本篇內(nèi)容介紹了“Spark Driver啟動流程是怎樣的”的有關(guān)知識，在實(shí)際案例的操作過程中，不少人都會遇到這樣的困境，接下來就讓小編帶領(lǐng)大家學(xué)習(xí)一下如何處理這些情況吧！希望大家仔細(xì)閱讀，能夠?qū)W有所成！

創(chuàng)新互聯(lián)主要從事成都網(wǎng)站制作、成都網(wǎng)站建設(shè)、網(wǎng)頁設(shè)計(jì)、企業(yè)做網(wǎng)站、公司建網(wǎng)站等業(yè)務(wù)。立足成都服務(wù)祁門,十余年網(wǎng)站建設(shè)經(jīng)驗(yàn),價(jià)格優(yōu)惠、服務(wù)專業(yè),歡迎來電咨詢建站服務(wù):18982081108

SparkContext.scala

val (sched, ts) = SparkContext.createTaskScheduler(this, master, deployMode)
    _schedulerBackend = sched
    _taskScheduler = ts
    _dagScheduler = new DAGScheduler(this)
    _heartbeatReceiver.ask[Boolean](TaskSchedulerIsSet)

    // start TaskScheduler after taskScheduler sets DAGScheduler reference in DAGScheduler's
    // constructor
    _taskScheduler.start()

我們再看一下 SparkContext.createTaskScheduler 當(dāng)中究竟做了些什么

      case SPARK_REGEX(sparkUrl) =>
        val scheduler = new TaskSchedulerImpl(sc)
        val masterUrls = sparkUrl.split(",").map("spark://" + _)
        val backend = new StandaloneSchedulerBackend(scheduler, sc, masterUrls)
        scheduler.initialize(backend)
        (backend, scheduler)

我們看到 _taskScheduler 是 TaskSchedulerImpl 的實(shí)例， _schedulerBackend 是 StandaloneSchedulerBackend 的實(shí)例，而會把 _schedulerBackend 通過 scheduler.initialize 給到 _taskScheduler。

然后再來看一下 _taskScheduler.start() 究竟干了些什么

  override def start() {
    backend.start()

    if (!isLocal && conf.getBoolean("spark.speculation", false)) {
      logInfo("Starting speculative execution thread")
      speculationScheduler.scheduleAtFixedRate(new Runnable {
        override def run(): Unit = Utils.tryOrStopSparkContext(sc) {
          checkSpeculatableTasks()
        }
      }, SPECULATION_INTERVAL_MS, SPECULATION_INTERVAL_MS, TimeUnit.MILLISECONDS)
    }
  }

我們看到首先是對 backend.start() 的調(diào)用，我們可以在 StandaloneSchedulerBackend 當(dāng)中找到start的實(shí)現(xiàn)：

override def start() {
    super.start()
    launcherBackend.connect()

    // The endpoint for executors to talk to us
    val driverUrl = RpcEndpointAddress(
      sc.conf.get("spark.driver.host"),
      sc.conf.get("spark.driver.port").toInt,
      CoarseGrainedSchedulerBackend.ENDPOINT_NAME).toString
    val args = Seq(
      "--driver-url", driverUrl,
      "--executor-id", "{{EXECUTOR_ID}}",
      "--hostname", "{{HOSTNAME}}",
      "--cores", "{{CORES}}",
      "--app-id", "{{APP_ID}}",
      "--worker-url", "{{WORKER_URL}}")
    val extraJavaOpts = sc.conf.getOption("spark.executor.extraJavaOptions")
      .map(Utils.splitCommandString).getOrElse(Seq.empty)
    val classPathEntries = sc.conf.getOption("spark.executor.extraClassPath")
      .map(_.split(java.io.File.pathSeparator).toSeq).getOrElse(Nil)
    val libraryPathEntries = sc.conf.getOption("spark.executor.extraLibraryPath")
      .map(_.split(java.io.File.pathSeparator).toSeq).getOrElse(Nil)

    // When testing, expose the parent class path to the child. This is processed by
    // compute-classpath.{cmd,sh} and makes all needed jars available to child processes
    // when the assembly is built with the "*-provided" profiles enabled.
    val testingClassPath =
      if (sys.props.contains("spark.testing")) {
        sys.props("java.class.path").split(java.io.File.pathSeparator).toSeq
      } else {
        Nil
      }

    // Start executors with a few necessary configs for registering with the scheduler
    val sparkJavaOpts = Utils.sparkJavaOpts(conf, SparkConf.isExecutorStartupConf)
    val javaOpts = sparkJavaOpts ++ extraJavaOpts
    val command = Command("org.apache.spark.executor.CoarseGrainedExecutorBackend",
      args, sc.executorEnvs, classPathEntries ++ testingClassPath, libraryPathEntries, javaOpts)
    val appUIAddress = sc.ui.map(_.appUIAddress).getOrElse("")
    val coresPerExecutor = conf.getOption("spark.executor.cores").map(_.toInt)
    // If we're using dynamic allocation, set our initial executor limit to 0 for now.
    // ExecutorAllocationManager will send the real initial limit to the Master later.
    val initialExecutorLimit =
      if (Utils.isDynamicAllocationEnabled(conf)) {
        Some(0)
      } else {
        None
      }
    val appDesc = new ApplicationDescription(sc.appName, maxCores, sc.executorMemory, command,
      appUIAddress, sc.eventLogDir, sc.eventLogCodec, coresPerExecutor, initialExecutorLimit)
    client = new StandaloneAppClient(sc.env.rpcEnv, masters, appDesc, this, conf)
    client.start()
    launcherBackend.setState(SparkAppHandle.State.SUBMITTED)
    waitForRegistration()
    launcherBackend.setState(SparkAppHandle.State.RUNNING)
  }

我們看一下client.start()當(dāng)中究竟了做了些什么：

  def start() {
    // Just launch an rpcEndpoint; it will call back into the listener.
    endpoint.set(rpcEnv.setupEndpoint("AppClient", new ClientEndpoint(rpcEnv)))
  }

endpoint是一個(gè)AtomicReference， rpcEnv.setupEndpoint 做了2件事，一個(gè)是注冊一個(gè)endpoint，另外把它的ref返回回來。這里哪里體現(xiàn)start了？我知道一定會進(jìn)到 ClientEndpoint 的start方法當(dāng)中去，可是究竟是怎么進(jìn)去的？？？？

下面這段代碼我們在Rpc機(jī)制的文章當(dāng)中提到過，紅色代碼部分，當(dāng)時(shí)并沒有太在意，現(xiàn)在看來，每個(gè)End Point注冊到Rpc Env當(dāng)中的時(shí)候，都會自動觸發(fā)它的start事件。

  def registerRpcEndpoint(name: String, endpoint: RpcEndpoint): NettyRpcEndpointRef = {
    val addr = RpcEndpointAddress(nettyEnv.address, name)
    val endpointRef = new NettyRpcEndpointRef(nettyEnv.conf, addr, nettyEnv)
    synchronized {
      if (stopped) {
        throw new IllegalStateException("RpcEnv has been stopped")
      }
      if (endpoints.putIfAbsent(name, new EndpointData(name, endpoint, endpointRef)) != null) {
        throw new IllegalArgumentException(s"There is already an RpcEndpoint called $name")
      }
      val data = endpoints.get(name)
      endpointRefs.put(data.endpoint, data.ref)
      receivers.offer(data)  // for the OnStart message
    }
    endpointRef
  }

之后，讓我們找到ClientEndpoint（ StandaloneAppClient的一個(gè)內(nèi)部類 ），看它的onStart方法：

    override def onStart(): Unit = {
      try {
        registerWithMaster(1)
      } catch {
        case e: Exception =>
          logWarning("Failed to connect to master", e)
          markDisconnected()
          stop()
      }
    }

registerWithMaster 當(dāng)中也存在遞歸調(diào)用，不過這個(gè)遞歸，是為了retry服務(wù)的，所以我們直接看 tryRegisterAllMasters() 。

    private def registerWithMaster(nthRetry: Int) {
      registerMasterFutures.set(tryRegisterAllMasters())
      registrationRetryTimer.set(registrationRetryThread.schedule(new Runnable {
        override def run(): Unit = {
          if (registered.get) {
            registerMasterFutures.get.foreach(_.cancel(true))
            registerMasterThreadPool.shutdownNow()
          } else if (nthRetry >= REGISTRATION_RETRIES) {
            markDead("All masters are unresponsive! Giving up.")
          } else {
            registerMasterFutures.get.foreach(_.cancel(true))
            registerWithMaster(nthRetry + 1)
          }
        }
      }, REGISTRATION_TIMEOUT_SECONDS, TimeUnit.SECONDS))
    }

    private def tryRegisterAllMasters(): Array[JFuture[_]] = {
      for (masterAddress <- masterRpcAddresses) yield {
        registerMasterThreadPool.submit(new Runnable {
          override def run(): Unit = try {
            if (registered.get) {
              return
            }
            logInfo("Connecting to master " + masterAddress.toSparkURL + "...")
            val masterRef = rpcEnv.setupEndpointRef(masterAddress, Master.ENDPOINT_NAME)
            masterRef.send(RegisterApplication(appDescription, self))
          } catch {
            case ie: InterruptedException => // Cancelled
            case NonFatal(e) => logWarning(s"Failed to connect to master $masterAddress", e)
          }
        })
      }
    }

紅色2行代碼，注冊End Point，并發(fā)送消息。這里的master end point，應(yīng)該是一個(gè)位于spark集群，master節(jié)點(diǎn)上的end point，相對于driver上的Rpc Env來講，應(yīng)該是一個(gè)remote的end point。

我們找到master.scala，先看它的類聲明：

private[deploy] class Master(
    override val rpcEnv: RpcEnv,
    address: RpcAddress,
    webUiPort: Int,
    val securityMgr: SecurityManager,
    val conf: SparkConf)
  extends ThreadSafeRpcEndpoint with Logging with LeaderElectable {

再找到它的receive方法：

override def receive: PartialFunction[Any, Unit]

只需要看其中一段：

    case RegisterApplication(description, driver) =>
      // TODO Prevent repeated registrations from some driver
      if (state == RecoveryState.STANDBY) {
        // ignore, don't send response
      } else {
        logInfo("Registering app " + description.name)
        val app = createApplication(description, driver)
        registerApplication(app)
        logInfo("Registered app " + description.name + " with ID " + app.id)
        persistenceEngine.addApplication(app)
        driver.send(RegisteredApplication(app.id, self))
        schedule()
      }

1、創(chuàng)建app，2、注冊app，3、持久化app，4、向driver的endpoint發(fā)送消息，5、schedule()

step4， 其中driver是跟著Rpc Message一起過來的，需要給driver發(fā)一個(gè)注冊app的響應(yīng)。

我們再回到 ClientEndpoint.receive，

    override def receive: PartialFunction[Any, Unit] = {
      case RegisteredApplication(appId_, masterRef) =>
        // FIXME How to handle the following cases?
        // 1. A master receives multiple registrations and sends back multiple
        // RegisteredApplications due to an unstable network.
        // 2. Receive multiple RegisteredApplication from different masters because the master is
        // changing.
        appId.set(appId_)
        registered.set(true)
        master = Some(masterRef)
        listener.connected(appId.get)

step5，我們看看當(dāng)中做了些什么事情：

  private def schedule(): Unit = {
    if (state != RecoveryState.ALIVE) {
      return
    }
    // Drivers take strict precedence over executors
    val shuffledAliveWorkers = Random.shuffle(workers.toSeq.filter(_.state == WorkerState.ALIVE))
    val numWorkersAlive = shuffledAliveWorkers.size
    var curPos = 0
    for (driver <- waitingDrivers.toList) { // iterate over a copy of waitingDrivers
      // We assign workers to each waiting driver in a round-robin fashion. For each driver, we
      // start from the last worker that was assigned a driver, and continue onwards until we have
      // explored all alive workers.
      var launched = false
      var numWorkersVisited = 0
      while (numWorkersVisited < numWorkersAlive && !launched) {
        val worker = shuffledAliveWorkers(curPos)
        numWorkersVisited += 1
        if (worker.memoryFree >= driver.desc.mem && worker.coresFree >= driver.desc.cores) {
          launchDriver(worker, driver)
          waitingDrivers -= driver
          launched = true
        }
        curPos = (curPos + 1) % numWorkersAlive
      }
    }
    startExecutorsOnWorkers()
  }

launchDriver(worker, driver) 我們理解為，在worder上為當(dāng)前的driver啟動一個(gè)線程。

再看一下 startExecutorsOnWorkers() ：

  private def startExecutorsOnWorkers(): Unit = {
    // Right now this is a very simple FIFO scheduler. We keep trying to fit in the first app
    // in the queue, then the second app, etc.
    for (app <- waitingApps if app.coresLeft > 0) {
      val coresPerExecutor: Option[Int] = app.desc.coresPerExecutor
      // Filter out workers that don't have enough resources to launch an executor
      val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE)
        .filter(worker => worker.memoryFree >= app.desc.memoryPerExecutorMB &&
          worker.coresFree >= coresPerExecutor.getOrElse(1))
        .sortBy(_.coresFree).reverse
      val assignedCores = scheduleExecutorsOnWorkers(app, usableWorkers, spreadOutApps)

      // Now that we've decided how many cores to allocate on each worker, let's allocate them
      for (pos <- 0 until usableWorkers.length if assignedCores(pos) > 0) {
        allocateWorkerResourceToExecutors(
          app, assignedCores(pos), coresPerExecutor, usableWorkers(pos))
      }
    }
  }

  private def allocateWorkerResourceToExecutors(
      app: ApplicationInfo,
      assignedCores: Int,
      coresPerExecutor: Option[Int],
      worker: WorkerInfo): Unit = {
    // If the number of cores per executor is specified, we divide the cores assigned
    // to this worker evenly among the executors with no remainder.
    // Otherwise, we launch a single executor that grabs all the assignedCores on this worker.
    val numExecutors = coresPerExecutor.map { assignedCores / _ }.getOrElse(1)
    val coresToAssign = coresPerExecutor.getOrElse(assignedCores)
    for (i <- 1 to numExecutors) {
      val exec = app.addExecutor(worker, coresToAssign)
      launchExecutor(worker, exec)
      app.state = ApplicationState.RUNNING
    }
  }

  private def launchExecutor(worker: WorkerInfo, exec: ExecutorDesc): Unit = {
    logInfo("Launching executor " + exec.fullId + " on worker " + worker.id)
    worker.addExecutor(exec)
    worker.endpoint.send(LaunchExecutor(masterUrl,
      exec.application.id, exec.id, exec.application.desc, exec.cores, exec.memory))
    exec.application.driver.send(
      ExecutorAdded(exec.id, worker.id, worker.hostPort, exec.cores, exec.memory))
  }

1、在一個(gè)本地的worker變量當(dāng)中添加一個(gè)exec

2、通知worker，啟動一個(gè)executor

3、通知driver，executor added

“Spark Driver啟動流程是怎樣的”的內(nèi)容就介紹到這里了，感謝大家的閱讀。如果想了解更多行業(yè)相關(guān)的知識可以關(guān)注創(chuàng)新互聯(lián)網(wǎng)站，小編將為大家輸出更多高質(zhì)量的實(shí)用文章！