本文转自Jerryshao Blog

Background

Spark在资源管理和调度方式上采用了类似于Hadoop YARN的方式,最上层是资源调度器,它负责分配资源和调度注册到Spark中的所有应用,Spark选用Mesos或是YARN等作为其资源调度框架。在每一个应用内部,Spark又实现了任务调度器,负责任务的调度和协调,类似于MapReduce。本质上,外层的资源调度和内层的任务调度相互独立,各司其职。本文对于Spark的源码分析主要集中在内层的任务调度器上,分析Spark任务调度器的实现。

Scheduler模块整体架构

scheduler模块主要分为两大部分:

  1. TaskSchedulerListenerTaskSchedulerListener部分的主要功能是监听用户提交的job,将job分解为不同的类型的stage以及相应的task,并向TaskScheduler提交task。
  2. TaskSchedulerTaskScheduler接收用户提交的task并执行。而TaskScheduler根据部署的不同又分为三个子模块:

    • ClusterScheduler
    • LocalScheduler
    • MesosScheduler

TaskSchedulerListener

Spark抽象了TaskSchedulerListener并在其上实现了DAGSchedulerDAGScheduler的主要功能是接收用户提交的job,将job根据类型划分为不同的stage,并在每一个stage内产生一系列的task,向TaskScheduler提交task。下面我们首先来看一下TaskSchedulerListener部分的类图:

DAGScheduler class chart

  • 用户所提交的job在得到DAGScheduler的调度后,会被包装成ActiveJob,同时会启动JobWaiter阻塞监听job的完成状况。
  • 于此同时依据job中RDD的dependency和dependency属性(NarrowDependencyShufflerDependecy),DAGScheduler会根据依赖关系的先后产生出不同的stage DAG(result stage, shuffle map stage)。
  • 在每一个stage内部,根据stage产生出相应的task,包括ResultTask或是ShuffleMapTask,这些task会根据RDD中partition的数量和分布,产生出一组相应的task,并将其包装为TaskSet提交到TaskScheduler上去。

RDD的依赖关系和Stage的分类

在Spark中,每一个RDD是对于数据集在某一状态下的表现形式,而这个状态有可能是从前一状态转换而来的,因此换句话说这一个RDD有可能与之前的RDD(s)有依赖关系。根据依赖关系的不同,可以将RDD分成两种不同的类型:Narrow DependencyWide Dependency

  • Narrow Dependency指的是 child RDD只依赖于parent RDD(s)固定数量的partition。
  • Wide Dependency指的是child RDD的每一个partition都依赖于parent RDD(s)所有partition。

它们之间的区别可参看下图:

RDD dependecies

根据RDD依赖关系的不同,Spark也将每一个job分为不同的stage,而stage之间的依赖关系则形成了DAG。对于Narrow Dependency,Spark会尽量多地将RDD转换放在同一个stage中;而对于Wide Dependency,由于Wide Dependency通常意味着shuffle操作,因此Spark会将此stage定义为ShuffleMapStage,以便于向MapOutputTracker注册shuffle操作。对于stage的划分可参看下图,Spark通常将shuffle操作定义为stage的边界。

different stage boundary

DAGScheduler

在用户创建SparkContext对象时,Spark会在内部创建DAGScheduler对象,并根据用户的部署情况,绑定不同的TaskSechduler,并启动DAGcheduler

private var taskScheduler: TaskScheduler = {
    //...
}
taskScheduler.start()

private var dagScheduler = new DAGScheduler(taskScheduler)
dagScheduler.start()

DAGScheduler的启动会在内部创建daemon线程,daemon线程调用run()从block queue中取出event进行处理。

private def run() {
  SparkEnv.set(env)

  while (true) {
    val event = eventQueue.poll(POLL_TIMEOUT, TimeUnit.MILLISECONDS)
    if (event != null) {
      logDebug("Got event of type " + event.getClass.getName)
    }

    if (event != null) {
      if (processEvent(event)) {
        return
      }
    }

    val time = System.currentTimeMillis() // TODO: use a pluggable clock for testability
    if (failed.size > 0 && time > lastFetchFailureTime + RESUBMIT_TIMEOUT) {
      resubmitFailedStages()
    } else {
      submitWaitingStages()
    }
  }
}

run()会调用processEvent来处理不同的event。

DAGScheduler处理的event包括:

  • JobSubmitted
  • CompletionEvent
  • ExecutorLost
  • TaskFailed
  • StopDAGScheduler

根据event的不同调用不同的方法去处理。

本质上DAGScheduler是一个生产者-消费者模型,用户和TaskSchduler产生event将其放入block queue,daemon线程消费event并处理相应事件。

Job的生与死

既然用户提交的job最终会交由DAGScheduler去处理,那么我们就来研究一下DAGScheduler处理job的整个流程。在这里我们分析两种不同类型的job的处理流程。

  1. 没有shuffle和reduce的job

    val textFile = sc.textFile(“README.md”) textFile.filter(line => line.contains(“Spark”)).count()
  2. 有shuffle和reduce的job

    val textFile = sc.textFile(“README.md”) textFile.flatMap(line => line.split(“ “)).map(word => (word, 1)).reduceByKey((a, b) => a + b)

首先在对RDDcount()reduceByKey()操作都会调用SparkContextrunJob()来提交job,而SparkContextrunJob()最终会调用DAGSchedulerrunJob()

def runJob[T, U: ClassManifest](
    finalRdd: RDD[T],
    func: (TaskContext, Iterator[T]) => U,
    partitions: Seq[Int],
    callSite: String,
    allowLocal: Boolean,
    resultHandler: (Int, U) => Unit)
{
  if (partitions.size == 0) {
    return
  }
  val (toSubmit, waiter) = prepareJob(
      finalRdd, func, partitions, callSite, allowLocal, resultHandler)
  eventQueue.put(toSubmit)
  waiter.awaitResult() match {
    case JobSucceeded => {}
    case JobFailed(exception: Exception) =>
      logInfo("Failed to run " + callSite)
      throw exception
  }
}

runJob()会调用prepareJob()对job进行预处理,封装成JobSubmitted事件,放入queue中,并阻塞等待job完成。

当daemon线程的processEvent()从queue中取出JobSubmitted事件后,会根据job划分出不同的stage,并且提交stage:

case JobSubmitted(finalRDD, func, partitions, allowLocal, callSite, listener) =>
  val runId = nextRunId.getAndIncrement()
  val finalStage = newStage(finalRDD, None, runId)
  val job = new ActiveJob(runId, finalStage, func, partitions, callSite, listener)
  clearCacheLocs()
  if (allowLocal && finalStage.parents.size == 0 && partitions.length == 1) {
    runLocally(job)
  } else {
    activeJobs += job
    resultStageToJob(finalStage) = job
    submitStage(finalStage)
  }

首先,对于任何的job都会产生出一个finalStage来产生和提交task。其次对于某些简单的job,它没有依赖关系,并且只有一个partition,这样的job会使用local thread处理而并非提交到TaskScheduler上处理。

接下来产生finalStage后,需要调用submitStage(),它根据stage之间的依赖关系得出stage DAG,并以依赖关系进行处理:

private def submitStage(stage: Stage) {
  if (!waiting(stage) && !running(stage) && !failed(stage)) {
    val missing = getMissingParentStages(stage).sortBy(_.id)
    if (missing == Nil) {
      submitMissingTasks(stage)
      running += stage
    } else {
      for (parent <- missing) {
        submitStage(parent)
      }
      waiting += stage
    }
  }
}

对于新提交的job,finalStage的parent stage还未获得,因此submitStage会调用getMissingParentStages()来获得依赖关系:

private def getMissingParentStages(stage: Stage): List[Stage] = {
  val missing = new HashSet[Stage]
  val visited = new HashSet[RDD[_]]
  def visit(rdd: RDD[_]) {
    if (!visited(rdd)) {
      visited += rdd
      if (getCacheLocs(rdd).contains(Nil)) {
        for (dep <- rdd.dependencies) {
          dep match {
            case shufDep: ShuffleDependency[_,_] =>
              val mapStage = getShuffleMapStage(shufDep, stage.priority)
              if (!mapStage.isAvailable) {
                missing += mapStage
              }
            case narrowDep: NarrowDependency[_] =>
              visit(narrowDep.rdd)
          }
        }
      }
    }
  }
  visit(stage.rdd)
  missing.toList
}

这里parent stage是通过RDD的依赖关系递归遍历获得。对于Wide Dependecy也就是Shuffle Dependecy,Spark会产生新的mapStage作为finalStage的parent,而对于Narrow Dependecy Spark则不会产生新的stage。这里对stage的划分是按照上面提到的作为划分依据的,因此对于本段开头提到的两种job,第一种job只会产生一个finalStage,而第二种job会产生finalStagemapStage

当stage DAG产生以后,针对每个stage需要产生task去执行,故在这会调用submitMissingTasks()

private def submitMissingTasks(stage: Stage) {
  val myPending = pendingTasks.getOrElseUpdate(stage, new HashSet)
  myPending.clear()
  var tasks = ArrayBuffer[Task[_]]()
  if (stage.isShuffleMap) {
    for (p <- 0 until stage.numPartitions if stage.outputLocs(p) == Nil) {
      val locs = getPreferredLocs(stage.rdd, p)
      tasks += new ShuffleMapTask(stage.id, stage.rdd, stage.shuffleDep.get, p, locs)
    }
  } else {
    val job = resultStageToJob(stage)
    for (id <- 0 until job.numPartitions if (!job.finished(id))) {
      val partition = job.partitions(id)
      val locs = getPreferredLocs(stage.rdd, partition)
      tasks += new ResultTask(stage.id, stage.rdd, job.func, partition, locs, id)
    }
  }
  if (tasks.size > 0) {
    myPending ++= tasks
    taskSched.submitTasks(
      new TaskSet(tasks.toArray, stage.id, stage.newAttemptId(), stage.priority))
    if (!stage.submissionTime.isDefined) {
      stage.submissionTime = Some(System.currentTimeMillis())
    }
  } else {
    running -= stage
  }
}

首先根据stage所依赖的RDD的partition的分布,会产生出与partition数量相等的task,这些task根据partition的locality进行分布;其次对于finalStage或是mapStage会产生不同的task;最后所有的task会封装到TaskSet内提交到TaskScheduler去执行。

至此job在DAGScheduler内的启动过程全部完成,交由TaskScheduler执行task,当task执行完后会将结果返回给DAGSchedulerDAGScheduler调用handleTaskComplete()处理task返回:

private def handleTaskCompletion(event: CompletionEvent) {
  val task = event.task
  val stage = idToStage(task.stageId)

  def markStageAsFinished(stage: Stage) = {
    val serviceTime = stage.submissionTime match {
      case Some(t) => "%.03f".format((System.currentTimeMillis() - t) / 1000.0)
      case _ => "Unkown"
    }
    logInfo("%s (%s) finished in %s s".format(stage, stage.origin, serviceTime))
    running -= stage
  }
  event.reason match {
    case Success =>
        ...
      task match {
        case rt: ResultTask[_, _] =>
          ...
        case smt: ShuffleMapTask =>
          ...
      }
    case Resubmitted =>
      ...

    case FetchFailed(bmAddress, shuffleId, mapId, reduceId) =>
      ...
    case other =>
      abortStage(idToStage(task.stageId), task + " failed: " + other)
  }
}

每个执行完成的task都会将结果返回给DAGSchedulerDAGScheduler根据返回结果来进行进一步的动作。

RDD的计算

RDD的计算是在task中完成的。我们之前提到task分为ResultTaskShuffleMapTask,我们分别来看一下这两种task具体的执行过程。

  • ResultTask

      override def run(attemptId: Long): U = {
        val context = new TaskContext(stageId, partition, attemptId)
        try {
          func(context, rdd.iterator(split, context))
        } finally {
          context.executeOnCompleteCallbacks()
        }
      }
    
  • ShuffleMapTask

      override def run(attemptId: Long): MapStatus = {
        val numOutputSplits = dep.partitioner.numPartitions
        
        val taskContext = new TaskContext(stageId, partition, attemptId)
        try {
          val buckets = Array.fill(numOutputSplits)(new ArrayBuffer[(Any, Any)])
          for (elem <- rdd.iterator(split, taskContext)) {
            val pair = elem.asInstanceOf[(Any, Any)]
            val bucketId = dep.partitioner.getPartition(pair._1)
            buckets(bucketId) += pair
          }
        
          val compressedSizes = new Array[Byte](numOutputSplits)
        
          val blockManager = SparkEnv.get.blockManager
          for (i <- 0 until numOutputSplits) {
            val blockId = "shuffle_" + dep.shuffleId + "_" + partition + "_" + i
            val iter: Iterator[(Any, Any)] = buckets(i).iterator
            val size = blockManager.put(blockId, iter, StorageLevel.DISK_ONLY, false)
            compressedSizes(i) = MapOutputTracker.compressSize(size)
          }
        
          return new MapStatus(blockManager.blockManagerId, compressedSizes)
        } finally {
          taskContext.executeOnCompleteCallbacks()
        }
      }
    

ResultTaskShuffleMapTask都会调用RDDiterator()来计算和转换RDD,不同的是:ResultTask转换完RDD后调用func()计算结果;而ShufflerMapTask则将其放入blockManager中用来shuffle。

RDD的计算调用iterator()iterator()在内部调用compute()RDD依赖关系的根开始计算:

final def iterator(split: Partition, context: TaskContext): Iterator[T] = {
  if (storageLevel != StorageLevel.NONE) {
    SparkEnv.get.cacheManager.getOrCompute(this, split, context, storageLevel)
  } else {
    computeOrReadCheckpoint(split, context)
  }
}

private[spark] def computeOrReadCheckpoint(split: Partition, context: TaskContext): Iterator[T] = {
  if (isCheckpointed) {
    firstParent[T].iterator(split, context)
  } else {
    compute(split, context)
  }
}

至此大致分析了TaskSchedulerListener,包括DAGScheduler内部的结构,job生命周期内的活动,RDD是何时何地计算的。接下来我们分析一下task在TaskScheduler内干了什么。

TaskScheduler

前面也提到了Spark实现了三种不同的TaskScheduler,包括LocalShedulerClusterSchedulerMesosSchedulerLocalSheduler是一个在本地执行的线程池,DAGScheduler提交的所有task会在线程池中被执行,并将结果返回给DAGSchedulerMesosScheduler依赖于Mesos进行调度,笔者对Mesos了解甚少,因此不做分析。故此章节主要分析ClusterScheduler模块。

ClusterScheduler模块与deploy模块和executor模块耦合较为紧密,因此在分析ClUsterScheduler时也会顺带介绍deploy和executor模块。

首先我们来看一下ClusterScheduler的类图:

ClusterScheduler

ClusterScheduler的启动会伴随SparkDeploySchedulerBackend的启动,而backend会将自己分为两个角色:首先是driver,driver是一个local运行的actor,负责与remote的executor进行通行,提交任务,控制executor;其次是StandaloneExecutorBackend,Spark会在每一个slave node上启动一个StandaloneExecutorBackend进程,负责执行任务,返回执行结果。

ClusterScheduler的启动

SparkContext实例化的过程中,ClusterScheduler被随之实例化,同时赋予其SparkDeploySchedulerBackend

  master match {
      ...

    case SPARK_REGEX(sparkUrl) =>
      val scheduler = new ClusterScheduler(this)
      val backend = new SparkDeploySchedulerBackend(scheduler, this, sparkUrl, appName)
      scheduler.initialize(backend)
      scheduler

    case LOCAL_CLUSTER_REGEX(numSlaves, coresPerSlave, memoryPerSlave) =>
      ...
    case _ =>
      ...
  }
}
taskScheduler.start()

ClusterScheduler的启动会启动SparkDeploySchedulerBackend,同时启动daemon进程来检查speculative task:

override def start() {
  backend.start()

  if (System.getProperty("spark.speculation", "false") == "true") {
    new Thread("ClusterScheduler speculation check") {
      setDaemon(true)

      override def run() {
        while (true) {
          try {
            Thread.sleep(SPECULATION_INTERVAL)
          } catch {
            case e: InterruptedException => {}
          }
          checkSpeculatableTasks()
        }
      }
    }.start()
  }
}

SparkDeploySchedulerBacked的启动首先会调用父类的start(),接着它会启动client,并由client连接到master向每一个node的worker发送请求启动StandaloneExecutorBackend。这里的client、master、worker涉及到了deploy模块,暂时不做具体介绍。而StandaloneExecutorBackend则涉及到了executor模块,它主要的功能是在每一个node创建task可以运行的环境,并让task在其环境中运行。

override def start() {
  super.start()

  val driverUrl = "akka://spark@%s:%s/user/%s".format(
    System.getProperty("spark.driver.host"), System.getProperty("spark.driver.port"),
    StandaloneSchedulerBackend.ACTOR_NAME)
  val args = Seq(driverUrl, "", "", "")
  val command = Command("spark.executor.StandaloneExecutorBackend", args, sc.executorEnvs)
  val sparkHome = sc.getSparkHome().getOrElse(
    throw new IllegalArgumentException("must supply spark home for spark standalone"))
  val appDesc = new ApplicationDescription(appName, maxCores, executorMemory, command, sparkHome)

  client = new Client(sc.env.actorSystem, master, appDesc, this)
  client.start()
}

StandaloneSchedulerBackend中会创建DriverActor,它就是local的driver,以actor的方式与remote的executor进行通信。

override def start() {
  val properties = new ArrayBuffer[(String, String)]
  val iterator = System.getProperties.entrySet.iterator
  while (iterator.hasNext) {
    val entry = iterator.next
    val (key, value) = (entry.getKey.toString, entry.getValue.toString)
    if (key.startsWith("spark.")) {
      properties += ((key, value))
    }
  }
  driverActor = actorSystem.actorOf(
    Props(new DriverActor(properties)), name = StandaloneSchedulerBackend.ACTOR_NAME)
}

在client实例化之前,会将StandaloneExecutorBackend的启动环境作为参数传递给client,而client启动时会将此提交给master,由master分发给所有node上的worker,worker会配置环境并创建进程启动StandaloneExecutorBackend

至此ClusterScheduler的启动,local driver的创建,remote executor环境的启动所有过程都已结束,ClusterScheduler等待DAGScheduler提交任务。

ClusterScheduler提交任务

DAGScheduler会调用ClusterScheduler提交任务,任务会被包装成TaskSetManager并等待调度:

override def submitTasks(taskSet: TaskSet) {
  val tasks = taskSet.tasks
  logInfo("Adding task set " + taskSet.id + " with " + tasks.length + " tasks")
  this.synchronized {
    val manager = new TaskSetManager(this, taskSet)
    activeTaskSets(taskSet.id) = manager
    activeTaskSetsQueue += manager
    taskSetTaskIds(taskSet.id) = new HashSet[Long]()

    if (hasReceivedTask == false) {
      starvationTimer.scheduleAtFixedRate(new TimerTask() {
        override def run() {
          if (!hasLaunchedTask) {
            logWarning("Initial job has not accepted any resources; " +
              "check your cluster UI to ensure that workers are registered")
          } else {
            this.cancel()
          }
        }
      }, STARVATION_TIMEOUT, STARVATION_TIMEOUT)
    }
    hasReceivedTask = true;
  }
  backend.reviveOffers()
}

在任务提交的同时会启动定时器,如果任务还未被执行,定时器持续发出警告直到任务被执行。同时会调用StandaloneSchedulerBackendreviveOffers(),而它则会通过actor向driver发送ReviveOffers,driver收到ReviveOffers后调用makeOffers()

// Make fake resource offers on just one executor
def makeOffers(executorId: String) {
  launchTasks(scheduler.resourceOffers(
    Seq(new WorkerOffer(executorId, executorHost(executorId), freeCores(executorId)))))
}

// Launch tasks returned by a set of resource offers
def launchTasks(tasks: Seq[Seq[TaskDescription]]) {
  for (task <- tasks.flatten) {
    freeCores(task.executorId) -= 1
    executorActor(task.executorId) ! LaunchTask(task)
  }
}

makeOffers()会向ClusterScheduler申请资源,并向executor提交LauchTask请求。

接下来LaunchTask会进入executor模块,StandaloneExecutorBackend在收到LaunchTask请求后会调用Executor执行task:

override def receive = {
  case RegisteredExecutor(sparkProperties) =>
    ...  
  case RegisterExecutorFailed(message) =>
    ...
  case LaunchTask(taskDesc) =>
    logInfo("Got assigned task " + taskDesc.taskId)
    executor.launchTask(this, taskDesc.taskId, taskDesc.serializedTask)

  case Terminated(_) | RemoteClientDisconnected(_, _) | RemoteClientShutdown(_, _) =>
    ...
}

def launchTask(context: ExecutorBackend, taskId: Long, serializedTask: ByteBuffer) {
  threadPool.execute(new TaskRunner(context, taskId, serializedTask))
}

Executor内部是一个线程池,每一个提交的task都会包装为TaskRunner交由threadpool执行:

class TaskRunner(context: ExecutorBackend, taskId: Long, serializedTask: ByteBuffer)
  extends Runnable {

  override def run() {
    SparkEnv.set(env)
    Thread.currentThread.setContextClassLoader(urlClassLoader)
    val ser = SparkEnv.get.closureSerializer.newInstance()
    logInfo("Running task ID " + taskId)
    context.statusUpdate(taskId, TaskState.RUNNING, EMPTY_BYTE_BUFFER)
    try {
      SparkEnv.set(env)
      Accumulators.clear()
      val (taskFiles, taskJars, taskBytes) = Task.deserializeWithDependencies(serializedTask)
      updateDependencies(taskFiles, taskJars)
      val task = ser.deserialize[Task[Any]](taskBytes, Thread.currentThread.getContextClassLoader)
      logInfo("Its generation is " + task.generation)
      env.mapOutputTracker.updateGeneration(task.generation)
      val value = task.run(taskId.toInt)
      val accumUpdates = Accumulators.values
      val result = new TaskResult(value, accumUpdates)
      val serializedResult = ser.serialize(result)
      logInfo("Serialized size of result for " + taskId + " is " + serializedResult.limit)
      context.statusUpdate(taskId, TaskState.FINISHED, serializedResult)
      logInfo("Finished task ID " + taskId)
    } catch {
      case ffe: FetchFailedException => {
        val reason = ffe.toTaskEndReason
        context.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason))
      }

      case t: Throwable => {
        val reason = ExceptionFailure(t)
        context.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason))

        // TODO: Should we exit the whole executor here? On the one hand, the failed task may
        // have left some weird state around depending on when the exception was thrown, but on
        // the other hand, maybe we could detect that when future tasks fail and exit then.
        logError("Exception in task ID " + taskId, t)
        //System.exit(1)
      }
    }
  }
}

其中task.run()则真正执行了task中的任务,如前RDD的计算章节所述。返回值被包装成TaskResult返回。

至此task在ClusterScheduler内运行的流程有了一个大致的介绍,当然这里略掉了许多异常处理的分支,但这不影响我们对主线的了解。

END

至此对Spark的Scheduler模块的主线做了一个顺藤摸瓜式的介绍,Scheduler模块作为Spark最核心的模块之一,充分体现了Spark与MapReduce的不同之处,体现了Spark DAG思想的精巧和设计的优雅。

当然Spark的代码仍然在积极开发之中,当前的源码分析在过不久后可能会变得没有意义,但重要的是体会Spark区别于MapReduce的设计理念,以及DAG思想的应用。DAG作为对MapReduce框架的改进越来越受到大数据界的重视,hortonworks也提出了类似DAG的框架tez作为对MapReduce的改进。