YARN源码分析之ApplicationMaster启动流程之RM端

任何一个计算框架或者说一个服务要运行在yarn上，都需要一个master来对job进行管理，这个master就是ApplicationMaster。

ApplicationMaster是一个job的大脑，下面就以MapReduce为例，介绍下ApplicationMaster的启动流程。

首先client向RM提交一个application请求，RM创建一个application，然后再创建一个appattempt，后期的调度和任务的拆解都是对这个appattempt进行的，当appattempt的状态从ALLOCATED_SAVING变成ALLOCATED时，由AttemptStoredTransition.transition调用appAttempt.launchAttempt()进行启动，下面来看下具体代码:

// RMAppAttemptImpl.java
private void launchAttempt(){
  // Send event to launch the AM Container
  // 通过异步调度器得到该事件注册的handle (在ResourceManager中注册)
  // AMLauncherEvent 对应的handle是ApplicationMasterLauncher
  eventHandler.handle(new AMLauncherEvent(AMLauncherEventType.LAUNCH, this));
}

AMLauncherEvent对应的handle是ApplicationMasterLauncher，事件类型是LAUNCH，在ApplicationMasterLauncher.handle中会调用launch(application)，代码如下:

private void launch(RMAppAttempt application) {
  // 创建一个线程
  Runnable launcher = createRunnableLauncher(application, 
      AMLauncherEventType.LAUNCH);
  // 将线程放入阻塞队列中
  masterEvents.add(launcher);
}

只从这个方法来分析，首先创建了一个launcher线程，然后将其放入一个队列中，等待另一个线程从队列中取出进行操作，这是典型的生产者消费者模型。那么我们就来看下ApplicationMasterLauncher(ApplicationMasterLauncher是一个事件也是一个服务)关于这块代码的具体实现。

先看下createRunnableLauncher，

protected Runnable createRunnableLauncher(RMAppAttempt application, 
    AMLauncherEventType event) {
  Runnable launcher =
      new AMLauncher(context, application, event, getConfig());
  return launcher;
}

这里只是new了一个AMLauncher，AMLauncher实现了Runnable接口，是执行AM操作的线程，只执行launch和cleanup。

launcher线程创建之后add到阻塞队列masterEvents中，那么必然会有另一个线程来队列中take launcher，这个线程是LauncherThread类型的launcherHandlingThread，launcherHandlingThread将launcher取出丢给线程池去执行，代码如下:

private class LauncherThread extends Thread {
  public LauncherThread() {
    super("ApplicationMaster Launcher");
  }
  @Override
  public void run() {
    while (!this.isInterrupted()) {
      Runnable toLaunch;
      try {
        // 从阻塞队列中取出 
        toLaunch = masterEvents.take();
        // 交给线程池执行
        // this.launcherPool = new ThreadPoolExecutor(10, 10, 1, TimeUnit.HOURS, new LinkedBlockingQueue<Runnable>());
        launcherPool.execute(toLaunch);
      } catch (InterruptedException e) {
        LOG.warn(this.getClass().getName() + " interrupted. Returning.");
        return;
      }
    }
  }
} 

放入线程池之后，launcher线程就开始执行，调用的是AMLauncher.run

public void run() {
  switch (eventType) {
  case LAUNCH:
    try {
      LOG.info("Launching master" + application.getAppAttemptId());
      launch();
      handler.handle(new RMAppAttemptEvent(application.getAppAttemptId(),
          RMAppAttemptEventType.LAUNCHED));
    } catch(Exception ie) {
      String message = "Error launching " + application.getAppAttemptId()
          + ". Got exception: " + StringUtils.stringifyException(ie);
      LOG.info(message);
      handler.handle(new RMAppAttemptLaunchFailedEvent(application
          .getAppAttemptId(), message));
    }
    break;
  case CLEANUP:
    ...
    break;
  default:
    LOG.warn("Received unknown event-type " + eventType + ". Ignoring.");
    break;
  }
}

之前放入阻塞队列masterEvents的事件类型是LAUNCH，则此处调用launch()方法:

private void launch() throws IOException, YarnException {
  // 得到对应node的rpc客户端
  connect();
  ContainerId masterContainerID = masterContainer.getId();
  ApplicationSubmissionContext applicationContext =
    application.getSubmissionContext();
  LOG.info("Setting up container " + masterContainer
      + " for AM " + application.getAppAttemptId());  
  ContainerLaunchContext launchContext =
      createAMContainerLaunchContext(applicationContext, masterContainerID);
  // 构建request
  StartContainerRequest scRequest =
      StartContainerRequest.newInstance(launchContext,
        masterContainer.getContainerToken());
  List<StartContainerRequest> list = new ArrayList<StartContainerRequest>();
  list.add(scRequest);
  StartContainersRequest allRequests =
      StartContainersRequest.newInstance(list);
  // 远程调用startContainers
  StartContainersResponse response =
      containerMgrProxy.startContainers(allRequests);
  if (response.getFailedRequests() != null
      && response.getFailedRequests().containsKey(masterContainerID)) {
    Throwable t =
        response.getFailedRequests().get(masterContainerID).deSerialize();
    parseAndThrowException(t);
  } else {
    LOG.info("Done launching container " + masterContainer + " for AM "
        + application.getAppAttemptId());
  }
}

AMLaunch.launch先在connect()中拿到对应node的rpc客户端containerMgrProxy，然后构造request，最后调用rpc函数startContainers()并返回response。

以上是RM端MRAppMaster的启动流程，下一篇将继续介绍nodemanager端端启动流程。