diff options
Diffstat (limited to 'core/src/main/scala/org/apache')
16 files changed, 529 insertions, 269 deletions
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 d36e1b13a6..fbda11f578 100644 --- a/core/src/main/scala/org/apache/spark/MapOutputTracker.scala +++ b/core/src/main/scala/org/apache/spark/MapOutputTracker.scala @@ -246,12 +246,12 @@ 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 @@ -276,6 +276,10 @@ 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 6fd7a0d15a..f3ce4c879d 100644 --- a/core/src/main/scala/org/apache/spark/SparkContext.scala +++ b/core/src/main/scala/org/apache/spark/SparkContext.scala @@ -83,7 +83,7 @@ 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]] = @@ -155,123 +155,11 @@ 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 "yarn-client" => - val scheduler = try { - val clazz = Class.forName("org.apache.spark.scheduler.cluster.YarnClientClusterScheduler") - val cons = clazz.getConstructor(classOf[SparkContext]) - cons.newInstance(this).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, this).asInstanceOf[CoarseGrainedSchedulerBackend] - } catch { - case th: Throwable => { - throw new SparkException("YARN mode not available ?", th) - } - } - - 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() @@ -1138,6 +1026,124 @@ 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/rdd/RDD.scala b/core/src/main/scala/org/apache/spark/rdd/RDD.scala index f80d3d601c..ea45566ad1 100644 --- a/core/src/main/scala/org/apache/spark/rdd/RDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/RDD.scala @@ -104,7 +104,7 @@ abstract class RDD[T: ClassTag]( 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 @@ -117,7 +117,7 @@ abstract class RDD[T: ClassTag]( 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) = { @@ -126,7 +126,7 @@ abstract class RDD[T: ClassTag]( } /** User-defined generator of this RDD*/ - var generator = Utils.getCallSiteInfo.firstUserClass + @transient var generator = Utils.getCallSiteInfo.firstUserClass /** Reset generator*/ def setGenerator(_generator: String) = { @@ -938,7 +938,7 @@ abstract class RDD[T: ClassTag]( 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 elementClassTag: ClassTag[T] = classTag[T] @@ -953,7 +953,7 @@ abstract class RDD[T: ClassTag]( 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 9313bf87ec..83be3c6eb4 100644 --- a/core/src/main/scala/org/apache/spark/rdd/ZippedPartitionsRDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/ZippedPartitionsRDD.scala @@ -23,7 +23,8 @@ import scala.reflect.ClassTag 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 @@ -48,27 +49,21 @@ abstract class ZippedPartitionsBaseRDD[V: ClassTag]( 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() { 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 201572d16a..963d15b76d 100644 --- a/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala +++ b/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala @@ -115,31 +115,7 @@ class DAGScheduler( // Warns the user if a stage contains a task with size greater than this value (in KB) val TASK_SIZE_TO_WARN = 100 - private val eventProcessActor: ActorRef = env.actorSystem.actorOf(Props(new Actor { - override def preStart() { - import context.dispatcher - context.system.scheduler.schedule(RESUBMIT_TIMEOUT, RESUBMIT_TIMEOUT) { - 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) - - if (!processEvent(event)) - submitWaitingStages() - else - context.stop(self) - } - })) + private var eventProcessActor: ActorRef = _ private[scheduler] val nextJobId = new AtomicInteger(0) @@ -147,9 +123,13 @@ 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[scheduler] val stageIdToJobIds = new TimeStampedHashMap[Int, HashSet[Int]] + + private[scheduler] val stageIdToStage = new TimeStampedHashMap[Int, Stage] - private val shuffleToMapStage = new TimeStampedHashMap[Int, Stage] + private[scheduler] val shuffleToMapStage = new TimeStampedHashMap[Int, Stage] private[spark] val stageToInfos = new TimeStampedHashMap[Stage, StageInfo] @@ -180,6 +160,57 @@ 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() { + import context.dispatcher + /** + * 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, RESUBMIT_TIMEOUT, 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) } @@ -208,16 +239,16 @@ 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[_], @@ -227,21 +258,45 @@ 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. */ @@ -293,6 +348,89 @@ 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. */ @@ -381,13 +519,25 @@ 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 + @@ -397,37 +547,31 @@ 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) @@ -458,7 +602,12 @@ 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 @@ -520,6 +669,7 @@ 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 @@ -534,31 +684,59 @@ 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) @@ -579,7 +757,7 @@ 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 @@ -661,6 +839,7 @@ class DAGScheduler( activeJobs -= job resultStageToJob -= stage markStageAsFinished(stage) + jobIdToStageIdsRemove(job.jobId) listenerBus.post(SparkListenerJobEnd(job, JobSucceeded)) } job.listener.taskSucceeded(rt.outputId, event.result) @@ -697,7 +876,7 @@ 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 + @@ -714,9 +893,12 @@ 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) } } } @@ -800,21 +982,42 @@ 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") @@ -885,25 +1088,24 @@ 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/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/cluster/ClusterScheduler.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala index 8056cb2597..66ab8ea4cd 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 @@ -99,8 +99,8 @@ 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 = { @@ -172,7 +172,9 @@ 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 8884ea85a3..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 @@ -574,7 +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/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/storage/StoragePerfTester.scala b/core/src/main/scala/org/apache/spark/storage/StoragePerfTester.scala index 1e4db4f66b..d52b3d8284 100644 --- a/core/src/main/scala/org/apache/spark/storage/StoragePerfTester.scala +++ b/core/src/main/scala/org/apache/spark/storage/StoragePerfTester.scala @@ -1,3 +1,20 @@ +/* + * 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.storage import java.util.concurrent.atomic.AtomicLong 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 fbd822867f..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) @@ -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} |