TaskScheduler的启动
《深入理解Spark:核心思想与源码分析》一书前言的内容请看链接《深入理解SPARK:核心思想与源码分析》一书正式出版上市
《深入理解Spark:核心思想与源码分析》一书第一章的内容请看链接《第1章 环境准备》
《深入理解Spark:核心思想与源码分析》一书第二章的内容请看链接《第2章 SPARK设计理念与基本架构》
由于本书的第3章内容较多,所以打算分别开辟四篇随笔分别展现。
《深入理解Spark:核心思想与源码分析》一书第三章第一部分的内容请看链接《深入理解Spark:核心思想与源码分析》——SparkContext的初始化(伯篇)》
《深入理解Spark:核心思想与源码分析》一书第三章第二部分的内容请看链接《深入理解Spark:核心思想与源码分析》——SparkContext的初始化(仲篇)》
本文展现第3章第三部分的内容:
3.8 TaskScheduler的启动
3.7节介绍了任务调度器TaskScheduler的创建,要想TaskScheduler发挥作用,必须要启动它,代码如下。
taskScheduler.start()
TaskScheduler在启动的时候,实际调用了backend的start方法。
override def start() { backend.start() }
以LocalBackend为例,启动LocalBackend时向actorSystem注册了LocalActor,见代码清单3-30所示(在《深入理解Spark:核心思想与源码分析》——SparkContext的初始化(中)》一文)。
3.8.1 创建LocalActor
创建LocalActor的过程主要是构建本地的Executor,见代码清单3-36。
代码清单3-36 LocalActor的实现
private[spark] class LocalActor(scheduler: TaskSchedulerImpl, executorBackend: LocalBackend, private val totalCores: Int) extends Actor with ActorLogReceive with Logging { import context.dispatcher // to use Akka's scheduler.scheduleOnce() private var freeCores = totalCores private val localExecutorId = SparkContext.DRIVER_IDENTIFIER private val localExecutorHostname = "localhost" val executor = new Executor( localExecutorId, localExecutorHostname, scheduler.conf.getAll, totalCores, isLocal = true) override def receiveWithLogging = { case ReviveOffers => reviveOffers() case StatusUpdate(taskId, state, serializedData) => scheduler.statusUpdate(taskId, state, serializedData) if (TaskState.isFinished(state)) { freeCores += scheduler.CPUS_PER_TASK reviveOffers() } case KillTask(taskId, interruptThread) => executor.killTask(taskId, interruptThread) case StopExecutor => executor.stop() } }
Executor的构建,见代码清单3-37,主要包括以下步骤:
1) 创建并注册ExecutorSource。ExecutorSource是做什么的呢?笔者将在3.10.2节详细介绍。
2) 获取SparkEnv。如果是非local模式,Worker上的CoarseGrainedExecutorBackend向Driver上的CoarseGrainedExecutorBackend注册Executor时,则需要新建SparkEnv。可以修改属性spark.executor.port(默认为0,表示随机生成)来配置Executor中的ActorSystem的端口号。
3) 创建并注册ExecutorActor。ExecutorActor负责接受发送给Executor的消息。
4) urlClassLoader的创建。为什么需要创建这个ClassLoader?在非local模式中,Driver或者Worker上都会有多个Executor,每个Executor都设置自身的urlClassLoader,用于加载任务上传的jar包中的类,有效对任务的类加载环境进行隔离。
5) 创建Executor执行TaskRunner任务(TaskRunner将在5.5节介绍)的线程池。此线程池是通过调用Utils.newDaemonCachedThreadPool创建的,具体实现请参阅附录A。
6) 启动Executor的心跳线程。此线程用于向Driver发送心跳。
此外,还包括Akka发送消息的帧大小(10485760字节)、结果总大小的字节限制(1073741824字节)、正在运行的task的列表、设置serializer的默认ClassLoader为创建的ClassLoader等。
代码清单3-37 Executor的构建
val executorSource = new ExecutorSource(this, executorId) private val env = { if (!isLocal) { val port = conf.getInt("spark.executor.port", 0) val _env = SparkEnv.createExecutorEnv( conf, executorId, executorHostname, port, numCores, isLocal, actorSystem) SparkEnv.set(_env) _env.metricsSystem.registerSource(executorSource) _env.blockManager.initialize(conf.getAppId) _env } else { SparkEnv.get } } private val executorActor = env.actorSystem.actorOf( Props(new ExecutorActor(executorId)), "ExecutorActor") private val urlClassLoader = createClassLoader() private val replClassLoader = addReplClassLoaderIfNeeded(urlClassLoader) env.serializer.setDefaultClassLoader(urlClassLoader) private val akkaFrameSize = AkkaUtils.maxFrameSizeBytes(conf) private val maxResultSize = Utils.getMaxResultSize(conf) val threadPool = Utils.newDaemonCachedThreadPool("Executor task launch worker") private val runningTasks = new ConcurrentHashMap[Long, TaskRunner] startDriverHeartbeater()
3.8.2 ExecutorSource的创建与注册
ExecutorSource用于测量系统。通过metricRegistry的register方法注册计量,这些计量信息包括threadpool.activeTasks、threadpool.completeTasks、threadpool.currentPool_size、threadpool.maxPool_size、filesystem.hdfs.write_bytes、filesystem.hdfs.read_ops、filesystem.file.write_bytes、filesystem.hdfs.largeRead_ops、filesystem.hdfs.write_ops等,ExecutorSource的实现见代码清单3-38。Metric接口的具体实现,参考附录D。
代码清单3-38 ExecutorSource的实现
private[spark] class ExecutorSource(val executor: Executor, executorId: String) extends Source { private def fileStats(scheme: String) : Option[FileSystem.Statistics] = FileSystem.getAllStatistics().filter(s => s.getScheme.equals(scheme)).headOption private def registerFileSystemStat[T]( scheme: String, name: String, f: FileSystem.Statistics => T, defaultValue: T) = { metricRegistry.register(MetricRegistry.name("filesystem", scheme, name), new Gauge[T] { override def getValue: T = fileStats(scheme).map(f).getOrElse(defaultValue) }) } override val metricRegistry = new MetricRegistry() override val sourceName = "executor" metricRegistry.register(MetricRegistry.name("threadpool", "activeTasks"), new Gauge[Int] { override def getValue: Int = executor.threadPool.getActiveCount() }) metricRegistry.register(MetricRegistry.name("threadpool", "completeTasks"), new Gauge[Long] { override def getValue: Long = executor.threadPool.getCompletedTaskCount() }) metricRegistry.register(MetricRegistry.name("threadpool", "currentPool_size"), new Gauge[Int] { override def getValue: Int = executor.threadPool.getPoolSize() }) metricRegistry.register(MetricRegistry.name("threadpool", "maxPool_size"), new Gauge[Int] { override def getValue: Int = executor.threadPool.getMaximumPoolSize() }) // Gauge for file system stats of this executor for (scheme <- Array("hdfs", "file")) { registerFileSystemStat(scheme, "read_bytes", _.getBytesRead(), 0L) registerFileSystemStat(scheme, "write_bytes", _.getBytesWritten(), 0L) registerFileSystemStat(scheme, "read_ops", _.getReadOps(), 0) registerFileSystemStat(scheme, "largeRead_ops", _.getLargeReadOps(), 0) registerFileSystemStat(scheme, "write_ops", _.getWriteOps(), 0) } }
创建完ExecutorSource后,调用MetricsSystem的registerSource方法将ExecutorSource注册到MetricsSystem。registerSource方法使用MetricRegistry的register方法,将Source注册到MetricRegistry,见代码清单3-39。关于MetricRegistry,具体参阅附录D。
代码清单3-39 MetricsSystem注册Source的实现
def registerSource(source: Source) { sources += source try { val regName = buildRegistryName(source) registry.register(regName, source.metricRegistry) } catch { case e: IllegalArgumentException => logInfo("Metrics already registered", e) } }
3.8.3 ExecutorActor的构建与注册
ExecutorActor很简单,当接收到SparkUI发来的消息时,将所有线程的栈信息发送回去,代码实现如下。
override def receiveWithLogging = { case TriggerThreadDump => sender ! Utils.getThreadDump() }
3.8.4 Spark自身ClassLoader的创建
获取要创建的ClassLoader的父加载器currentLoader,然后根据currentJars生成URL数组,spark.files.userClassPathFirst属性指定加载类时是否先从用户的classpath下加载,最后创建ExecutorURLClassLoader或者ChildExecutorURLClassLoader,见代码清单3-40。
代码清单3-40 Spark自身ClassLoader的创建
private def createClassLoader(): MutableURLClassLoader = { val currentLoader = Utils.getContextOrSparkClassLoader val urls = currentJars.keySet.map { uri => new File(uri.split("/").last).toURI.toURL }.toArray val userClassPathFirst = conf.getBoolean("spark.files.userClassPathFirst", false) userClassPathFirst match { case true => new ChildExecutorURLClassLoader(urls, currentLoader) case false => new ExecutorURLClassLoader(urls, currentLoader) } }
Utils.getContextOrSparkClassLoader的实现见附录A。ExecutorURLClassLoader或者ChildExecutorURLClassLoader实际上都继承了URLClassLoader,见代码清单3-41。
代码清单3-41 ChildExecutorURLClassLoader与ExecutorURLClassLoader的实现
private[spark] class ChildExecutorURLClassLoader(urls: Array[URL], parent: ClassLoader) extends MutableURLClassLoader { private object userClassLoader extends URLClassLoader(urls, null){ override def addURL(url: URL) { super.addURL(url) } override def findClass(name: String): Class[_] = { super.findClass(name) } } private val parentClassLoader = new ParentClassLoader(parent) override def findClass(name: String): Class[_] = { try { userClassLoader.findClass(name) } catch { case e: ClassNotFoundException => { parentClassLoader.loadClass(name) } } } def addURL(url: URL) { userClassLoader.addURL(url) } def getURLs() = { userClassLoader.getURLs() } } private[spark] class ExecutorURLClassLoader(urls: Array[URL], parent: ClassLoader) extends URLClassLoader(urls, parent) with MutableURLClassLoader { override def addURL(url: URL) { super.addURL(url) } }
如果需要REPL交互,还会调用addReplClassLoaderIfNeeded创建replClassLoader,见代码清单3-42。
代码清单3-42 addReplClassLoaderIfNeeded的实现
private def addReplClassLoaderIfNeeded(parent: ClassLoader): ClassLoader = { val classUri = conf.get("spark.repl.class.uri", null) if (classUri != null) { logInfo("Using REPL class URI: " + classUri) val userClassPathFirst: java.lang.Boolean = conf.getBoolean("spark.files.userClassPathFirst", false) try { val klass = Class.forName("org.apache.spark.repl.ExecutorClassLoader") .asInstanceOf[Class[_ <: ClassLoader]] val constructor = klass.getConstructor(classOf[SparkConf], classOf[String], classOf[ClassLoader], classOf[Boolean]) constructor.newInstance(conf, classUri, parent, userClassPathFirst) } catch { case _: ClassNotFoundException => logError("Could not find org.apache.spark.repl.ExecutorClassLoader on classpath!") System.exit(1) null } } else { parent } }
3.8.5 启动Executor的心跳线程
Executor的心跳由startDriverHeartbeater启动,见代码清单3-43。Executor心跳线程的间隔由属性spark.executor.heartbeatInterval配置,默认是10000毫秒。此外,超时时间是30秒,超时重试次数是3次,重试间隔是3000毫秒,使用actorSystem.actorSelection (url)方法查找到匹配的Actor引用, url是akka.tcp://sparkDriver@ $driverHost:$driverPort/user/HeartbeatReceiver,最终创建一个运行过程中,每次会休眠10000到20000毫秒的线程。此线程从runningTasks获取最新的有关Task的测量信息,将其与executorId、blockManagerId封装为Heartbeat消息,向HeartbeatReceiver发送Heartbeat消息。
代码清单3-43 启动Executor的心跳线程
def startDriverHeartbeater() { val interval = conf.getInt("spark.executor.heartbeatInterval", 10000) val timeout = AkkaUtils.lookupTimeout(conf) val retryAttempts = AkkaUtils.numRetries(conf) val retryIntervalMs = AkkaUtils.retryWaitMs(conf) val heartbeatReceiverRef = AkkaUtils.makeDriverRef("HeartbeatReceiver", conf,env.actorSystem) val t = new Thread() { override def run() { // Sleep a random interval so the heartbeats don't end up in sync Thread.sleep(interval + (math.random * interval).asInstanceOf[Int]) while (!isStopped) { val tasksMetrics = new ArrayBuffer[(Long, TaskMetrics)]() val curGCTime = gcTime for (taskRunner <- runningTasks.values()) { if (!taskRunner.attemptedTask.isEmpty) { Option(taskRunner.task).flatMap(_.metrics).foreach { metrics => metrics.updateShuffleReadMetrics metrics.jvmGCTime = curGCTime - taskRunner.startGCTime if (isLocal) { val copiedMetrics = Utils.deserialize[TaskMetrics](Utils.serialize(metrics)) tasksMetrics += ((taskRunner.taskId, copiedMetrics)) } else { // It will be copied by serialization tasksMetrics += ((taskRunner.taskId, metrics)) } } } } val message = Heartbeat(executorId, tasksMetrics.toArray, env.blockManager.blockManagerId) try { val response = AkkaUtils.askWithReply[HeartbeatResponse](message, heartbeatReceiverRef, retryAttempts, retryIntervalMs, timeout) if (response.reregisterBlockManager) { logWarning("Told to re-register on heartbeat") env.blockManager.reregister() } } catch { case NonFatal(t) => logWarning("Issue communicating with driver in heartbeater", t) } Thread.sleep(interval) } } } t.setDaemon(true) t.setName("Driver Heartbeater") t.start() }
这个心跳线程的作用是什么呢?其作用有两个:
q 更新正在处理的任务的测量信息;
q 通知BlockManagerMaster,此Executor上的BlockManager依然活着。
下面对心跳线程的实现详细分析下,读者可以自行选择是否需要阅读。
初始化TaskSchedulerImpl后会创建心跳接收器HeartbeatReceiver。HeartbeatReceiver接受所有分配给当前Driver Application的Executor的心跳,并将Task、Task计量信息、心跳等交给TaskSchedulerImpl和DAGScheduler作进一步处理。创建心跳接收器的代码如下。
private val heartbeatReceiver = env.actorSystem.actorOf( Props(new HeartbeatReceiver(taskScheduler)), "HeartbeatReceiver")
HeartbeatReceiver在收到心跳消息后,会调用TaskScheduler的executorHeartbeatReceived方法,代码如下。
override def receiveWithLogging = { case Heartbeat(executorId, taskMetrics, blockManagerId) => val response = HeartbeatResponse( !scheduler.executorHeartbeatReceived(executorId, taskMetrics, blockManagerId)) sender ! response }
executorHeartbeatReceived的实现代码如下。
val metricsWithStageIds: Array[(Long, Int, Int, TaskMetrics)] = synchronized { taskMetrics.flatMap { case (id, metrics) => taskIdToTaskSetId.get(id) .flatMap(activeTaskSets.get) .map(taskSetMgr => (id, taskSetMgr.stageId, taskSetMgr.taskSet.attempt, metrics)) } } dagScheduler.executorHeartbeatReceived(execId, metricsWithStageIds, blockManagerId)
这段程序通过遍历taskMetrics,依据taskIdToTaskSetId和activeTaskSets找到TaskSetManager。然后将taskId、TaskSetManager.stageId、TaskSetManager .taskSet.attempt、TaskMetrics封装到Array[(Long, Int, Int, TaskMetrics)]的数组metricsWithStageIds中。最后调用了dagScheduler的executorHeartbeatReceived方法,其实现如下。
listenerBus.post(SparkListenerExecutorMetricsUpdate(execId, taskMetrics)) implicit val timeout = Timeout(600 seconds) Await.result( blockManagerMaster.driverActor ? BlockManagerHeartbeat(blockManagerId), timeout.duration).asInstanceOf[Boolean]
dagScheduler将executorId、metricsWithStageIds封装为SparkListenerExecutorMetricsUpdate事件,并post到listenerBus中,此事件用于更新Stage的各种测量数据。最后给BlockManagerMaster持有的BlockManagerMasterActor发送BlockManagerHeartbeat消息。BlockManagerMasterActor在收到消息后会匹配执行heartbeatReceived方法(会在4.3.1节介绍)。heartbeatReceived最终更新BlockManagerMaster对BlockManager最后可见时间(即更新BlockManagerId对应的BlockManagerInfo的_lastSeenMs,见代码清单3-44)。
代码清单3-44 BlockManagerMasterActor的心跳处理
private def heartbeatReceived(blockManagerId: BlockManagerId): Boolean = { if (!blockManagerInfo.contains(blockManagerId)) { blockManagerId.isDriver && !isLocal } else { blockManagerInfo(blockManagerId).updateLastSeenMs() true } }
local模式下Executor的心跳通信过程,可以用图3-3来表示。
图3-3 Executor的心跳通信过程
注意:在非local模式中Executor发送心跳的过程是一样的,主要的区别是Executor进程与Driver不在同一个进程,甚至不在同一个节点上。
接下来会初始化块管理器BlockManager,代码如下。
env.blockManager.initialize(applicationId)
具体的初始化过程,请参阅第4章。
未完待续。。。
后记:自己牺牲了7个月的周末和下班空闲时间,通过研究Spark源码和原理,总结整理的《深入理解Spark:核心思想与源码分析》一书现在已经正式出版上市,目前亚马逊、京东、当当、天猫等网站均有销售,欢迎感兴趣的同学购买。我开始研究源码时的Spark版本是1.2.0,经过7个多月的研究和出版社近4个月的流程,Spark自身的版本迭代也很快,如今最新已经是1.6.0。目前市面上另外2本源码研究的Spark书籍的版本分别是0.9.0版本和1.2.0版本,看来这些书的作者都与我一样,遇到了这种问题。由于研究和出版都需要时间,所以不能及时跟上Spark的脚步,还请大家见谅。但是Spark核心部分的变化相对还是很少的,如果对版本不是过于追求,依然可以选择本书。
京东(现有满100减30活动):http://item.jd.com/11846120.html