spark 源码分析之四 -- TaskScheduler的创建和启动过程

在 spark 源码分析之二 -- SparkContext 的初始化过程 中，第 14 步 和 16 步分别描述了 TaskScheduler的 初始化 和 启动过程。

话分两头，先说 TaskScheduler的初始化过程

TaskScheduler的实例化

 val (sched, ts) = SparkContext.createTaskScheduler(this, master, deployMode)

其调用了org.apache.spark.SparkContext#createTaskScheduler ， 源码如下：

 /**
    * Create a task scheduler based on a given master URL.
    * Return a 2-tuple of the scheduler backend and the task scheduler.
    */
   private def createTaskScheduler(
       sc: SparkContext,
       master: String,
       deployMode: String): (SchedulerBackend, TaskScheduler) = {
     import SparkMasterRegex._

     // When running locally, don't try to re-execute tasks on failure.
     val MAX_LOCAL_TASK_FAILURES = 1

     master match {
       case "local" =>
         val scheduler = new TaskSchedulerImpl(sc, MAX_LOCAL_TASK_FAILURES, isLocal = true)
         val backend = new LocalSchedulerBackend(sc.getConf, scheduler, 1)
         scheduler.initialize(backend)
         (backend, scheduler)

       case LOCAL_N_REGEX(threads) =>
         def localCpuCount: Int = Runtime.getRuntime.availableProcessors()
         // local[*] estimates the number of cores on the machine; local[N] uses exactly N threads.
         val threadCount = if (threads == "*") localCpuCount else threads.toInt
         if (threadCount <= 0) {
           throw new SparkException(s"Asked to run locally with $threadCount threads")
         }
         val scheduler = new TaskSchedulerImpl(sc, MAX_LOCAL_TASK_FAILURES, isLocal = true)
         val backend = new LocalSchedulerBackend(sc.getConf, scheduler, threadCount)
         scheduler.initialize(backend)
         (backend, scheduler)

       case LOCAL_N_FAILURES_REGEX(threads, maxFailures) =>
         def localCpuCount: Int = Runtime.getRuntime.availableProcessors()
         // local[*, M] means the number of cores on the computer with M failures
         // local[N, M] means exactly N threads with M failures
         val threadCount = if (threads == "*") localCpuCount else threads.toInt
         val scheduler = new TaskSchedulerImpl(sc, maxFailures.toInt, isLocal = true)
         val backend = new LocalSchedulerBackend(sc.getConf, scheduler, threadCount)
         scheduler.initialize(backend)
         (backend, scheduler)

       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)

       case LOCAL_CLUSTER_REGEX(numSlaves, coresPerSlave, memoryPerSlave) =>
         // Check to make sure memory requested <= memoryPerSlave. Otherwise Spark will just hang.
         val memoryPerSlaveInt = memoryPerSlave.toInt
         if (sc.executorMemory > memoryPerSlaveInt) {
           throw new SparkException(
             "Asked to launch cluster with %d MB RAM / worker but requested %d MB/worker".format(
               memoryPerSlaveInt, sc.executorMemory))
         }

         val scheduler = new TaskSchedulerImpl(sc)
         val localCluster = new LocalSparkCluster(
           numSlaves.toInt, coresPerSlave.toInt, memoryPerSlaveInt, sc.conf)
         val masterUrls = localCluster.start()
         val backend = new StandaloneSchedulerBackend(scheduler, sc, masterUrls)
         scheduler.initialize(backend)
         backend.shutdownCallback = (backend: StandaloneSchedulerBackend) => {
           localCluster.stop()
         }
         (backend, scheduler)

       case masterUrl =>
         val cm = getClusterManager(masterUrl) match {
           case Some(clusterMgr) => clusterMgr
           case None => throw new SparkException("Could not parse Master URL: '" + master + "'")
         }
         try {
           val scheduler = cm.createTaskScheduler(sc, masterUrl)
           val backend = cm.createSchedulerBackend(sc, masterUrl, scheduler)
           cm.initialize(scheduler, backend)
           (backend, scheduler)
         } catch {
           case se: SparkException => throw se
           case NonFatal(e) =>
             throw new SparkException("External scheduler cannot be instantiated", e)
         }
     }
   }

不同的实现如下：

　　

实例化部分剖析完毕，下半部分重点剖析yarn-client mode 下 TaskScheduler 的启动过程

yarn-client模式TaskScheduler 启动过程

初始化调度池

yarn-client 模式下，TaskScheduler的实现是 org.apache.spark.scheduler.cluster.YarnScheduler, TaskSchedulerBackend的实现是org.apache.spark.scheduler.cluster.YarnClientSchedulerBackend

在org.apache.spark.SparkContext#createTaskScheduler 方法中，有如下调用：

 case masterUrl =>
         val cm = getClusterManager(masterUrl) match {
           case Some(clusterMgr) => clusterMgr
           case None => throw new SparkException("Could not parse Master URL: '" + master + "'")
         }
         try {
           val scheduler = cm.createTaskScheduler(sc, masterUrl)
           val backend = cm.createSchedulerBackend(sc, masterUrl, scheduler)
           cm.initialize(scheduler, backend)
           (backend, scheduler)
         } catch {
           case se: SparkException => throw se
           case NonFatal(e) =>
             throw new SparkException("External scheduler cannot be instantiated", e)
         }

其中的，cm.initialize(scheduler, backend)中的cm 是org.apache.spark.scheduler.cluster.YarnClusterManager，TaskScheduler的实现是 org.apache.spark.scheduler.cluster.YarnScheduler, TaskSchedulerBackend的实现是org.apache.spark.scheduler.cluster.YarnClientSchedulerBackend。YarnClusterManager 的 initialize 方法实现如下：

   override def initialize(scheduler: TaskScheduler, backend: SchedulerBackend): Unit = {
     scheduler.asInstanceOf[TaskSchedulerImpl].initialize(backend)
   }

其并没有实现 initialize， 父类TaskSchedulerImpl 的实现如下：

 def initialize(backend: SchedulerBackend) {
     this.backend = backend
     schedulableBuilder = {
       schedulingMode match {
         case SchedulingMode.FIFO =>
           new FIFOSchedulableBuilder(rootPool)
         case SchedulingMode.FAIR =>
           new FairSchedulableBuilder(rootPool, conf)
         case _ =>
           throw new IllegalArgumentException(s"Unsupported $SCHEDULER_MODE_PROPERTY: " +
           s"$schedulingMode")
       }
     }
     schedulableBuilder.buildPools()
   }

可以看出，其重要作用就是设置 TaskScheduler 的 TaskSchedulerBackend 引用。

调度模式主要有FIFO和FAIR两种模式。默认是FIFO模式，可以使用spark.scheduler.mode 参数来设定。使用建造者模式来创建 Pool 对象。

其中，org.apache.spark.scheduler.FIFOSchedulableBuilder#buildPools是一个空实现，即没有做任何的操作；而 org.apache.spark.scheduler.FairSchedulableBuilder#buildPools会加载 相应调度分配策略文件；策略文件可以使用 spark.scheduler.allocation.file 参数来设定，如果没有设定会进一步加载默认的 fairscheduler.xml 文件，如果还没有，则不加载。如果有调度池的配置，则根据配置配置调度pool并将其加入到 root 池中。最后初始化 default 池并将其加入到 root 池中。

在HeartBeatReceiver 中设定 taskscheduler 变量

 _heartbeatReceiver.ask[Boolean](TaskSchedulerIsSet)

首先，_heartbeatReceiver 是一个 RpcEndPointRef 对象，其请求最终会被 HeartbeatReceiver（Endpoint）接收并处理。即org.apache.spark.HeartbeatReceiver#receiveAndReply方法：

　　

 case TaskSchedulerIsSet =>
       scheduler = sc.taskScheduler
       context.reply(true)

具体的关于RPC的相关解释，会在后面有专门的文章篇幅介绍。在这里就不做过多解释。 // TODO

启动TaskScheduler

org.apache.spark.SparkContext 的初始化方法有如下代码启动 TaskScheduler：

 _taskScheduler.start()

yarn-client模式下，运行中调用了 org.apache.spark.scheduler.cluster.YarnScheduler 的 start 方法，它沿用了父类 TaskSchedulerImpl 的实现：

 override def start() {
     // 1. 启动 task scheduler backend
     backend.start()
     // 2. 设定 speculationScheduler 定时任务
     if (!isLocal && conf.getBoolean("spark.speculation", false)) {
       logInfo("Starting speculative execution thread")
       speculationScheduler.scheduleWithFixedDelay(new Runnable {
         override def run(): Unit = Utils.tryOrStopSparkContext(sc) {
           checkSpeculatableTasks()
         }
       }, SPECULATION_INTERVAL_MS, SPECULATION_INTERVAL_MS, TimeUnit.MILLISECONDS)
     }
   }

第1步：task scheduler backend 的启动：org.apache.spark.scheduler.cluster.YarnClientSchedulerBackend#start的方法如下：

 /**
    * Create a Yarn client to submit an application to the ResourceManager.
    * This waits until the application is running.
    */
   override def start() {
     // 1. 获取driver 的 host 和 port
     val driverHost = conf.get("spark.driver.host")
     val driverPort = conf.get("spark.driver.port")
     val hostport = driverHost + ":" + driverPort
     // 2. 设定 driver 的 web UI 地址
     sc.ui.foreach { ui => conf.set("spark.driver.appUIAddress", ui.webUrl) }

     val argsArrayBuf = new ArrayBuffer[String]()
     argsArrayBuf += ("--arg", hostport)

     logDebug("ClientArguments called with: " + argsArrayBuf.mkString(" "))
     val args = new ClientArguments(argsArrayBuf.toArray)
     totalExpectedExecutors = SchedulerBackendUtils.getInitialTargetExecutorNumber(conf)
     // 3. 启动 deploy client，并切初始化 driverClient 的 Rpc environment，并在该RPC 环境中初始化master 和 driver 的rpc endpoint
     client = new Client(args, conf)
     // 4. 将 application id 绑定到 yarn 上
     bindToYarn(client.submitApplication(), None)

     // SPARK-8687: Ensure all necessary properties have already been set before
     // we initialize our driver scheduler backend, which serves these properties
     // to the executors
     super.start()
    // 5. 检查 yarn application的状态，不能为 kill， finished等等
     waitForApplication()
    // 6. 监控线程
     monitorThread = asyncMonitorApplication()
     monitorThread.start()
   }

重点解释一下第三步，涉及的源码步如下：

 object Client {
   def main(args: Array[String]) {
     // scalastyle:off println
     if (!sys.props.contains("SPARK_SUBMIT")) {
       println("WARNING: This client is deprecated and will be removed in a future version of Spark")
       println("Use ./bin/spark-submit with \"--master spark://host:port\"")
     }
     // scalastyle:on println
     new ClientApp().start(args, new SparkConf())
   }
 }

 private[spark] class ClientApp extends SparkApplication {

   override def start(args: Array[String], conf: SparkConf): Unit = {
     val driverArgs = new ClientArguments(args)

     if (!conf.contains("spark.rpc.askTimeout")) {
       conf.set("spark.rpc.askTimeout", "10s")
     }
     Logger.getRootLogger.setLevel(driverArgs.logLevel)

     val rpcEnv =
       RpcEnv.create("driverClient", Utils.localHostName(), 0, conf, new SecurityManager(conf))

     val masterEndpoints = driverArgs.masters.map(RpcAddress.fromSparkURL).
       map(rpcEnv.setupEndpointRef(_, Master.ENDPOINT_NAME))
     rpcEnv.setupEndpoint("client", new ClientEndpoint(rpcEnv, driverArgs, masterEndpoints, conf))

     rpcEnv.awaitTermination()
   }

 }

可以看到，在Client 的main方法中，初始化了ClientApp 对象，并调用了其 start 方法，在start 方法中， 首先解析了 driver的 参数。然后创建了 driver 端的 RPC environment，然后 根据解析的 master 的信息，初始化 master 的endpointref，并且建立了 client endpoint 并返回 client endpoint ref。

定时执行推测任务

下面继续看 org.apache.spark.scheduler.cluster.YarnScheduler 的 start 方法 的 第二步方法，首先 spark 推测任务 feature 默认是关闭的，原因如果有很多任务都延迟了，那么它会再启动一个相同的任务，这样可能会消耗掉所有的资源，对集群资源和提交到集群上的任务造成不可控的影响。启动了一个延迟定时器，定时地执行 checkSpeculatableTasks 方法，如下：

 // Check for speculatable tasks in all our active jobs.
   def checkSpeculatableTasks() {
     var shouldRevive = false
     synchronized {
       shouldRevive = rootPool.checkSpeculatableTasks(MIN_TIME_TO_SPECULATION) // 1. 推测是否应该跑一个新任务
     }
     if (shouldRevive) {
       backend.reviveOffers() // 2. 跑一个新任务
     }
   }

其中，第一步推断任务，有两个实现一个是Pool 的实现，一个是TaskSetManager 的实现，Pool 会递归调用子Pool来获取 speculatable tasks。如果需要推测，则运行task scheduler backend 的 reviveOffers方法，大致思路如下，首先获取 executor 上的空闲资源，然后将这些资源分配给 推测的 task，供其使用。

总结，本篇源码剖析了在Spark Context 启动过程中， 以 yarn-client 模式为例，剖析了task scheduler 是如何启动的。

其中关于RpcEnv的介绍直接略过了，下一篇会专门讲解Spark 中内置的Rpc 机制的整体架构以及其是如何运行的。