【Spark Core】任务执行机制和Task源码浅析1

引言

1. Task类及其相关

1.1 Task类

/**  * A unit of execution. We have two kinds of Task's in Spark:  * - [[org.apache.spark.scheduler.ShuffleMapTask]]  * - [[org.apache.spark.scheduler.ResultTask]]  *  * A Spark job consists of one or more stages. The very last stage in a job consists of multiple  * ResultTasks, while earlier stages consist of ShuffleMapTasks. A ResultTask executes the task  * and sends the task output back to the driver application. A ShuffleMapTask executes the task  * and divides the task output to multiple buckets (based on the task's partitioner).  *  * @paramstageId id of the stage this task belongs to  * @parampartitionId index of the number in the RDD  */ private[spark]abstractclassTask[T](val stageId: Int, var partitionId: Int)extendsSerializable 

除了上述基本接口和变量，Task的伴生对象提供了序列化和反序列化应用依赖的jar包的方法。原因是Task需要保证工作节点具备本次Task需要的其他依赖，注册到SparkContext下，所以提供了把依赖转成流写入写出的方法。

1.2 ShuffleMapTask

对应于ShuffleMap Stage, 产生的结果作为其他stage的输入。ShuffleMapTask复写了MapStatus向外读写的方法，因为向外读写的内容包括：stageId，rdd，dep，partitionId，epoch和split(某个partition)。对于其中的stageId，rdd，dep有统一的序列化和反序列化操作并会cache在内存里，再放到ObjectOutput里写出去。序列化操作使用的是Gzip，序列化信息会维护在serializedInfoCache = newHashMap[Int, Array[Byte]]。这部分需要序列化并保存的原因是：stageId，rdd，dep真正代表了本次Shuffle Task的信息，为了减轻master节点负担，把这部分序列化结果cache了起来。

/** * A ShuffleMapTask divides the elements of an RDD into multiple buckets (based on a partitioner * specified in the ShuffleDependency). * * See [[org.apache.spark.scheduler.Task]] for more information. *  * @paramstageId id of the stage this task belongs to  * @paramtaskBinary broadcast version of of the RDD and the ShuffleDependency. Once deserialized,  * the type should be (RDD[_], ShuffleDependency[_, _, _]).  * @parampartition partition of the RDD this task is associated with  * @paramlocs preferred task execution locations for locality scheduling  */ private[spark]classShuffleMapTask(  stageId: Int,  taskBinary: Broadcast[Array[Byte]],  partition: Partition, @transientprivate var locs: Seq[TaskLocation]) extendsTask[MapStatus](stageId, partition.index)withLogging { 

1.3 ResultTask

对应于Result Stage直接产生结果。

/**  * A task that sends back the output to the driver application.  *  * See [[Task]] for more information.  *  * @paramstageId id of the stage this task belongs to  * @paramtaskBinary broadcasted version of the serialized RDD and the function to apply on each  * partition of the given RDD. Once deserialized, the type should be  * (RDD[T], (TaskContext, Iterator[T]) => U).  * @parampartition partition of the RDD this task is associated with  * @paramlocs preferred task execution locations for locality scheduling  * @paramoutputId index of the task in this job (a job can launch tasks on only a subset of the  * input RDD's partitions).  */ private[spark]classResultTask[T,U](  stageId: Int,  taskBinary: Broadcast[Array[Byte]],  partition: Partition, @transientlocs: Seq[TaskLocation],  val outputId: Int) extendsTask[U](stageId, partition.index)withSerializable { 

1.4 TaskSet

TaskSet是一个数据结构，用于封装一个stage的所有的tasks, 以提交给TaskScheduler。TaskSet就是可以做pipeline的一组完全相同的task，每个task的处理逻辑完全相同，不同的是处理数据，每个task负责处理一个partition。pipeline，可以称为大数据处理的基石，只有数据进行pipeline处理，才能将其放到集群中去运行。对于一个task来说，它从数据源获得逻辑，然后按照拓扑顺序，顺序执行（实际上是调用rdd的compute）。

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

2. Executor注册到Driver

Driver发送LaunchTask消息被Executor接收，Executor会使用launchTask对消息进行处理。不过在这个过程之前，我们要知道，如果Executor没有注册到Driver，即便接收到LaunchTask指令，也不会做任务处理。所以我们要先搞清楚，Executor是如何在Driver侧进行注册的。

2.1 Application注册

Executor的注册是发生在Application的注册过程中的，我们以Standalone模式为例：

下面是SparkDeploySchedulerBackend的start函数中的部分注册Application的代码：

// Start executors with a few necessary configs for registering with the scheduler valsparkJavaOpts = Utils.sparkJavaOpts(conf, SparkConf.isExecutorStartupConf) valjavaOpts = sparkJavaOpts ++ extraJavaOpts valcommand = Command("org.apache.spark.executor.CoarseGrainedExecutorBackend",  args, sc.executorEnvs, classPathEntries ++ testingClassPath, libraryPathEntries, javaOpts) valappUIAddress = sc.ui.map(_.appUIAddress).getOrElse("") valappDesc =newApplicationDescription(sc.appName, maxCores, sc.executorMemory, command,  appUIAddress, sc.eventLogDir, sc.eventLogCodec)  client = newAppClient(sc.env.actorSystem, masters, appDesc,this, conf) client.start() 

下面是RegisterApplication(appDescription)消息的相关处理代码（在Master.scala中的receiveWithLogging部分代码）：

caseRegisterApplication(description) => { if(state == RecoveryState.STANDBY) { // ignore, don't send response  } else{  logInfo("Registering app "+ description.name) valapp = createApplication(description, sender)  registerApplication(app)  logInfo("Registered app "+ description.name +" with ID "+ app.id)  persistenceEngine.addApplication(app)  sender ! RegisteredApplication(app.id, masterUrl)  schedule()//为处于待分配资源的Application分配资源。在每次有新的Application加入或者新的资源加入时都会调用schedule进行调度  } } 

这段代码做了以下几件事：

2.2 Master中的schedule函数

schedule()为处于待分配资源的Application分配资源。在每次有新的Application加入或者新的资源加入时都会调用schedule进行调度。为Application分配资源选择worker（executor），现在有两种策略：

对于同一个Application，它在一个worker上只能拥有一个executor；当然了，这个executor可能拥有多于1个core。对于策略1，任务的部署会慢于策略2，但是GC的时间会更快。

schedule函数的源码，解释在中文注释中：

/*  * Schedule the currently available resources among waiting apps. This method will be called  * every time a new app joins or resource availability changes.  */ privatedefschedule() { if(state != RecoveryState.ALIVE) {return}  // First schedule drivers, they take strict precedence over applications // Randomization helps balance drivers valshuffledAliveWorkers = Random.shuffle(workers.toSeq.filter(_.state == WorkerState.ALIVE)) valnumWorkersAlive = shuffledAliveWorkers.size varcurPos =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. varlaunched =false varnumWorkersVisited =0 while(numWorkersVisited < numWorkersAlive && !launched) { valworker = 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  }  }  // 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. if(spreadOutApps) {//尽量的打散负载，如有可能，每个executor分配一个core // Try to spread out each app among all the nodes, until it has all its cores for(app <- waitingAppsifapp.coresLeft >0) { // 可用的worker的标准：State是Alive，其上并没有该Application的executor，可用内存满足要求。  // 在可用的worker中，优先选择可用core数多的。  valusableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE)  .filter(canUse(app, _)).sortBy(_.coresFree).reverse valnumUsable = usableWorkers.length valassigned =newArray[Int](numUsable)// Number of cores to give on each node vartoAssign = math.min(app.coresLeft, usableWorkers.map(_.coresFree).sum) varpos =0 while(toAssign >0) { if(usableWorkers(pos).coresFree - assigned(pos) >0) {  toAssign -= 1  assigned(pos) += 1  }  pos = (pos + 1) % numUsable  } // Now that we've decided how many cores to give on each node, let's actually give them for(pos <-0until numUsable) { if(assigned(pos) >0) { valexec = app.addExecutor(usableWorkers(pos), assigned(pos))  launchExecutor(usableWorkers(pos), exec)  app.state = ApplicationState.RUNNING  }  }  }  } else{//尽可能多的利用worker的core // Pack each app into as few nodes as possible until we've assigned all its cores for(worker <- workersifworker.coresFree >0&& worker.state == WorkerState.ALIVE) { for(app <- waitingAppsifapp.coresLeft >0) { if(canUse(app, worker)) { valcoresToUse = math.min(worker.coresFree, app.coresLeft) if(coresToUse >0) { valexec = app.addExecutor(worker, coresToUse)  launchExecutor(worker, exec)  app.state = ApplicationState.RUNNING  }  }  }  }  } } 

2.3 launchExecutor函数

在选择了worker和确定了worker上得executor需要的CPU core数后，Master会调用 launchExecutor(worker: WorkerInfo, exec: ExecutorInfo)向Worker发送请求，向AppClient发送executor已经添加的消息。同时会更新master保存的worker的信息，包括增加executor，减少可用的CPU core数和memory数。Master不会等到真正在worker上成功启动executor后再更新worker的信息。如果worker启动executor失败，那么它会发送FAILED的消息给Master，Master收到该消息时再次更新worker的信息即可。

deflaunchExecutor(worker: WorkerInfo, exec: ExecutorDesc) {  logInfo("Launching executor "+ exec.fullId +" on worker "+ worker.id)  worker.addExecutor(exec)  worker.actor ! LaunchExecutor(masterUrl,  exec.application.id, exec.id, exec.application.desc, exec.cores, exec.memory)  exec.application.driver ! ExecutorAdded(  exec.id, worker.id, worker.hostPort, exec.cores, exec.memory) } 

2.4 Executor的创建

CoarseGrainedExecutorBackend启动后，preStart函数执行的相关操作：

overridedefpreStart() {  logInfo("Connecting to driver: "+ driverUrl)  driver = context.actorSelection(driverUrl)  driver ! RegisterExecutor(executorId, hostPort, cores, extractLogUrls)  context.system.eventStream.subscribe(self, classOf[RemotingLifecycleEvent]) } 

CoarseGrainedExecutorBackend接收RegisteredExecutor消息后，创建Executor的操作：

overridedefreceiveWithLogging = { caseRegisteredExecutor =>  logInfo("Successfully registered with driver") val(hostname, _) = Utils.parseHostPort(hostPort)  executor = newExecutor(executorId, hostname, env, userClassPath, isLocal =false)   ...... 

参考资料