這篇博文，我們就來講講Executor啟動后，是如何在Executor上執行Task的，以及其后續處理。

執行Task

我們在《深入理解Spark 2.1 Core （三）：任務調度器的原理與源碼分析 》中提到了，任務調度完成后，CoarseGrainedSchedulerBackend.DriverEndpoint會調用launchTasks向CoarseGrainedExecutorBackend發送帶著serializedTask的LaunchTask信號。接下來，我們就來講講CoarseGrainedExecutorBackend接收到LaunchTask信號后，是如何執行Task的。

調用棧如下：

• CoarseGrainedExecutorBackend.receive
  
  • Executor.launchTask
    
    • Executor.TaskRunner.run
      
      • Executor.updateDependencies
      • Task.run
        
        • ShuffleMapTask.runTask
        • ResultTask.runTask

CoarseGrainedExecutorBackend.receive

case LaunchTask(data) =>if (executor == null) {exitExecutor(1, "Received LaunchTask command but executor was null")} else {// 反序列話task描述val taskDesc = ser.deserialize[TaskDescription](data.value) logInfo("Got assigned task " + taskDesc.taskId)// 調用executor.launchTaskexecutor.launchTask(this, taskId = taskDesc.taskId, attemptNumber = taskDesc.attemptNumber,taskDesc.name, taskDesc.serializedTask)}

Executor.launchTask

def launchTask(context: ExecutorBackend,taskId: Long,attemptNumber: Int,taskName: String,serializedTask: ByteBuffer): Unit = {// 創建TaskRunnerval tr = new TaskRunner(context, taskId = taskId, attemptNumber = attemptNumber, taskName,serializedTask)// 把taskID 以及 對應的 TaskRunner，// 加入到 ConcurrentHashMap[Long, TaskRunner]runningTasks.put(taskId, tr)// 線程池 執行 TaskRunnerthreadPool.execute(tr)}

• 1

Executor.TaskRunner.run

override def run(): Unit = {val threadMXBean = ManagementFactory.getThreadMXBeanval taskMemoryManager = new TaskMemoryManager(env.memoryManager, taskId)// 記錄開始反序列化的時間val deserializeStartTime = System.currentTimeMillis()// 記錄開始反序列化的時的Cpu時間val deserializeStartCpuTime = if (threadMXBean.isCurrentThreadCpuTimeSupported) {threadMXBean.getCurrentThreadCpuTime} else 0LThread.currentThread.setContextClassLoader(replClassLoader)val ser = env.closureSerializer.newInstance()logInfo(s"Running $taskName (TID $taskId)")execBackend.statusUpdate(taskId, TaskState.RUNNING, EMPTY_BYTE_BUFFER)var taskStart: Long = 0var taskStartCpu: Long = 0// 開始GC的時間startGCTime = computeTotalGcTime()try {//反序列化任務信息val (taskFiles, taskJars, taskProps, taskBytes) =Task.deserializeWithDependencies(serializedTask)// 根據taskProps設置executor屬性Executor.taskDeserializationProps.set(taskProps)// 根據taskFiles和taskJars，// 下載任務所需的File 和 加載所需的Jar包updateDependencies(taskFiles, taskJars)// 根據taskBytes生成tasktask = ser.deserialize[Task[Any]](taskBytes, Thread.currentThread.getContextClassLoader)//設置task屬性task.localProperties = taskProps//設置task內存管理task.setTaskMemoryManager(taskMemoryManager)// 若在反序列話之前Task就被kill了，// 拋出異常if (killed) {throw new TaskKilledException}logDebug("Task " + taskId + "'s epoch is " + task.epoch)//更新mapOutputTracker Epoch 為task epochenv.mapOutputTracker.updateEpoch(task.epoch)// 記錄任務開始時間taskStart = System.currentTimeMillis()// 記錄任務開始時的cpu時間taskStartCpu = if (threadMXBean.isCurrentThreadCpuTimeSupported) {threadMXBean.getCurrentThreadCpuTime} else 0Lvar threwException = trueval value = try {// 運行Taskval res = task.run(taskAttemptId = taskId,attemptNumber = attemptNumber,metricsSystem = env.metricsSystem)threwException = falseres} finally {val releasedLocks = env.blockManager.releaseAllLocksForTask(taskId)val freedMemory = taskMemoryManager.cleanUpAllAllocatedMemory()if (freedMemory > 0 && !threwException) {val errMsg = s"Managed memory leak detected; size = $freedMemory bytes, TID = $taskId"if (conf.getBoolean("spark.unsafe.exceptionOnMemoryLeak", false)) {throw new SparkException(errMsg)} else {logWarning(errMsg)}}if (releasedLocks.nonEmpty && !threwException) {val errMsg =s"${releasedLocks.size} block locks were not released by TID = $taskId:\n" +releasedLocks.mkString("[", ", ", "]")if (conf.getBoolean("spark.storage.exceptionOnPinLeak", false)) {throw new SparkException(errMsg)} else {logWarning(errMsg)}}}// 記錄任務結束時間val taskFinish = System.currentTimeMillis()// 記錄任務結束時的cpu時間val taskFinishCpu = if (threadMXBean.isCurrentThreadCpuTimeSupported) {threadMXBean.getCurrentThreadCpuTime} else 0L// 若task在運行中被kill了// 則拋出異常if (task.killed) {throw new TaskKilledException}val resultSer = env.serializer.newInstance()// 結果記錄序列化開始的系統時間val beforeSerialization = System.currentTimeMillis()// 序列化結果val valueBytes = resultSer.serialize(value)// 結果記錄序列化完成的系統時間val afterSerialization = System.currentTimeMillis()// 反序列話發生在兩個地方：// 1. 在該函數下反序列化Task信息以及Task實例。// 2. 在任務啟動后，Task.run 反序列化 RDD 和 函數// 計算task的反序列化費時task.metrics.setExecutorDeserializeTime((taskStart - deserializeStartTime) + task.executorDeserializeTime)// 計算task的反序列化cpu費時task.metrics.setExecutorDeserializeCpuTime((taskStartCpu - deserializeStartCpuTime) + task.executorDeserializeCpuTime)// 計算task運行費時task.metrics.setExecutorRunTime((taskFinish - taskStart) - task.executorDeserializeTime)// 計算task運行cpu費時task.metrics.setExecutorCpuTime((taskFinishCpu - taskStartCpu) - task.executorDeserializeCpuTime)// 計算GC時間task.metrics.setJvmGCTime(computeTotalGcTime() - startGCTime)//計算結果序列化時間 task.metrics.setResultSerializationTime(afterSerialization - beforeSerialization)val accumUpdates = task.collectAccumulatorUpdates()// 這里代碼存在缺陷：// value相當于被序列化了兩次val directResult = new DirectTaskResult(valueBytes, accumUpdates)val serializedDirectResult = ser.serialize(directResult)// 得到結果的大小val resultSize = serializedDirectResult.limit// 對于計算結果，會根據結果的大小有不同的策略：// 1.生成結果在(正無窮,1GB)：// 超過1GB的部分結果直接丟棄，// 可以通過spark.driver.maxResultSize實現// 默認為1G// 2.生成結果大小在$[1GB,128MB - 200KB]// 會把該結果以taskId為編號存入BlockManager中，// 然后把該編號通過Netty發送給Driver，// 該閾值是Netty框架傳輸的最大值// spark.akka.frameSize（默認為128MB）和Netty的預留空間reservedSizeBytes（200KB）的差值// 3.生成結果大小在（128MB - 200KB,0）：// 直接通過Netty發送到Driverval serializedResult: ByteBuffer = {if (maxResultSize > 0 && resultSize > maxResultSize) {logWarning(s"Finished $taskName (TID $taskId). Result is larger than maxResultSize " +s"(${Utils.bytesToString(resultSize)} > ${Utils.bytesToString(maxResultSize)}), " +s"dropping it.")ser.serialize(new IndirectTaskResult[Any](TaskResultBlockId(taskId), resultSize))} else if (resultSize > maxDirectResultSize) {val blockId = TaskResultBlockId(taskId)env.blockManager.putBytes(blockId,new ChunkedByteBuffer(serializedDirectResult.duplicate()),StorageLevel.MEMORY_AND_DISK_SER)logInfo(s"Finished $taskName (TID $taskId). $resultSize bytes result sent via BlockManager)")ser.serialize(new IndirectTaskResult[Any](blockId, resultSize))} else {logInfo(s"Finished $taskName (TID $taskId). $resultSize bytes result sent to driver")serializedDirectResult}}// 更新execBackend 狀態execBackend.statusUpdate(taskId, TaskState.FINISHED, serializedResult)} catch {case ffe: FetchFailedException =>val reason = ffe.toTaskFailedReasonsetTaskFinishedAndClearInterruptStatus()execBackend.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason))case _: TaskKilledException =>logInfo(s"Executor killed $taskName (TID $taskId)")setTaskFinishedAndClearInterruptStatus()execBackend.statusUpdate(taskId, TaskState.KILLED, ser.serialize(TaskKilled))case _: InterruptedException if task.killed =>logInfo(s"Executor interrupted and killed $taskName (TID $taskId)")setTaskFinishedAndClearInterruptStatus()execBackend.statusUpdate(taskId, TaskState.KILLED, ser.serialize(TaskKilled))case CausedBy(cDE: CommitDeniedException) =>val reason = cDE.toTaskFailedReasonsetTaskFinishedAndClearInterruptStatus()execBackend.statusUpdate(taskId, TaskState.FAILED, ser.serialize(reason))case t: Throwable =>logError(s"Exception in $taskName (TID $taskId)", t)val accums: Seq[AccumulatorV2[_, _]] =if (task != null) {task.metrics.setExecutorRunTime(System.currentTimeMillis() - taskStart)task.metrics.setJvmGCTime(computeTotalGcTime() - startGCTime)task.collectAccumulatorUpdates(taskFailed = true)} else {Seq.empty}val accUpdates = accums.map(acc => acc.toInfo(Some(acc.value), None))val serializedTaskEndReason = {try {ser.serialize(new ExceptionFailure(t, accUpdates).withAccums(accums))} catch {case _: NotSerializableException =>ser.serialize(new ExceptionFailure(t, accUpdates, false).withAccums(accums))}}setTaskFinishedAndClearInterruptStatus()execBackend.statusUpdate(taskId, TaskState.FAILED, serializedTaskEndReason)if (Utils.isFatalError(t)) {SparkUncaughtExceptionHandler.uncaughtException(t)}} finally {// 任務結束后移除runningTasks.remove(taskId)}}

• 1

Executor.updateDependencies

接下來，我們來看看更新executor的依賴，即下載任務所需的File和加載所需的Jar包：

private def updateDependencies(newFiles: HashMap[String, Long], newJars: HashMap[String, Long]) {lazy val hadoopConf = SparkHadoopUtil.get.newConfiguration(conf)synchronized {// 下載任務所需的Filefor ((name, timestamp) <- newFiles if currentFiles.getOrElse(name, -1L) < timestamp) {logInfo("Fetching " + name + " with timestamp " + timestamp)Utils.fetchFile(name, new File(SparkFiles.getRootDirectory()), conf,env.securityManager, hadoopConf, timestamp, useCache = !isLocal)currentFiles(name) = timestamp}// 加載所需的Jar包for ((name, timestamp) <- newJars) {val localName = name.split("/").lastval currentTimeStamp = currentJars.get(name).orElse(currentJars.get(localName)).getOrElse(-1L)if (currentTimeStamp < timestamp) {logInfo("Fetching " + name + " with timestamp " + timestamp)Utils.fetchFile(name, new File(SparkFiles.getRootDirectory()), conf,env.securityManager, hadoopConf, timestamp, useCache = !isLocal)currentJars(name) = timestamp// 把它加入到 class loaderval url = new File(SparkFiles.getRootDirectory(), localName).toURI.toURLif (!urlClassLoader.getURLs().contains(url)) {logInfo("Adding " + url + " to class loader")urlClassLoader.addURL(url)}}}}}

• 1

Task.run

接下來，我們來看看這篇博文最核心的部分——task運行：

final def run(taskAttemptId: Long,attemptNumber: Int,metricsSystem: MetricsSystem): T = {SparkEnv.get.blockManager.registerTask(taskAttemptId)//創建TaskContextImplcontext = new TaskContextImpl(stageId,partitionId,taskAttemptId,attemptNumber,taskMemoryManager,localProperties,metricsSystem,metrics)//在TaskContext中設置TaskContextImplTaskContext.setTaskContext(context)taskThread = Thread.currentThread()if (_killed) {kill(interruptThread = false)}new CallerContext("TASK", appId, appAttemptId, jobId, Option(stageId), Option(stageAttemptId),Option(taskAttemptId), Option(attemptNumber)).setCurrentContext()try {// 調用runTaskrunTask(context)} catch {case e: Throwable =>try {context.markTaskFailed(e)} catch {case t: Throwable =>e.addSuppressed(t)}throw e} finally {// 標記Task完成context.markTaskCompleted()try {Utils.tryLogNonFatalError {// 釋放內存SparkEnv.get.blockManager.memoryStore.releaseUnrollMemoryForThisTask(MemoryMode.ON_HEAP)SparkEnv.get.blockManager.memoryStore.releaseUnrollMemoryForThisTask(MemoryMode.OFF_HEAP)val memoryManager = SparkEnv.get.memoryManagermemoryManager.synchronized { memoryManager.notifyAll() }}} finally {//取消TaskContext設置TaskContext.unset()}}}

• 1

Task有兩個子類，一個是非最后的Stage的Task，ShuffleMapTask；一個是最后的Stage的Task，ResultTask。它們都覆蓋了Task的runTask方法，接下來我們就分別來講下它們的runTask方法。

ShuffleMapTask.runTask

根據每個Stage的partition數量來生成ShuffleMapTask，ShuffleMapTask會根據下游的Partition數量和Shuffle的策略來生成一系列文件。

override def runTask(context: TaskContext): MapStatus = {val threadMXBean = ManagementFactory.getThreadMXBean// 記錄反序列化開始時間val deserializeStartTime = System.currentTimeMillis()// 記錄反序列化開始時的Cpu時間val deserializeStartCpuTime = if (threadMXBean.isCurrentThreadCpuTimeSupported) {threadMXBean.getCurrentThreadCpuTime} else 0Lval ser = SparkEnv.get.closureSerializer.newInstance()// 反序列化rdd 及其 依賴val (rdd, dep) = ser.deserialize[(RDD[_], ShuffleDependency[_, _, _])](ByteBuffer.wrap(taskBinary.value), Thread.currentThread.getContextClassLoader)// 計算 反序列化費時_executorDeserializeTime = System.currentTimeMillis() - deserializeStartTime// 計算 反序列化Cpu費時_executorDeserializeCpuTime = if (threadMXBean.isCurrentThreadCpuTimeSupported) {threadMXBean.getCurrentThreadCpuTime - deserializeStartCpuTime} else 0Lvar writer: ShuffleWriter[Any, Any] = nulltry {//獲取shuffleManagerval manager = SparkEnv.get.shuffleManager// writerwriter = manager.getWriter[Any, Any](dep.shuffleHandle, partitionId, context)// 調用writer.write 開始計算RDD，// 這部分 我們會在后續博文講解writer.write(rdd.iterator(partition, context).asInstanceOf[Iterator[_ <: Product2[Any, Any]]])// 停止計算，并返回結果writer.stop(success = true).get} catch {case e: Exception =>try {if (writer != null) {writer.stop(success = false)}} catch {case e: Exception =>log.debug("Could not stop writer", e)}throw e}}

• 1

ResultTask.runTask

override def runTask(context: TaskContext): U = {val threadMXBean = ManagementFactory.getThreadMXBean// 記錄反序列化開始時間val deserializeStartTime = System.currentTimeMillis()// 記錄反序列化開始時的Cpu時間val deserializeStartCpuTime = if (threadMXBean.isCurrentThreadCpuTimeSupported) {threadMXBean.getCurrentThreadCpuTime} else 0Lval ser = SparkEnv.get.closureSerializer.newInstance()// 反序列化rdd 及其 作用于RDD的結果函數val (rdd, func) = ser.deserialize[(RDD[T], (TaskContext, Iterator[T]) => U)](ByteBuffer.wrap(taskBinary.value), Thread.currentThread.getContextClassLoader)// 計算 反序列化費時_executorDeserializeTime = System.currentTimeMillis() - deserializeStartTime// 計算 反序列化Cpu費時_executorDeserializeCpuTime = if (threadMXBean.isCurrentThreadCpuTimeSupported) {threadMXBean.getCurrentThreadCpuTime - deserializeStartCpuTime} else 0L// 這部分 我們會在后續博文講解func(context, rdd.iterator(partition, context))}

后續處理

計量統計

對各個費時的統計，上章已經講解。

回收內存

這在上章Task.run也已經講解。

處理執行結果

Executor.TaskRunner.run的execBackend.statusUpdate在《深入理解Spark 2.1 Core （四）：運算結果處理和容錯的原理與源碼分析 》中我們已經講解過。

總結

以上是生活随笔為你收集整理的深入理解Spark 2.1 Core （七）：Standalone模式任务执行的原理与源码分析的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。