(read_with_length(stream)) - if type(obj) == Batch: # We don't care about inheritance - for item in obj.items: - yield item - else: - yield obj - except EOFError: - return + stream.write(obj)
\ No newline at end of file diff --git a/python/pyspark/tests.py b/python/pyspark/tests.py index 29d6a128f6..3987642bf4 100644 --- a/python/pyspark/tests.py +++ b/python/pyspark/tests.py @@ -19,6 +19,8 @@ Unit tests for PySpark; additional tests are implemented as doctests in individual modules. """ +from fileinput import input +from glob import glob import os import shutil import sys @@ -86,7 +88,8 @@ class TestCheckpoint(PySparkTestCase): time.sleep(1) # 1 second self.assertTrue(flatMappedRDD.getCheckpointFile() is not None) - recovered = self.sc._checkpointFile(flatMappedRDD.getCheckpointFile()) + recovered = self.sc._checkpointFile(flatMappedRDD.getCheckpointFile(), + flatMappedRDD._jrdd_deserializer) self.assertEquals([1, 2, 3, 4], recovered.collect()) @@ -137,6 +140,19 @@ class TestAddFile(PySparkTestCase): self.assertEqual("Hello World from inside a package!", UserClass().hello()) +class TestRDDFunctions(PySparkTestCase): + + def test_save_as_textfile_with_unicode(self): + # Regression test for SPARK-970 + x = u"\u00A1Hola, mundo!" + data = self.sc.parallelize([x]) + tempFile = NamedTemporaryFile(delete=True) + tempFile.close() + data.saveAsTextFile(tempFile.name) + raw_contents = ''.join(input(glob(tempFile.name + "/part-0000*"))) + self.assertEqual(x, unicode(raw_contents.strip(), "utf-8")) + + class TestIO(PySparkTestCase): def test_stdout_redirection(self): diff --git a/python/pyspark/worker.py b/python/pyspark/worker.py index d63c2aaef7..f2b3f3c142 100644 --- a/python/pyspark/worker.py +++ b/python/pyspark/worker.py @@ -23,23 +23,22 @@ import sys import time import socket import traceback -from base64 import standard_b64decode # CloudPickler needs to be imported so that depicklers are registered using the # copy_reg module. from pyspark.accumulators import _accumulatorRegistry 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, write_long, read_int, dump_pickle, load_pickle, read_from_pickle_file +from pyspark.serializers import write_with_length, write_int, read_long, \ + write_long, read_int, SpecialLengths, MUTF8Deserializer, PickleSerializer -def load_obj(infile): - return load_pickle(standard_b64decode(infile.readline().strip())) +pickleSer = PickleSerializer() +mutf8_deserializer = MUTF8Deserializer() def report_times(outfile, boot, init, finish): - write_int(-3, outfile) + write_int(SpecialLengths.TIMING_DATA, outfile) write_long(1000 * boot, outfile) write_long(1000 * init, outfile) write_long(1000 * finish, outfile) @@ -52,7 +51,7 @@ def main(infile, outfile): return # fetch name of workdir - spark_files_dir = load_pickle(read_with_length(infile)) + spark_files_dir = mutf8_deserializer.loads(infile) SparkFiles._root_directory = spark_files_dir SparkFiles._is_running_on_worker = True @@ -60,38 +59,33 @@ def main(infile, outfile): num_broadcast_variables = read_int(infile) for _ in range(num_broadcast_variables): bid = read_long(infile) - value = read_with_length(infile) - _broadcastRegistry[bid] = Broadcast(bid, load_pickle(value)) + value = pickleSer._read_with_length(infile) + _broadcastRegistry[bid] = Broadcast(bid, value) # fetch names of includes (*.zip and *.egg files) and construct PYTHONPATH sys.path.append(spark_files_dir) # *.py files that were added will be copied here num_python_includes = read_int(infile) for _ in range(num_python_includes): - sys.path.append(os.path.join(spark_files_dir, load_pickle(read_with_length(infile)))) + filename = mutf8_deserializer.loads(infile) + sys.path.append(os.path.join(spark_files_dir, filename)) - # now load function - func = load_obj(infile) - bypassSerializer = load_obj(infile) - if bypassSerializer: - dumps = lambda x: x - else: - dumps = dump_pickle + command = pickleSer._read_with_length(infile) + (func, deserializer, serializer) = command init_time = time.time() - iterator = read_from_pickle_file(infile) try: - for obj in func(split_index, iterator): - write_with_length(dumps(obj), outfile) + iterator = deserializer.load_stream(infile) + serializer.dump_stream(func(split_index, iterator), outfile) except Exception as e: - write_int(-2, outfile) + write_int(SpecialLengths.PYTHON_EXCEPTION_THROWN, 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, outfile) - for aid, accum in _accumulatorRegistry.items(): - write_with_length(dump_pickle((aid, accum._value)), outfile) - write_int(-1, outfile) + write_int(SpecialLengths.END_OF_DATA_SECTION, outfile) + write_int(len(_accumulatorRegistry), outfile) + for (aid, accum) in _accumulatorRegistry.items(): + pickleSer._write_with_length((aid, accum._value), outfile) if __name__ == '__main__': diff --git a/python/run-tests b/python/run-tests index cbc554ea9d..d4dad672d2 100755 --- a/python/run-tests +++ b/python/run-tests @@ -37,6 +37,7 @@ run_test "pyspark/rdd.py" run_test "pyspark/context.py" run_test "-m doctest pyspark/broadcast.py" run_test "-m doctest pyspark/accumulators.py" +run_test "-m doctest pyspark/serializers.py" run_test "pyspark/tests.py" if [[ $FAILED != 0 ]]; then diff --git a/streaming/pom.xml b/streaming/pom.xml index 7a9ae6a97b..40892937b8 100644 --- a/streaming/pom.xml +++ b/streaming/pom.xml @@ -110,15 +110,8 @@ <artifactId>scala-library</artifactId> </dependency> <dependency> - <groupId>com.typesafe.akka</groupId> + <groupId>${akka.group}</groupId> <artifactId>akka-zeromq</artifactId> - <version>2.0.3</version> - <exclusions> - <exclusion> - <groupId>org.jboss.netty</groupId> - <artifactId>netty</artifactId> - </exclusion> - </exclusions> </dependency> <dependency> <groupId>org.scalatest</groupId> diff --git a/streaming/src/main/scala/org/apache/spark/streaming/Checkpoint.scala b/streaming/src/main/scala/org/apache/spark/streaming/Checkpoint.scala index bb9febad38..9271914eb5 100644 --- a/streaming/src/main/scala/org/apache/spark/streaming/Checkpoint.scala +++ b/streaming/src/main/scala/org/apache/spark/streaming/Checkpoint.scala @@ -94,7 +94,7 @@ class CheckpointWriter(checkpointDir: String) extends Logging { fs.delete(file, false) fs.rename(writeFile, file) - val finishTime = System.currentTimeMillis(); + val finishTime = System.currentTimeMillis() logInfo("Checkpoint for time " + checkpointTime + " saved to file '" + file + "', took " + bytes.length + " bytes and " + (finishTime - startTime) + " milliseconds") return @@ -124,7 +124,9 @@ class CheckpointWriter(checkpointDir: String) extends Logging { def stop() { synchronized { - if (stopped) return ; + if (stopped) { + return + } stopped = true } executor.shutdown() diff --git a/streaming/src/main/scala/org/apache/spark/streaming/api/java/JavaStreamingContext.scala b/streaming/src/main/scala/org/apache/spark/streaming/api/java/JavaStreamingContext.scala index cf30b541e1..7f9dab0ef9 100644 --- a/streaming/src/main/scala/org/apache/spark/streaming/api/java/JavaStreamingContext.scala +++ b/streaming/src/main/scala/org/apache/spark/streaming/api/java/JavaStreamingContext.scala @@ -17,7 +17,7 @@ package org.apache.spark.streaming.api.java -import java.lang.{Long => JLong, Integer => JInt} +import java.lang.{Integer => JInt} import java.io.InputStream import java.util.{Map => JMap, List => JList} @@ -33,10 +33,9 @@ import twitter4j.auth.Authorization import org.apache.spark.rdd.RDD import org.apache.spark.storage.StorageLevel import org.apache.spark.api.java.function.{Function => JFunction, Function2 => JFunction2} -import org.apache.spark.api.java.{JavaPairRDD, JavaRDDLike, JavaSparkContext, JavaRDD} +import org.apache.spark.api.java.{JavaPairRDD, JavaSparkContext, JavaRDD} import org.apache.spark.streaming._ import org.apache.spark.streaming.dstream._ -import org.apache.spark.streaming.receivers.{ActorReceiver, ReceiverSupervisorStrategy} /** * A StreamingContext is the main entry point for Spark Streaming functionality. Besides the basic @@ -311,7 +310,7 @@ class JavaStreamingContext(val ssc: StreamingContext) { implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[V]] implicit val cmf: ClassManifest[F] = implicitly[ClassManifest[AnyRef]].asInstanceOf[ClassManifest[F]] - ssc.fileStream[K, V, F](directory); + ssc.fileStream[K, V, F](directory) } /** diff --git a/streaming/src/main/scala/org/apache/spark/streaming/dstream/FlumeInputDStream.scala b/streaming/src/main/scala/org/apache/spark/streaming/dstream/FlumeInputDStream.scala index 18de772946..a0189eca04 100644 --- a/streaming/src/main/scala/org/apache/spark/streaming/dstream/FlumeInputDStream.scala +++ b/streaming/src/main/scala/org/apache/spark/streaming/dstream/FlumeInputDStream.scala @@ -137,8 +137,8 @@ class FlumeReceiver( protected override def onStart() { val responder = new SpecificResponder( - classOf[AvroSourceProtocol], new FlumeEventServer(this)); - val server = new NettyServer(responder, new InetSocketAddress(host, port)); + classOf[AvroSourceProtocol], new FlumeEventServer(this)) + val server = new NettyServer(responder, new InetSocketAddress(host, port)) blockGenerator.start() server.start() logInfo("Flume receiver started") diff --git a/streaming/src/test/scala/org/apache/spark/streaming/InputStreamsSuite.scala b/streaming/src/test/scala/org/apache/spark/streaming/InputStreamsSuite.scala index a559db468a..7dc82decef 100644 --- a/streaming/src/test/scala/org/apache/spark/streaming/InputStreamsSuite.scala +++ b/streaming/src/test/scala/org/apache/spark/streaming/InputStreamsSuite.scala @@ -124,9 +124,9 @@ class InputStreamsSuite extends TestSuiteBase with BeforeAndAfter { val clock = ssc.scheduler.clock.asInstanceOf[ManualClock] val input = Seq(1, 2, 3, 4, 5) Thread.sleep(1000) - val transceiver = new NettyTransceiver(new InetSocketAddress("localhost", testPort)); + val transceiver = new NettyTransceiver(new InetSocketAddress("localhost", testPort)) val client = SpecificRequestor.getClient( - classOf[AvroSourceProtocol], transceiver); + classOf[AvroSourceProtocol], transceiver) for (i <- 0 until input.size) { val event = new AvroFlumeEvent diff --git a/streaming/src/test/scala/org/apache/spark/streaming/TestSuiteBase.scala b/streaming/src/test/scala/org/apache/spark/streaming/TestSuiteBase.scala index be140699c2..8c8c359e6e 100644 --- a/streaming/src/test/scala/org/apache/spark/streaming/TestSuiteBase.scala +++ b/streaming/src/test/scala/org/apache/spark/streaming/TestSuiteBase.scala @@ -251,7 +251,7 @@ trait TestSuiteBase extends FunSuite with BeforeAndAfter with Logging { Thread.sleep(500) // Give some time for the forgetting old RDDs to complete } catch { - case e: Exception => e.printStackTrace(); throw e; + case e: Exception => {e.printStackTrace(); throw e} } finally { ssc.stop() } diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala index 89b00415da..240ed8b32a 100644 --- a/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala @@ -21,9 +21,13 @@ import java.io.IOException import java.net.Socket import java.util.concurrent.CopyOnWriteArrayList import java.util.concurrent.atomic.{AtomicInteger, AtomicReference} + +import scala.collection.JavaConversions._ + import org.apache.hadoop.conf.Configuration import org.apache.hadoop.fs.{FileSystem, Path} import org.apache.hadoop.net.NetUtils +import org.apache.hadoop.security.UserGroupInformation import org.apache.hadoop.util.ShutdownHookManager import org.apache.hadoop.yarn.api._ import org.apache.hadoop.yarn.api.records._ @@ -31,48 +35,49 @@ 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 org.apache.spark.{SparkContext, Logging} import org.apache.spark.util.Utils -import scala.collection.JavaConversions._ + 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 var resourceManager: AMRMProtocol = _ + private var appAttemptId: ApplicationAttemptId = _ + private var userThread: Thread = _ private val yarnConf: YarnConfiguration = new YarnConfiguration(conf) private val fs = FileSystem.get(yarnConf) - private var yarnAllocator: YarnAllocationHandler = null - private var isFinished:Boolean = false - private var uiAddress: String = "" + private var yarnAllocator: YarnAllocationHandler = _ + private var isFinished: Boolean = false + private var uiAddress: String = _ private val maxAppAttempts: Int = conf.getInt(YarnConfiguration.RM_AM_MAX_RETRIES, YarnConfiguration.DEFAULT_RM_AM_MAX_RETRIES) private var isLastAMRetry: Boolean = true - // default to numWorkers * 2, with minimum of 3 + // default to numWorkers * 2, with minimum of 3 private val maxNumWorkerFailures = System.getProperty("spark.yarn.max.worker.failures", math.max(args.numWorkers * 2, 3).toString()).toInt def run() { - // setup the directories so things go to yarn approved directories rather - // then user specified and /tmp + // Setup the directories so things go to yarn approved directories rather + // then user specified and /tmp. System.setProperty("spark.local.dir", getLocalDirs()) - // use priority 30 as its higher then HDFS. Its same priority as MapReduce is using + // Use priority 30 as its higher then HDFS. Its same priority as MapReduce is using. ShutdownHookManager.get().addShutdownHook(new AppMasterShutdownHook(this), 30) appAttemptId = getApplicationAttemptId() - isLastAMRetry = appAttemptId.getAttemptId() >= maxAppAttempts; + isLastAMRetry = appAttemptId.getAttemptId() >= maxAppAttempts resourceManager = registerWithResourceManager() // Workaround until hadoop moves to something which has // https://issues.apache.org/jira/browse/HADOOP-8406 - fixed in (2.0.2-alpha but no 0.23 line) - // ignore result + // 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 ? + // Hence args.workerCores = numCore disabled above. Any better option? // Compute number of threads for akka //val minimumMemory = appMasterResponse.getMinimumResourceCapability().getMemory() @@ -98,7 +103,7 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e waitForSparkContextInitialized() - // do this after spark master is up and SparkContext is created so that we can register UI Url + // Do this after spark master is up and SparkContext is created so that we can register UI Url val appMasterResponse: RegisterApplicationMasterResponse = registerApplicationMaster() // Allocate all containers @@ -117,12 +122,12 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e // LOCAL_DIRS => 2.X, YARN_LOCAL_DIRS => 0.23.X val localDirs = Option(System.getenv("YARN_LOCAL_DIRS")) .getOrElse(Option(System.getenv("LOCAL_DIRS")) - .getOrElse("")) + .getOrElse("")) if (localDirs.isEmpty()) { throw new Exception("Yarn Local dirs can't be empty") } - return localDirs + localDirs } private def getApplicationAttemptId(): ApplicationAttemptId = { @@ -131,7 +136,7 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e val containerId = ConverterUtils.toContainerId(containerIdString) val appAttemptId = containerId.getApplicationAttemptId() logInfo("ApplicationAttemptId: " + appAttemptId) - return appAttemptId + appAttemptId } private def registerWithResourceManager(): AMRMProtocol = { @@ -139,7 +144,7 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e YarnConfiguration.RM_SCHEDULER_ADDRESS, YarnConfiguration.DEFAULT_RM_SCHEDULER_ADDRESS)) logInfo("Connecting to ResourceManager at " + rmAddress) - return rpc.getProxy(classOf[AMRMProtocol], rmAddress, conf).asInstanceOf[AMRMProtocol] + rpc.getProxy(classOf[AMRMProtocol], rmAddress, conf).asInstanceOf[AMRMProtocol] } private def registerApplicationMaster(): RegisterApplicationMasterResponse = { @@ -147,12 +152,13 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e 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. + // 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) appMasterRequest.setTrackingUrl(uiAddress) - return resourceManager.registerApplicationMaster(appMasterRequest) + resourceManager.registerApplicationMaster(appMasterRequest) } private def waitForSparkMaster() { @@ -166,21 +172,25 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e try { val socket = new Socket(driverHost, driverPort.toInt) socket.close() - logInfo("Driver now available: " + driverHost + ":" + driverPort) + logInfo("Driver now available: %s:%s".format(driverHost, driverPort)) driverUp = true } catch { - case e: Exception => - logWarning("Failed to connect to driver at " + driverHost + ":" + driverPort + ", retrying") - Thread.sleep(100) - tries = tries + 1 + case e: Exception => { + logWarning("Failed to connect to driver at %s:%s, retrying ...". + format(driverHost, driverPort)) + Thread.sleep(100) + tries = tries + 1 + } } } } 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 mainMethod = Class.forName( + args.userClass, + false /* initialize */, + Thread.currentThread.getContextClassLoader).getMethod("main", classOf[Array[String]]) val t = new Thread { override def run() { var successed = false @@ -195,7 +205,7 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e successed = true } finally { logDebug("finishing main") - isLastAMRetry = true; + isLastAMRetry = true if (successed) { ApplicationMaster.this.finishApplicationMaster(FinalApplicationStatus.SUCCEEDED) } else { @@ -205,7 +215,7 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e } } t.start() - return t + t } // this need to happen before allocateWorkers @@ -227,13 +237,20 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e if (null != sparkContext) { uiAddress = sparkContext.ui.appUIAddress - this.yarnAllocator = YarnAllocationHandler.newAllocator(yarnConf, resourceManager, - appAttemptId, args, sparkContext.preferredNodeLocationData) + this.yarnAllocator = YarnAllocationHandler.newAllocator( + yarnConf, + resourceManager, + appAttemptId, + args, + sparkContext.preferredNodeLocationData) } else { - logWarning("Unable to retrieve sparkContext inspite of waiting for " + count * waitTime + - ", numTries = " + numTries) - this.yarnAllocator = YarnAllocationHandler.newAllocator(yarnConf, resourceManager, - appAttemptId, args) + logWarning("Unable to retrieve sparkContext inspite of waiting for %d, numTries = %d". + format(count * waitTime, numTries)) + this.yarnAllocator = YarnAllocationHandler.newAllocator( + yarnConf, + resourceManager, + appAttemptId, + args) } } } finally { @@ -249,36 +266,39 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e // 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) { - if (yarnAllocator.getNumWorkersFailed >= maxNumWorkerFailures) { - finishApplicationMaster(FinalApplicationStatus.FAILED, - "max number of worker failures reached") - } - yarnAllocator.allocateContainers(math.max(args.numWorkers - yarnAllocator.getNumWorkersRunning, 0)) - ApplicationMaster.incrementAllocatorLoop(1) - Thread.sleep(100) + + // Exists the loop if the user thread exits. + while (yarnAllocator.getNumWorkersRunning < args.numWorkers && userThread.isAlive) { + if (yarnAllocator.getNumWorkersFailed >= maxNumWorkerFailures) { + finishApplicationMaster(FinalApplicationStatus.FAILED, + "max number of worker failures reached") + } + 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 + // In case of exceptions, etc - ensure that count is at least 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 + // Launch a progress reporter thread, else the app will get killed after expiration + // (def: 10mins) timeout. + // TODO(harvey): Verify the timeout if (userThread.isAlive) { - // ensure that progress is sent before YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS elapse. - + // Ensure that progress is sent before YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS elapses. val timeoutInterval = yarnConf.getInt(YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS, 120000) // we want to be reasonably responsive without causing too many requests to RM. - val schedulerInterval = + val schedulerInterval = System.getProperty("spark.yarn.scheduler.heartbeat.interval-ms", "5000").toLong // must be <= timeoutInterval / 2. val interval = math.min(timeoutInterval / 2, schedulerInterval) + launchReporterThread(interval) } } @@ -290,12 +310,13 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e override def run() { while (userThread.isAlive) { if (yarnAllocator.getNumWorkersFailed >= maxNumWorkerFailures) { - finishApplicationMaster(FinalApplicationStatus.FAILED, + finishApplicationMaster(FinalApplicationStatus.FAILED, "max number of worker failures reached") } val missingWorkerCount = args.numWorkers - yarnAllocator.getNumWorkersRunning if (missingWorkerCount > 0) { - logInfo("Allocating " + missingWorkerCount + " containers to make up for (potentially ?) lost containers") + logInfo("Allocating %d containers to make up for (potentially) lost containers". + format(missingWorkerCount)) yarnAllocator.allocateContainers(missingWorkerCount) } else sendProgress() @@ -303,16 +324,16 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e } } } - // setting to daemon status, though this is usually not a good idea. + // 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 + t } private def sendProgress() { logDebug("Sending progress") - // simulated with an allocate request with no nodes requested ... + // Simulated with an allocate request with no nodes requested ... yarnAllocator.allocateContainers(0) } @@ -332,7 +353,6 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e */ def finishApplicationMaster(status: FinalApplicationStatus, diagnostics: String = "") { - synchronized { if (isFinished) { return @@ -346,14 +366,13 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e finishReq.setAppAttemptId(appAttemptId) finishReq.setFinishApplicationStatus(status) finishReq.setDiagnostics(diagnostics) - // set tracking url to empty since we don't have a history server + // Set tracking url to empty since we don't have a history server. finishReq.setTrackingUrl("") resourceManager.finishApplicationMaster(finishReq) - } /** - * clean up the staging directory. + * Clean up the staging directory. */ private def cleanupStagingDir() { var stagingDirPath: Path = null @@ -369,13 +388,12 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e fs.delete(stagingDirPath, true) } } catch { - case e: IOException => - logError("Failed to cleanup staging dir " + stagingDirPath, e) + case ioe: IOException => + logError("Failed to cleanup staging dir " + stagingDirPath, ioe) } } - // The shutdown hook that runs when a signal is received AND during normal - // close of the JVM. + // The shutdown hook that runs when a signal is received AND during normal close of the JVM. class AppMasterShutdownHook(appMaster: ApplicationMaster) extends Runnable { def run() { @@ -385,15 +403,14 @@ class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) e if (appMaster.isLastAMRetry) appMaster.cleanupStagingDir() } } - } object ApplicationMaster { - // number of times to wait for the allocator loop to complete. - // each loop iteration waits for 100ms, so maximum of 3 seconds. + // 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. + // 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) @@ -411,7 +428,8 @@ object ApplicationMaster { applicationMasters.add(master) } - val sparkContextRef: AtomicReference[SparkContext] = new AtomicReference[SparkContext](null) + val sparkContextRef: AtomicReference[SparkContext] = + new AtomicReference[SparkContext](null /* initialValue */) val yarnAllocatorLoop: AtomicInteger = new AtomicInteger(0) def sparkContextInitialized(sc: SparkContext): Boolean = { @@ -421,19 +439,21 @@ object ApplicationMaster { 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. - // Note that this will unfortunately not properly clean up the staging files because it gets called to - // late and the filesystem is already shutdown. + // 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 the Client from declaring failure even though we exited properly. + // Note that this will unfortunately not properly clean up the staging files because it gets + // called too late, after the filesystem is already shutdown. 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' + // This is not only logs, but also ensures that log system is initialized for this instance + // when we are actually 'run'-ing. logInfo("Adding shutdown hook for context " + sc) override def run() { logInfo("Invoking sc stop from shutdown hook") sc.stop() - // best case ... + // Best case ... for (master <- applicationMasters) { master.finishApplicationMaster(FinalApplicationStatus.SUCCEEDED) } @@ -441,7 +461,7 @@ object ApplicationMaster { } ) } - // Wait for initialization to complete and atleast 'some' nodes can get allocated + // Wait for initialization to complete and atleast 'some' nodes can get allocated. yarnAllocatorLoop.synchronized { while (yarnAllocatorLoop.get() <= ALLOCATOR_LOOP_WAIT_COUNT) { yarnAllocatorLoop.wait(1000L) diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala index 1078d5b826..79dd038065 100644 --- a/yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala @@ -17,49 +17,52 @@ package org.apache.spark.deploy.yarn -import java.net.{InetAddress, InetSocketAddress, UnknownHostException, URI} +import java.net.{InetAddress, UnknownHostException, URI} import java.nio.ByteBuffer +import scala.collection.JavaConversions._ +import scala.collection.mutable.HashMap +import scala.collection.mutable.Map + import org.apache.hadoop.conf.Configuration import org.apache.hadoop.fs.{FileContext, FileStatus, FileSystem, Path, FileUtil} -import org.apache.hadoop.fs.permission.FsPermission +import org.apache.hadoop.fs.permission.FsPermission; +import org.apache.hadoop.io.DataOutputBuffer import org.apache.hadoop.mapred.Master import org.apache.hadoop.net.NetUtils -import org.apache.hadoop.io.DataOutputBuffer import org.apache.hadoop.security.UserGroupInformation import org.apache.hadoop.yarn.api._ import org.apache.hadoop.yarn.api.ApplicationConstants.Environment -import org.apache.hadoop.yarn.api.records._ import org.apache.hadoop.yarn.api.protocolrecords._ +import org.apache.hadoop.yarn.api.records._ import org.apache.hadoop.yarn.client.YarnClientImpl import org.apache.hadoop.yarn.conf.YarnConfiguration import org.apache.hadoop.yarn.ipc.YarnRPC import org.apache.hadoop.yarn.util.{Apps, Records} -import scala.collection.mutable.HashMap -import scala.collection.mutable.Map -import scala.collection.JavaConversions._ - import org.apache.spark.Logging import org.apache.spark.util.Utils import org.apache.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) val credentials = UserGroupInformation.getCurrentUser().getCredentials() private val SPARK_STAGING: String = ".sparkStaging" private val distCacheMgr = new ClientDistributedCacheManager() - // staging directory is private! -> rwx-------- + // Staging directory is private! -> rwx-------- val STAGING_DIR_PERMISSION: FsPermission = FsPermission.createImmutable(0700:Short) - // app files are world-wide readable and owner writable -> rw-r--r-- + + // App files are world-wide readable and owner writable -> rw-r--r-- val APP_FILE_PERMISSION: FsPermission = FsPermission.createImmutable(0644:Short) - def run() { + // for client user who want to monitor app status by itself. + def runApp() = { validateArgs() init(yarnConf) @@ -81,21 +84,26 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl appContext.setUser(UserGroupInformation.getCurrentUser().getShortUserName()) submitApp(appContext) - + appId + } + + def run() { + val appId = runApp() monitorApplication(appId) System.exit(0) } def validateArgs() = { - Map((System.getenv("SPARK_JAR") == null) -> "Error: You must set SPARK_JAR environment variable!", + Map( + (System.getenv("SPARK_JAR") == null) -> "Error: You must set SPARK_JAR environment variable!", (args.userJar == null) -> "Error: You must specify a user jar!", (args.userClass == null) -> "Error: You must specify a user class!", (args.numWorkers <= 0) -> "Error: You must specify atleast 1 worker!", - (args.amMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> - ("Error: AM memory size must be greater then: " + YarnAllocationHandler.MEMORY_OVERHEAD), - (args.workerMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> - ("Error: Worker memory size must be greater then: " + YarnAllocationHandler.MEMORY_OVERHEAD.toString())) - .foreach { case(cond, errStr) => + (args.amMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> ("Error: AM memory size must be " + + "greater than: " + YarnAllocationHandler.MEMORY_OVERHEAD), + (args.workerMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> ("Error: Worker memory size " + + "must be greater than: " + YarnAllocationHandler.MEMORY_OVERHEAD) + ).foreach { case(cond, errStr) => if (cond) { logError(errStr) args.printUsageAndExit(1) @@ -109,21 +117,26 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl def logClusterResourceDetails() { val clusterMetrics: YarnClusterMetrics = super.getYarnClusterMetrics - logInfo("Got Cluster metric info from ASM, numNodeManagers=" + clusterMetrics.getNumNodeManagers) + 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) + logInfo("""Queue info ... queueName = %s, queueCurrentCapacity = %s, queueMaxCapacity = %s, + queueApplicationCount = %s, queueChildQueueCount = %s""".format( + queueInfo.getQueueName, + queueInfo.getCurrentCapacity, + queueInfo.getMaximumCapacity, + queueInfo.getApplications.size, + queueInfo.getChildQueues.size)) } - + def verifyClusterResources(app: GetNewApplicationResponse) = { val maxMem = app.getMaximumResourceCapability().getMemory() logInfo("Max mem capabililty of a single resource in this cluster " + maxMem) - - // if we have requested more then the clusters max for a single resource then exit. + + // If we have requested more then the clusters max for a single resource then exit. if (args.workerMemory > maxMem) { - logError("the worker size is to large to run on this cluster " + args.workerMemory); + logError("the worker size is to large to run on this cluster " + args.workerMemory) System.exit(1) } val amMem = args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD @@ -132,10 +145,10 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl System.exit(1) } - // We could add checks to make sure the entire cluster has enough resources but that involves getting - // all the node reports and computing ourselves + // We could add checks to make sure the entire cluster has enough resources but that involves + // getting all the node reports and computing ourselves } - + def createApplicationSubmissionContext(appId: ApplicationId): ApplicationSubmissionContext = { logInfo("Setting up application submission context for ASM") val appContext = Records.newRecord(classOf[ApplicationSubmissionContext]) @@ -144,9 +157,7 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl return appContext } - /* - * see if two file systems are the same or not. - */ + /** See if two file systems are the same or not. */ private def compareFs(srcFs: FileSystem, destFs: FileSystem): Boolean = { val srcUri = srcFs.getUri() val dstUri = destFs.getUri() @@ -160,8 +171,8 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl var dstHost = dstUri.getHost() if ((srcHost != null) && (dstHost != null)) { try { - srcHost = InetAddress.getByName(srcHost).getCanonicalHostName(); - dstHost = InetAddress.getByName(dstHost).getCanonicalHostName(); + srcHost = InetAddress.getByName(srcHost).getCanonicalHostName() + dstHost = InetAddress.getByName(dstHost).getCanonicalHostName() } catch { case e: UnknownHostException => return false @@ -178,30 +189,27 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl if (srcUri.getPort() != dstUri.getPort()) { return false } - return true; + return true } - /** - * Copy the file into HDFS if needed. - */ + /** Copy the file into HDFS if needed. */ private def copyRemoteFile( dstDir: Path, originalPath: Path, replication: Short, setPerms: Boolean = false): Path = { val fs = FileSystem.get(conf) - val remoteFs = originalPath.getFileSystem(conf); + val remoteFs = originalPath.getFileSystem(conf) var newPath = originalPath if (! compareFs(remoteFs, fs)) { newPath = new Path(dstDir, originalPath.getName()) logInfo("Uploading " + originalPath + " to " + newPath) - FileUtil.copy(remoteFs, originalPath, fs, newPath, false, conf); - fs.setReplication(newPath, replication); + FileUtil.copy(remoteFs, originalPath, fs, newPath, false, conf) + fs.setReplication(newPath, replication) if (setPerms) fs.setPermission(newPath, new FsPermission(APP_FILE_PERMISSION)) } - // resolve any symlinks in the URI path so using a "current" symlink - // to point to a specific version shows the specific version - // in the distributed cache configuration + // Resolve any symlinks in the URI path so using a "current" symlink to point to a specific + // version shows the specific version in the distributed cache configuration val qualPath = fs.makeQualified(newPath) val fc = FileContext.getFileContext(qualPath.toUri(), conf) val destPath = fc.resolvePath(qualPath) @@ -210,11 +218,11 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl def prepareLocalResources(appStagingDir: String): HashMap[String, LocalResource] = { logInfo("Preparing Local resources") - // Upload Spark and the application JAR to the remote file system if necessary - // Add them as local resources to the AM + // Upload Spark and the application JAR to the remote file system if necessary. Add them as + // local resources to the AM. val fs = FileSystem.get(conf) - val delegTokenRenewer = Master.getMasterPrincipal(conf); + val delegTokenRenewer = Master.getMasterPrincipal(conf) if (UserGroupInformation.isSecurityEnabled()) { if (delegTokenRenewer == null || delegTokenRenewer.length() == 0) { logError("Can't get Master Kerberos principal for use as renewer") @@ -226,7 +234,7 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl if (UserGroupInformation.isSecurityEnabled()) { val dstFs = dst.getFileSystem(conf) - dstFs.addDelegationTokens(delegTokenRenewer, credentials); + dstFs.addDelegationTokens(delegTokenRenewer, credentials) } val localResources = HashMap[String, LocalResource]() FileSystem.mkdirs(fs, dst, new FsPermission(STAGING_DIR_PERMISSION)) @@ -241,7 +249,7 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl var localURI = new URI(localPath) // if not specified assume these are in the local filesystem to keep behavior like Hadoop if (localURI.getScheme() == null) { - localURI = new URI(FileSystem.getLocal(conf).makeQualified(new Path(localPath)).toString()) + localURI = new URI(FileSystem.getLocal(conf).makeQualified(new Path(localPath)).toString) } val setPermissions = if (destName.equals(Client.APP_JAR)) true else false val destPath = copyRemoteFile(dst, new Path(localURI), replication, setPermissions) @@ -286,10 +294,10 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl } } - UserGroupInformation.getCurrentUser().addCredentials(credentials); + UserGroupInformation.getCurrentUser().addCredentials(credentials) return localResources } - + def setupLaunchEnv( localResources: HashMap[String, LocalResource], stagingDir: String): HashMap[String, String] = { @@ -302,16 +310,16 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl env("SPARK_YARN_MODE") = "true" env("SPARK_YARN_STAGING_DIR") = stagingDir - // set the environment variables to be passed on to the Workers + // Set the environment variables to be passed on to the Workers. distCacheMgr.setDistFilesEnv(env) distCacheMgr.setDistArchivesEnv(env) - // allow users to specify some environment variables + // Allow users to specify some environment variables. Apps.setEnvFromInputString(env, System.getenv("SPARK_YARN_USER_ENV")) - // Add each SPARK-* key to the environment + // Add each SPARK-* key to the environment. System.getenv().filterKeys(_.startsWith("SPARK")).foreach { case (k,v) => env(k) = v } - return env + env } def userArgsToString(clientArgs: ClientArguments): String = { @@ -321,13 +329,13 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl 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 = { + 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) @@ -335,8 +343,10 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl val minResMemory: Int = newApp.getMinimumResourceCapability().getMemory() + // TODO(harvey): This can probably be a val. var amMemory = ((args.amMemory / minResMemory) * minResMemory) + - (if (0 != (args.amMemory % minResMemory)) minResMemory else 0) - YarnAllocationHandler.MEMORY_OVERHEAD + ((if ((args.amMemory % minResMemory) == 0) 0 else minResMemory) - + YarnAllocationHandler.MEMORY_OVERHEAD) // Extra options for the JVM var JAVA_OPTS = "" @@ -347,13 +357,18 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl JAVA_OPTS += " -Djava.io.tmpdir=" + new Path(Environment.PWD.$(), YarnConfiguration.DEFAULT_CONTAINER_TEMP_DIR) + " " - // 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 + // 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. + val useConcurrentAndIncrementalGC = env.isDefinedAt("SPARK_USE_CONC_INCR_GC") && + java.lang.Boolean.parseBoolean(env("SPARK_USE_CONC_INCR_GC")) + if (useConcurrentAndIncrementalGC) { + // 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 " @@ -366,7 +381,7 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl } // Command for the ApplicationMaster - var javaCommand = "java"; + var javaCommand = "java" val javaHome = System.getenv("JAVA_HOME") if ((javaHome != null && !javaHome.isEmpty()) || env.isDefinedAt("JAVA_HOME")) { javaCommand = Environment.JAVA_HOME.$() + "/bin/java" @@ -375,7 +390,7 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl val commands = List[String](javaCommand + " -server " + JAVA_OPTS + - " org.apache.spark.deploy.yarn.ApplicationMaster" + + " " + args.amClass + " --class " + args.userClass + " --jar " + args.userJar + userArgsToString(args) + @@ -386,28 +401,28 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl " 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 + // Memory for the ApplicationMaster. capability.setMemory(args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD) amContainer.setResource(capability) - // Setup security tokens + // Setup security tokens. val dob = new DataOutputBuffer() credentials.writeTokenStorageToStream(dob) amContainer.setContainerTokens(ByteBuffer.wrap(dob.getData())) - return amContainer + amContainer } - + def submitApp(appContext: ApplicationSubmissionContext) = { - // Submit the application to the applications manager + // Submit the application to the applications manager. logInfo("Submitting application to ASM") super.submitApplication(appContext) } - + def monitorApplication(appId: ApplicationId): Boolean = { - while(true) { + while (true) { Thread.sleep(1000) val report = super.getApplicationReport(appId) @@ -425,16 +440,16 @@ class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl "\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 } } - return true + true } } @@ -467,7 +482,7 @@ object Client { Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() + Path.SEPARATOR + LOG4J_PROP) } - // normally the users app.jar is last in case conflicts with spark jars + // Normally the users app.jar is last in case conflicts with spark jars val userClasspathFirst = System.getProperty("spark.yarn.user.classpath.first", "false") .toBoolean if (userClasspathFirst) { diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala index 852dbd7dab..b9dbc3fb87 100644 --- a/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala @@ -35,6 +35,7 @@ class ClientArguments(val args: Array[String]) { var numWorkers = 2 var amQueue = System.getProperty("QUEUE", "default") var amMemory: Int = 512 + var amClass: String = "org.apache.spark.deploy.yarn.ApplicationMaster" var appName: String = "Spark" // TODO var inputFormatInfo: List[InputFormatInfo] = null @@ -62,18 +63,22 @@ class ClientArguments(val args: Array[String]) { userArgsBuffer += value args = tail - case ("--master-memory") :: MemoryParam(value) :: tail => - amMemory = value + case ("--master-class") :: value :: tail => + amClass = value args = tail - case ("--num-workers") :: IntParam(value) :: tail => - numWorkers = value + case ("--master-memory") :: MemoryParam(value) :: tail => + amMemory = value args = tail case ("--worker-memory") :: MemoryParam(value) :: tail => workerMemory = value args = tail + case ("--num-workers") :: IntParam(value) :: tail => + numWorkers = value + args = tail + case ("--worker-cores") :: IntParam(value) :: tail => workerCores = value args = tail @@ -119,19 +124,20 @@ class ClientArguments(val args: Array[String]) { System.err.println( "Usage: org.apache.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" + - " --name NAME The name of your application (Default: Spark)\n" + - " --queue QUEUE The hadoop queue to use for allocation requests (Default: 'default')\n" + - " --addJars jars Comma separated list of local jars that want SparkContext.addJar to work with.\n" + - " --files files Comma separated list of files to be distributed with the job.\n" + - " --archives archives Comma separated list of archives to be distributed with the job." + " --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-class CLASS_NAME Class Name for Master (Default: spark.deploy.yarn.ApplicationMaster)\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" + + " --name NAME The name of your application (Default: Spark)\n" + + " --queue QUEUE The hadoop queue to use for allocation requests (Default: 'default')\n" + + " --addJars jars Comma separated list of local jars that want SparkContext.addJar to work with.\n" + + " --files files Comma separated list of files to be distributed with the job.\n" + + " --archives archives Comma separated list of archives to be distributed with the job." ) System.exit(exitCode) } diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala index 674c8f8112..5f159b073f 100644 --- a/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala @@ -197,11 +197,11 @@ class ClientDistributedCacheManager() extends Logging { */ def checkPermissionOfOther(fs: FileSystem, path: Path, action: FsAction, statCache: Map[URI, FileStatus]): Boolean = { - val status = getFileStatus(fs, path.toUri(), statCache); + val status = getFileStatus(fs, path.toUri(), statCache) val perms = status.getPermission() val otherAction = perms.getOtherAction() if (otherAction.implies(action)) { - return true; + return true } return false } diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala new file mode 100644 index 0000000000..421a83c87a --- /dev/null +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala @@ -0,0 +1,246 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.deploy.yarn + +import java.net.Socket +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 akka.actor._ +import akka.remote.{RemoteClientShutdown, RemoteClientDisconnected, RemoteClientLifeCycleEvent} +import akka.remote.RemoteClientShutdown +import akka.actor.Terminated +import akka.remote.RemoteClientDisconnected +import org.apache.spark.{SparkContext, Logging} +import org.apache.spark.util.{Utils, AkkaUtils} +import org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend +import org.apache.spark.scheduler.SplitInfo + +class WorkerLauncher(args: ApplicationMasterArguments, conf: Configuration) extends Logging { + + def this(args: ApplicationMasterArguments) = this(args, new Configuration()) + + private val rpc: YarnRPC = YarnRPC.create(conf) + private var resourceManager: AMRMProtocol = null + private var appAttemptId: ApplicationAttemptId = null + private var reporterThread: Thread = null + private val yarnConf: YarnConfiguration = new YarnConfiguration(conf) + + private var yarnAllocator: YarnAllocationHandler = null + private var driverClosed:Boolean = false + + val actorSystem : ActorSystem = AkkaUtils.createActorSystem("sparkYarnAM", Utils.localHostName, 0)._1 + var actor: ActorRef = null + + // This actor just working as a monitor to watch on Driver Actor. + class MonitorActor(driverUrl: String) extends Actor { + + var driver: ActorRef = null + + override def preStart() { + logInfo("Listen to driver: " + driverUrl) + driver = context.actorFor(driverUrl) + driver ! "hello" + context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent]) + context.watch(driver) // Doesn't work with remote actors, but useful for testing + } + + override def receive = { + case Terminated(_) | RemoteClientDisconnected(_, _) | RemoteClientShutdown(_, _) => + logInfo("Driver terminated or disconnected! Shutting down.") + driverClosed = true + } + } + + def run() { + + 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 + } + } + + waitForSparkMaster() + + // Allocate all containers + allocateWorkers() + + // Launch a progress reporter thread, else app will get killed after expiration (def: 10mins) timeout + // 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)) + reporterThread = launchReporterThread(interval) + + // Wait for the reporter thread to Finish. + reporterThread.join() + + finishApplicationMaster(FinalApplicationStatus.SUCCEEDED) + actorSystem.shutdown() + + logInfo("Exited") + 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 + val hostport = args.userArgs(0) + val (driverHost, driverPort) = Utils.parseHostPort(hostport) + while(!driverUp) { + try { + val socket = new Socket(driverHost, driverPort) + 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) + } + } + System.setProperty("spark.driver.host", driverHost) + System.setProperty("spark.driver.port", driverPort.toString) + + val driverUrl = "akka://spark@%s:%s/user/%s".format( + driverHost, driverPort.toString, CoarseGrainedSchedulerBackend.ACTOR_NAME) + + actor = actorSystem.actorOf(Props(new MonitorActor(driverUrl)), name = "YarnAM") + } + + + private def allocateWorkers() { + + // Fixme: should get preferredNodeLocationData from SparkContext, just fake a empty one for now. + val preferredNodeLocationData: scala.collection.Map[String, scala.collection.Set[SplitInfo]] = scala.collection.immutable.Map() + + yarnAllocator = YarnAllocationHandler.newAllocator(yarnConf, resourceManager, appAttemptId, args, 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) { + yarnAllocator.allocateContainers(math.max(args.numWorkers - yarnAllocator.getNumWorkersRunning, 0)) + Thread.sleep(100) + } + + logInfo("All workers have launched.") + + } + + // 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 (!driverClosed) { + 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 finishApplicationMaster(status: FinalApplicationStatus) { + + logInfo("finish ApplicationMaster with " + status) + val finishReq = Records.newRecord(classOf[FinishApplicationMasterRequest]) + .asInstanceOf[FinishApplicationMasterRequest] + finishReq.setAppAttemptId(appAttemptId) + finishReq.setFinishApplicationStatus(status) + resourceManager.finishApplicationMaster(finishReq) + } + +} + + +object WorkerLauncher { + def main(argStrings: Array[String]) { + val args = new ApplicationMasterArguments(argStrings) + new WorkerLauncher(args).run() + } +} diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala index 7a66532254..6a90cc51cf 100644 --- a/yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala @@ -21,53 +21,59 @@ import java.net.URI import java.nio.ByteBuffer import java.security.PrivilegedExceptionAction +import scala.collection.JavaConversions._ +import scala.collection.mutable.HashMap + import org.apache.hadoop.conf.Configuration -import org.apache.hadoop.fs.{FileStatus, FileSystem, Path} +import org.apache.hadoop.fs.Path import org.apache.hadoop.io.DataOutputBuffer import org.apache.hadoop.net.NetUtils import org.apache.hadoop.security.UserGroupInformation import org.apache.hadoop.yarn.api._ +import org.apache.hadoop.yarn.api.ApplicationConstants.Environment 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, ProtoUtils} -import org.apache.hadoop.yarn.api.ApplicationConstants.Environment - -import scala.collection.JavaConversions._ -import scala.collection.mutable.HashMap import org.apache.spark.Logging -import org.apache.spark.util.Utils -class WorkerRunnable(container: Container, conf: Configuration, masterAddress: String, - slaveId: String, hostname: String, workerMemory: Int, workerCores: Int) - extends Runnable with Logging { - + +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 @@ -80,17 +86,21 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S JAVA_OPTS += " -Djava.io.tmpdir=" + new Path(Environment.PWD.$(), YarnConfiguration.DEFAULT_CONTAINER_TEMP_DIR) + " " - - // 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. + // 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 + // 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 " @@ -108,7 +118,7 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S credentials.writeTokenStorageToStream(dob) ctx.setContainerTokens(ByteBuffer.wrap(dob.getData())) - var javaCommand = "java"; + var javaCommand = "java" val javaHome = System.getenv("JAVA_HOME") if ((javaHome != null && !javaHome.isEmpty()) || env.isDefinedAt("JAVA_HOME")) { javaCommand = Environment.JAVA_HOME.$() + "/bin/java" @@ -117,8 +127,10 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S val commands = List[String](javaCommand + " -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 ? + // 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 + " org.apache.spark.executor.CoarseGrainedExecutorBackend " + @@ -130,7 +142,7 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S " 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] @@ -138,7 +150,8 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S cm.startContainer(startReq) } - private def setupDistributedCache(file: String, + private def setupDistributedCache( + file: String, rtype: LocalResourceType, localResources: HashMap[String, LocalResource], timestamp: String, @@ -153,12 +166,11 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S amJarRsrc.setSize(size.toLong) localResources(uri.getFragment()) = amJarRsrc } - - + def prepareLocalResources: HashMap[String, LocalResource] = { logInfo("Preparing Local resources") val localResources = HashMap[String, LocalResource]() - + if (System.getenv("SPARK_YARN_CACHE_FILES") != null) { val timeStamps = System.getenv("SPARK_YARN_CACHE_FILES_TIME_STAMPS").split(',') val fileSizes = System.getenv("SPARK_YARN_CACHE_FILES_FILE_SIZES").split(',') @@ -180,32 +192,32 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S timeStamps(i), fileSizes(i), visibilities(i)) } } - + logInfo("Prepared Local resources " + localResources) return localResources } - + def prepareEnvironment: HashMap[String, String] = { val env = new HashMap[String, String]() Client.populateClasspath(yarnConf, System.getenv("SPARK_YARN_LOG4J_PATH") != null, env) - // allow users to specify some environment variables + // Allow users to specify some environment variables Apps.setEnvFromInputString(env, System.getenv("SPARK_YARN_USER_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) - // use doAs and remoteUser here so we can add the container token and not - // pollute the current users credentials with all of the individual container tokens - val user = UserGroupInformation.createRemoteUser(container.getId().toString()); - val containerToken = container.getContainerToken(); + // Use doAs and remoteUser here so we can add the container token and not pollute the current + // users credentials with all of the individual container tokens + val user = UserGroupInformation.createRemoteUser(container.getId().toString()) + val containerToken = container.getContainerToken() if (containerToken != null) { user.addToken(ProtoUtils.convertFromProtoFormat(containerToken, cmAddress)) } @@ -216,8 +228,8 @@ class WorkerRunnable(container: Container, conf: Configuration, masterAddress: S return rpc.getProxy(classOf[ContainerManager], cmAddress, conf).asInstanceOf[ContainerManager] } - }); - return proxy; + }) + proxy } - + } diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala index 507a0743fd..f15f3c7c11 100644 --- a/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala @@ -17,55 +17,70 @@ package org.apache.spark.deploy.yarn +import java.lang.{Boolean => JBoolean} +import java.util.{Collections, Set => JSet} +import java.util.concurrent.{CopyOnWriteArrayList, ConcurrentHashMap} +import java.util.concurrent.atomic.AtomicInteger + +import scala.collection +import scala.collection.JavaConversions._ +import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet} + import org.apache.spark.Logging -import org.apache.spark.util.Utils import org.apache.spark.scheduler.SplitInfo -import scala.collection -import org.apache.hadoop.yarn.api.records.{AMResponse, ApplicationAttemptId, ContainerId, Priority, Resource, ResourceRequest, ContainerStatus, Container} import org.apache.spark.scheduler.cluster.{ClusterScheduler, CoarseGrainedSchedulerBackend} +import org.apache.spark.util.Utils + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.yarn.api.AMRMProtocol +import org.apache.hadoop.yarn.api.records.{AMResponse, ApplicationAttemptId} +import org.apache.hadoop.yarn.api.records.{Container, ContainerId, ContainerStatus} +import org.apache.hadoop.yarn.api.records.{Priority, Resource, ResourceRequest} 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. +// TODO: +// Too many params. +// 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]) +// 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]]() + // 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 + + // 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. + // Containers which have been released. private val releasedContainerList = new CopyOnWriteArrayList[ContainerId]() - // containers to be released in next request to RM + // Containers to be released in next request to RM private val pendingReleaseContainers = new ConcurrentHashMap[ContainerId, Boolean] private val numWorkersRunning = new AtomicInteger() @@ -83,23 +98,31 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM } def allocateContainers(workersToRequest: Int) { - // We need to send the request only once from what I understand ... but for now, not modifying this much. + // 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) + if (_allocatedContainers.size > 0) { + logDebug(""" + Allocated containers: %d + Current worker count: %d + Containers released: %s + Containers to be released: %s + Cluster resources: %s + """.format( + _allocatedContainers.size, + numWorkersRunning.get(), + releasedContainerList, + pendingReleaseContainers, + amResp.getAvailableResources)) val hostToContainers = new HashMap[String, ArrayBuffer[Container]]() - // ignore if not satisfying constraints { + // Ignore if not satisfying constraints { for (container <- _allocatedContainers) { if (isResourceConstraintSatisfied(container)) { // allocatedContainers += container @@ -113,8 +136,7 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM else releasedContainerList.add(container.getId()) } - // Find the appropriate containers to use - // Slightly non trivial groupBy I guess ... + // Find the appropriate containers to use. Slightly non trivial groupBy ... val dataLocalContainers = new HashMap[String, ArrayBuffer[Container]]() val rackLocalContainers = new HashMap[String, ArrayBuffer[Container]]() val offRackContainers = new HashMap[String, ArrayBuffer[Container]]() @@ -134,21 +156,22 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM 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) + // 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) + // 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 + // Now rack local if (remainingContainers != null){ val rack = YarnAllocationHandler.lookupRack(conf, candidateHost) @@ -161,15 +184,17 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM if (requiredRackCount >= remainingContainers.size){ // Add all to dataLocalContainers dataLocalContainers.put(rack, remainingContainers) - // all consumed + // 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]()) + // 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 @@ -185,8 +210,8 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM // 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. + // 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) @@ -199,33 +224,39 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM val workerHostname = container.getNodeId.getHost val containerId = container.getId - assert (container.getResource.getMemory >= (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)) + 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") + logInfo("""Ignoring container %s at host %s, since we already have the required number of + containers for it.""".format(containerId, workerHostname)) 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) + // 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"), CoarseGrainedSchedulerBackend.ACTOR_NAME) logInfo("launching container on " + containerId + " host " + workerHostname) - // just to be safe, simply remove it from pendingReleaseContainers. Should not be there, but .. + // 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]()) + 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) + if (rack != null) { + allocatedRackCount.put(rack, allocatedRackCount.getOrElse(rack, 0) + 1) + } } new Thread( @@ -234,17 +265,23 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM ).start() } } - logDebug("After allocated " + allocatedContainers.size + " containers (orig : " + - _allocatedContainers.size + "), current count " + numWorkersRunning.get() + - ", to-be-released " + releasedContainerList + ", pendingReleaseContainers : " + pendingReleaseContainers) + logDebug(""" + Finished processing %d containers. + Current number of workers running: %d, + releasedContainerList: %s, + pendingReleaseContainers: %s + """.format( + allocatedContainers.size, + numWorkersRunning.get(), + releasedContainerList, + pendingReleaseContainers)) } val completedContainers = amResp.getCompletedContainersStatuses() if (completedContainers.size > 0){ - logDebug("Completed " + completedContainers.size + " containers, current count " + numWorkersRunning.get() + - ", to-be-released " + releasedContainerList + ", pendingReleaseContainers : " + pendingReleaseContainers) - + logDebug("Completed %d containers, to-be-released: %s".format( + completedContainers.size, releasedContainerList)) for (completedContainer <- completedContainers){ val containerId = completedContainer.getContainerId @@ -253,16 +290,17 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM pendingReleaseContainers.remove(containerId) } else { - // simply decrement count - next iteration of ReporterThread will take care of allocating ! + // Simply decrement count - next iteration of ReporterThread will take care of allocating. numWorkersRunning.decrementAndGet() - logInfo("Container completed not by us ? nodeId: " + containerId + ", state " + completedContainer.getState + - " httpaddress: " + completedContainer.getDiagnostics + " exit status: " + completedContainer.getExitStatus()) - + logInfo("Completed container %s (state: %s, exit status: %s)".format( + containerId, + completedContainer.getState, + completedContainer.getExitStatus())) // Hadoop 2.2.X added a ContainerExitStatus we should switch to use // there are some exit status' we shouldn't necessarily count against us, but for // now I think its ok as none of the containers are expected to exit if (completedContainer.getExitStatus() != 0) { - logInfo("Container marked as failed: " + containerId) + logInfo("Container marked as failed: " + containerId) numWorkersFailed.incrementAndGet() } } @@ -281,7 +319,7 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM allocatedContainerToHostMap -= containerId - // doing this within locked context, sigh ... move to outside ? + // 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 @@ -291,9 +329,16 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM } } } - logDebug("After completed " + completedContainers.size + " containers, current count " + - numWorkersRunning.get() + ", to-be-released " + releasedContainerList + - ", pendingReleaseContainers : " + pendingReleaseContainers) + logDebug(""" + Finished processing %d completed containers. + Current number of workers running: %d, + releasedContainerList: %s, + pendingReleaseContainers: %s + """.format( + completedContainers.size, + numWorkersRunning.get(), + releasedContainerList, + pendingReleaseContainers)) } } @@ -347,7 +392,7 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM // default. if (numWorkers <= 0 || preferredHostToCount.isEmpty) { - logDebug("numWorkers: " + numWorkers + ", host preferences ? " + preferredHostToCount.isEmpty) + logDebug("numWorkers: " + numWorkers + ", host preferences: " + preferredHostToCount.isEmpty) resourceRequests = List( createResourceRequest(AllocationType.ANY, null, numWorkers, YarnAllocationHandler.PRIORITY)) } @@ -360,17 +405,24 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM val requiredCount = candidateCount - allocatedContainersOnHost(candidateHost) if (requiredCount > 0) { - hostContainerRequests += - createResourceRequest(AllocationType.HOST, candidateHost, requiredCount, YarnAllocationHandler.PRIORITY) + hostContainerRequests += createResourceRequest( + AllocationType.HOST, + candidateHost, + requiredCount, + YarnAllocationHandler.PRIORITY) } } - val rackContainerRequests: List[ResourceRequest] = createRackResourceRequests(hostContainerRequests.toList) + val rackContainerRequests: List[ResourceRequest] = createRackResourceRequests( + hostContainerRequests.toList) - val anyContainerRequests: ResourceRequest = - createResourceRequest(AllocationType.ANY, null, numWorkers, YarnAllocationHandler.PRIORITY) + val anyContainerRequests: ResourceRequest = createResourceRequest( + AllocationType.ANY, + resource = null, + numWorkers, + YarnAllocationHandler.PRIORITY) - val containerRequests: ArrayBuffer[ResourceRequest] = - new ArrayBuffer[ResourceRequest](hostContainerRequests.size() + rackContainerRequests.size() + 1) + val containerRequests: ArrayBuffer[ResourceRequest] = new ArrayBuffer[ResourceRequest]( + hostContainerRequests.size + rackContainerRequests.size + 1) containerRequests ++= hostContainerRequests containerRequests ++= rackContainerRequests @@ -389,52 +441,59 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM req.addAllReleases(releasedContainerList) if (numWorkers > 0) { - logInfo("Allocating " + numWorkers + " worker containers with " + (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD) + " of memory each.") + logInfo("Allocating %d worker containers with %d of memory each.".format(numWorkers, + workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)) } 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) + for (request <- resourceRequests) { + logInfo("ResourceRequest (host : %s, num containers: %d, priority = %s , capability : %s)". + format( + request.getHostName, + request.getNumContainers, + request.getPriority, + request.getCapability)) } resourceManager.allocate(req) } - private def createResourceRequest(requestType: AllocationType.AllocationType, - resource:String, numWorkers: Int, priority: Int): ResourceRequest = { + 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) - + assert(YarnAllocationHandler.ANY_HOST != resource) val hostname = resource val nodeLocal = createResourceRequestImpl(hostname, numWorkers, priority) - // add to host->rack mapping + // 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) + 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 = { + private def createResourceRequestImpl( + hostname:String, + numWorkers: Int, + priority: Int): ResourceRequest = { val rsrcRequest = Records.newRecord(classOf[ResourceRequest]) val memCapability = Records.newRecord(classOf[Resource]) @@ -455,11 +514,11 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM def createReleasedContainerList(): ArrayBuffer[ContainerId] = { val retval = new ArrayBuffer[ContainerId](1) - // iterator on COW list ... + // Iterator on COW list ... for (container <- releasedContainerList.iterator()){ retval += container } - // remove from the original list. + // Remove from the original list. if (! retval.isEmpty) { releasedContainerList.removeAll(retval) for (v <- retval) pendingReleaseContainers.put(v, true) @@ -474,14 +533,14 @@ private[yarn] class YarnAllocationHandler(val conf: Configuration, val resourceM object YarnAllocationHandler { val ANY_HOST = "*" - // all requests are issued with same priority : we do not (yet) have any distinction between + // 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 + // 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. @@ -489,38 +548,69 @@ object YarnAllocationHandler { private val rackToHostSet = new ConcurrentHashMap[String, JSet[String]]() - def newAllocator(conf: Configuration, - resourceManager: AMRMProtocol, appAttemptId: ApplicationAttemptId, - args: ApplicationMasterArguments): YarnAllocationHandler = { - - new YarnAllocationHandler(conf, resourceManager, appAttemptId, args.numWorkers, - args.workerMemory, args.workerCores, Map[String, Int](), Map[String, Int]()) + def newAllocator( + conf: Configuration, + resourceManager: AMRMProtocol, + appAttemptId: ApplicationAttemptId, + args: ApplicationMasterArguments): YarnAllocationHandler = { + + new YarnAllocationHandler( + conf, + resourceManager, + appAttemptId, + args.numWorkers, + args.workerMemory, + args.workerCores, + Map[String, Int](), + Map[String, Int]()) } - def newAllocator(conf: Configuration, - resourceManager: AMRMProtocol, appAttemptId: ApplicationAttemptId, - args: ApplicationMasterArguments, - map: collection.Map[String, collection.Set[SplitInfo]]): YarnAllocationHandler = { + 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) + 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 = { + 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) + 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]]) : + private def generateNodeToWeight( + conf: Configuration, + input: collection.Map[String, collection.Set[SplitInfo]]) : // host to count, rack to count (Map[String, Int], Map[String, Int]) = { @@ -544,7 +634,7 @@ object YarnAllocationHandler { } def lookupRack(conf: Configuration, host: String): String = { - if (! hostToRack.contains(host)) populateRackInfo(conf, host) + if (!hostToRack.contains(host)) populateRackInfo(conf, host) hostToRack.get(host) } @@ -567,10 +657,12 @@ object YarnAllocationHandler { val rack = rackInfo.getNetworkLocation hostToRack.put(hostname, rack) if (! rackToHostSet.containsKey(rack)) { - rackToHostSet.putIfAbsent(rack, Collections.newSetFromMap(new ConcurrentHashMap[String, JBoolean]())) + rackToHostSet.putIfAbsent(rack, + Collections.newSetFromMap(new ConcurrentHashMap[String, JBoolean]())) } rackToHostSet.get(rack).add(hostname) + // TODO(harvey): Figure out this comment... // 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 ... diff --git a/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala b/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala index ca2f1e2565..2ba2366ead 100644 --- a/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala +++ b/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala @@ -18,13 +18,10 @@ package org.apache.spark.deploy.yarn import org.apache.spark.deploy.SparkHadoopUtil -import collection.mutable.HashMap import org.apache.hadoop.mapred.JobConf 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. @@ -40,7 +37,7 @@ class YarnSparkHadoopUtil extends SparkHadoopUtil { // add any user credentials to the job conf which are necessary for running on a secure Hadoop cluster override def addCredentials(conf: JobConf) { - val jobCreds = conf.getCredentials(); + val jobCreds = conf.getCredentials() jobCreds.mergeAll(UserGroupInformation.getCurrentUser().getCredentials()) } } diff --git a/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala b/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala new file mode 100644 index 0000000000..63a0449e5a --- /dev/null +++ b/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala @@ -0,0 +1,47 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.scheduler.cluster + +import org.apache.spark._ +import org.apache.hadoop.conf.Configuration +import org.apache.spark.deploy.yarn.YarnAllocationHandler +import org.apache.spark.util.Utils + +/** + * + * This scheduler launch worker through Yarn - by call into Client to launch WorkerLauncher as AM. + */ +private[spark] class YarnClientClusterScheduler(sc: SparkContext, conf: Configuration) extends ClusterScheduler(sc) { + + def this(sc: SparkContext) = this(sc, new Configuration()) + + // 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 + } + + override def postStartHook() { + + // The yarn application is running, but the worker might not yet ready + // Wait for a few seconds for the slaves to bootstrap and register with master - best case attempt + Thread.sleep(2000L) + logInfo("YarnClientClusterScheduler.postStartHook done") + } +} diff --git a/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala b/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala new file mode 100644 index 0000000000..b206780c78 --- /dev/null +++ b/yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala @@ -0,0 +1,109 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.scheduler.cluster + +import org.apache.hadoop.yarn.api.records.{ApplicationId, YarnApplicationState} +import org.apache.spark.{SparkException, Logging, SparkContext} +import org.apache.spark.deploy.yarn.{Client, ClientArguments} + +private[spark] class YarnClientSchedulerBackend( + scheduler: ClusterScheduler, + sc: SparkContext) + extends CoarseGrainedSchedulerBackend(scheduler, sc.env.actorSystem) + with Logging { + + var client: Client = null + var appId: ApplicationId = null + + override def start() { + super.start() + + val defalutWorkerCores = "2" + val defalutWorkerMemory = "512m" + val defaultWorkerNumber = "1" + + val userJar = System.getenv("SPARK_YARN_APP_JAR") + var workerCores = System.getenv("SPARK_WORKER_CORES") + var workerMemory = System.getenv("SPARK_WORKER_MEMORY") + var workerNumber = System.getenv("SPARK_WORKER_INSTANCES") + + if (userJar == null) + throw new SparkException("env SPARK_YARN_APP_JAR is not set") + + if (workerCores == null) + workerCores = defalutWorkerCores + if (workerMemory == null) + workerMemory = defalutWorkerMemory + if (workerNumber == null) + workerNumber = defaultWorkerNumber + + val driverHost = System.getProperty("spark.driver.host") + val driverPort = System.getProperty("spark.driver.port") + val hostport = driverHost + ":" + driverPort + + val argsArray = Array[String]( + "--class", "notused", + "--jar", userJar, + "--args", hostport, + "--worker-memory", workerMemory, + "--worker-cores", workerCores, + "--num-workers", workerNumber, + "--master-class", "org.apache.spark.deploy.yarn.WorkerLauncher" + ) + + val args = new ClientArguments(argsArray) + client = new Client(args) + appId = client.runApp() + waitForApp() + } + + def waitForApp() { + + // TODO : need a better way to find out whether the workers are ready or not + // maybe by resource usage report? + while(true) { + val report = client.getApplicationReport(appId) + + logInfo("Application report from ASM: \n" + + "\t appMasterRpcPort: " + report.getRpcPort() + "\n" + + "\t appStartTime: " + report.getStartTime() + "\n" + + "\t yarnAppState: " + report.getYarnApplicationState() + "\n" + ) + + // Ready to go, or already gone. + val state = report.getYarnApplicationState() + if (state == YarnApplicationState.RUNNING) { + return + } else if (state == YarnApplicationState.FINISHED || + state == YarnApplicationState.FAILED || + state == YarnApplicationState.KILLED) { + throw new SparkException("Yarn application already ended," + + "might be killed or not able to launch application master.") + } + + Thread.sleep(1000) + } + } + + override def stop() { + super.stop() + client.stop() + logInfo("Stoped") + } + +} |