Master.preStart(){ webUi.bind() context.system.scheduler.schedule(0 millis, WORKER_TIMEOUT millis, self, CheckForWorkerTimeOut) //定时任务检测是否有DEAD WORKER需要移除 case CheckForWorkerTimeOut => { timeOutDeadWorkers() } /** Check for, and remove, any timed-out workers */ def timeOutDeadWorkers() { ... if (worker.lastHeartbeat < currentTime - ((REAPER_ITERATIONS + 1) * WORKER_TIMEOUT)) { workers -= worker } } }
Worker.preStart(){ override def preStart() { webUi = new WorkerWebUI(this, workDir, Some(webUiPort)) webUi.bind() registerWithMaster() //注册该Worker到Master } def tryRegisterAllMasters() { for (masterUrl <- masterUrls) { logInfo("Connecting to master " + masterUrl + "...") val actor = context.actorSelection(Master.toAkkaUrl(masterUrl)) actor ! RegisterWorker(workerId, host, port, cores, memory, webUi.boundPort, publicAddress) } } }
Master.scala case RegisterWorker(){ persistenceEngine.addWorker(worker) sender ! RegisteredWorker(masterUrl, masterWebUiUrl) //向Worker发送Worker注册成功事件 schedule() //调度部分后续章节分析 }
Worker.scala case RegisteredWorker(){ registered = true context.system.scheduler.schedule(0 millis, HEARTBEAT_MILLIS millis, self, SendHeartbeat) //Worker注册成功后，定时向Master发送心跳信息 } case SendHeartbeat => masterLock.synchronized { if (connected) { master ! Heartbeat(workerId) } }
Master.scala case Heartbeat(workerId) => { idToWorker.get(workerId) match { case Some(workerInfo) => workerInfo.lastHeartbeat = System.currentTimeMillis() //更新该worker的上次发送心跳信息的时间 case None => logWarning("Got heartbeat from unregistered worker " + workerId) } }
=================如上步骤完成了Worker到Master的连接===============================================
SparkContext启动时：
SparkContext.createTaskScheduler() ==>new SparkDeploySchedulerBackend() ==>创建AppClient并启动 ==>ClientActor.preStart():registerWithMaster(){actor ! RegisterApplication(appDescription)} //向Master发起RegisterApplication事件
Master.scala case RegisterApplication(description) { val app = createApplication(description, sender) registerApplication(app) persistenceEngine.addApplication(app) sender ! RegisteredApplication(app.id, masterUrl) //向Worker发起RegisteredApplication事件表示该Application已经注册成功 schedule() //调度部分后续章节分析
}
=======================如上步骤完成了Application到Master的连接===============================================
小结：
1、Master的主要功能：
1）Master Leader选举；
2）Master对Worker、Application等的管理(接收worker的注册并管理所有的worker，接收client提交的application，(FIFO)调度等待的application并向worker提交)；
2、Worker的主要功能：
1）通过RegisterWorker注册到Master；
2）定时发送心跳给Master；
3）根据master发送的application配置进程环境，并启动StandaloneExecutorBackend
3、运行spark-shell：
1）ClientActor通过RegisterApplication注册到Master；
2）Master收到RegisterApplication后，通过scheduler方法进行调度，如有满足要求的Worker，则发送LaunchExecutor给相应的Worker；