Spark修煉之道（高級篇）——Spark源碼閱讀：第六節 Task提交

Task提交

在上一節中的 Stage提交中我們提到，最終stage被封裝成TaskSet，使用taskScheduler.submitTasks提交，具體代碼如下：

taskScheduler.submitTasks(new TaskSet(
        tasks.toArray, stage.id, stage.latestInfo.attemptId, stage.firstJobId, properties))

Stage由一系列的tasks組成，這些task被封裝成TaskSet，TaskSet類定義如下：

/**
 * A set of tasks submitted together to the low-level TaskScheduler, usually representing
 * missing partitions of a particular stage.
 */
private[spark] class TaskSet(
    val tasks: Array[Task[_]],
    val stageId: Int,
    val stageAttemptId: Int,
    val priority: Int,
    val properties: Properties) {
    val id: String = stageId + "." + stageAttemptId

  override def toString: String = "TaskSet " + id
}

submitTasks方法定義在TaskScheduler Trait當中，目前TaskScheduler 只有一個子類TaskSchedulerImpl，其submitTasks方法源碼如下：

//TaskSchedulerImpl類中的submitTasks方法
override def submitTasks(taskSet: TaskSet) {
    val tasks = taskSet.tasks
    logInfo("Adding task set " + taskSet.id + " with " + tasks.length + " tasks")
    this.synchronized {
      //創建TaskSetManager，TaskSetManager用於對TaskSet中的Task進行調度，包括跟蹤Task的運行、Task失敗重試等
      val manager = createTaskSetManager(taskSet, maxTaskFailures)
      val stage = taskSet.stageId
      val stageTaskSets =
        taskSetsByStageIdAndAttempt.getOrElseUpdate(stage, new HashMap[Int, TaskSetManager])
      stageTaskSets(taskSet.stageAttemptId) = manager
      val conflictingTaskSet = stageTaskSets.exists { case (_, ts) =>
        ts.taskSet != taskSet && !ts.isZombie
      }
      if (conflictingTaskSet) {
        throw new IllegalStateException(s"more than one active taskSet for stage $stage:" +
          s" ${stageTaskSets.toSeq.map{_._2.taskSet.id}.mkString(",")}")
      }
      //schedulableBuilder中添加TaskSetManager，用於完成所有TaskSet的調度，即整個Spark程序生成的DAG圖對應Stage的TaskSet調度
      schedulableBuilder.addTaskSetManager(manager, manager.taskSet.properties)

      if (!isLocal && !hasReceivedTask) {
        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 " +
                "and have sufficient resources")
            } else {
              this.cancel()
            }
          }
        }, STARVATION_TIMEOUT_MS, STARVATION_TIMEOUT_MS)
      }
      hasReceivedTask = true
    }
    //爲Task分配運行資源
    backend.reviveOffers()
  }

SchedulerBackend有多種實現，如下圖所示：

我們以SparkDeploySchedulerBackend爲例進行說明，SparkDeploySchedulerBackend繼承自CoarseGrainedSchedulerBackend中的reviveOffers方法，具有代碼如下：

//CoarseGrainedSchedulerBackend中定義的reviveOffers方法
  override def reviveOffers() {
    //driverEndpoint發送ReviveOffers消息，由DriverEndPoint接受處理
    driverEndpoint.send(ReviveOffers)
  }

driverEndpoint的類型是RpcEndpointRef

//CoarseGrainedSchedulerBackend中的成員變量driverEndpoint
var driverEndpoint: RpcEndpointRef = null

它具有如下定義形式：

//RpcEndpointRef是遠程RpcEndpoint的引用，它是一個抽象類，有一個子類AkkaRpcEndpointRef
/**
 * A reference for a remote [[RpcEndpoint]]. [[RpcEndpointRef]] is thread-safe.
 */
private[spark] abstract class RpcEndpointRef(@transient conf: SparkConf)
  extends Serializable with Logging 

//在底層採用的是Akka進行實現
private[akka] class AkkaRpcEndpointRef(
    @transient defaultAddress: RpcAddress,
    @transient _actorRef: => ActorRef,
    @transient conf: SparkConf,
    @transient initInConstructor: Boolean = true)
  extends RpcEndpointRef(conf) with Logging {

  lazy val actorRef = _actorRef

  override lazy val address: RpcAddress = {
    val akkaAddress = actorRef.path.address
    RpcAddress(akkaAddress.host.getOrElse(defaultAddress.host),
      akkaAddress.port.getOrElse(defaultAddress.port))
  }

  override lazy val name: String = actorRef.path.name

  private[akka] def init(): Unit = {
    // Initialize the lazy vals
    actorRef
    address
    name
  }

  if (initInConstructor) {
    init()
  }

  override def send(message: Any): Unit = {
    actorRef ! AkkaMessage(message, false)
  }
//其它代碼省略

DriverEndpoint中的receive方法接收driverEndpoint.send(ReviveOffers)發來的消息，DriverEndpoint繼承了ThreadSafeRpcEndpoint trait，具體如下：

class DriverEndpoint(override val rpcEnv: RpcEnv, sparkProperties: Seq[(String, String)])
    extends ThreadSafeRpcEndpoint with Logging

ThreadSafeRpcEndpoint 繼承 RpcEndpoint trait，RpcEndpoint對receive方法進行了描述，具體如下：

/**
   * Process messages from [[RpcEndpointRef.send]] or [[RpcCallContext.reply)]]. If receiving a
   * unmatched message, [[SparkException]] will be thrown and sent to `onError`.
   */
  def receive: PartialFunction[Any, Unit] = {
    case _ => throw new SparkException(self + " does not implement 'receive'")
  }

DriverEndpoint 中的對其receive方法進行了重寫，具體實現如下：

 override def receive: PartialFunction[Any, Unit] = {
      case StatusUpdate(executorId, taskId, state, data) =>
        scheduler.statusUpdate(taskId, state, data.value)
        if (TaskState.isFinished(state)) {
          executorDataMap.get(executorId) match {
            case Some(executorInfo) =>
              executorInfo.freeCores += scheduler.CPUS_PER_TASK
              makeOffers(executorId)
            case None =>
              // Ignoring the update since we don't know about the executor.
              logWarning(s"Ignored task status update ($taskId state $state) " +
                s"from unknown executor with ID $executorId")
          }
        }
      //重要！處理髮送來的ReviveOffers消息
      case ReviveOffers =>
        makeOffers()

      case KillTask(taskId, executorId, interruptThread) =>
        executorDataMap.get(executorId) match {
          case Some(executorInfo) =>
            executorInfo.executorEndpoint.send(KillTask(taskId, executorId, interruptThread))
          case None =>
            // Ignoring the task kill since the executor is not registered.
            logWarning(s"Attempted to kill task $taskId for unknown executor $executorId.")
        }

    }

從上面的代碼可以看到，處理ReviveOffers消息時，調用的是makeOffers方法

  // Make fake resource offers on all executors
    private def makeOffers() {
      // Filter out executors under killing
      //所有可用的Executor
      val activeExecutors = executorDataMap.filterKeys(!executorsPendingToRemove.contains(_))
      //WorkOffer表示Executor上可用的資源，
      val workOffers = activeExecutors.map { case (id, executorData) =>
        new WorkerOffer(id, executorData.executorHost, executorData.freeCores)
      }.toSeq
      //先調用TaskSchedulerImpl的resourceOffers方法，爲Task的運行分配資源
      //再調用CoarseGrainedSchedulerBackend中的launchTasks方法啓動Task的運行，最終Task被提交到Worker節點上的Executor上運行
      launchTasks(scheduler.resourceOffers(workOffers))
    }

上面的代碼邏輯全部是在Driver端進行的，調用完launchTasks方法後，Task的執行便在Worker節點上運行了，至此完成Task的提交。
關於resourceOffers方法及launchTasks方法的具體內容，在後續章節中將進行進一步的解析。