YARN源码分析(一)-----ApplicationMaster

前言

在之前两周主要学了HDFS中的一些模块知识,其中的许多都或多或少有我们借鉴学习的地方,现在将目光转向另外一个块,被誉为MRv2,就是yarn,在Yarn中,解决了MR中JobTracker单点的问题,将此拆分成了ResourceManager和NodeManager这样的结构,在每个节点上,还会有ApplicationMaster来管理应用程序的整个生命周期,的确在Yarn中,多了许多优秀的设计,而今天,我主要分享的就是这个ApplicationMaster相关的一整套服务,他是隶属于ResoureManager的内部服务中的.了解了AM的启动机制,你将会更进一步了解Yarn的任务启动过程.


ApplicationMaster管理涉及类

ApplicationMaster管理涉及到了4大类,ApplicationMasterLauncher,AMLivelinessMonitor,ApplicationMasterService,以及ApplicationMaster自身类.下面介绍一下这些类的用途,在Yarn中,每个类都会有自己明确的功能模块的区分.

1.ApplicationMasterLauncher--姑且叫做AM启动关闭事件处理器,他既是一个服务也是一个处理器,在这个类中,只处理2类事件,launch和cleanup事件.分别对应启动应用和关闭应用的情形.

2.AMLivelinessMonitor--这个类从名字上可以看出他是监控类,监控的对象是AM存活状态的监控类,检测的方法与之前的HDFS一样,都是采用heartbeat的方式,如果有节点过期了,将会触发一次过期事件.

3.ApplicationMasterService--AM请求服务处理类.AMS存在于ResourceManager,中,服务的对象是各个节点上的ApplicationMaster,负责接收各个AM的注册请求,更新心跳包信息等.

4.ApplicationMaster--节点应用管理类,简单的说,ApplicationMaster负责管理整个应用的生命周期.

简答的描述完AM管理的相关类,下面从源码级别分析一下几个流程.


AM启动

要想让AM启动,启动的背景当然是有用户提交了新的Application的时候,之后ApplicationMasterLauncher会生成Launch事件,与对应的nodemanager通信,让其准备启动的新的AM的Container.在这里,就用到了ApplicationMasterLauncher这个类,之前在上文中已经提到,此类就处理2类事件,Launch启动和Cleanup清洗事件,先来看看这个类的基本变量设置

//Application应用事件处理器
public class ApplicationMasterLauncher extends AbstractService implements
    EventHandler<AMLauncherEvent> {
  private static final Log LOG = LogFactory.getLog(
      ApplicationMasterLauncher.class);
  private final ThreadPoolExecutor launcherPool;
  private LauncherThread launcherHandlingThread;
  
  //事件队列
  private final BlockingQueue<Runnable> masterEvents
    = new LinkedBlockingQueue<Runnable>();
  //资源管理器上下文
  protected final RMContext context;
  
  public ApplicationMasterLauncher(RMContext context) {
    super(ApplicationMasterLauncher.class.getName());
    this.context = context;
    //初始化线程池
    this.launcherPool = new ThreadPoolExecutor(10, 10, 1, 
        TimeUnit.HOURS, new LinkedBlockingQueue<Runnable>());
    //新建处理线程
    this.launcherHandlingThread = new LauncherThread();
  }

还算比较简单,有一个masterEvents事件队列,还有执行线程以及所需的线程池执行环境。在RM相关的服务中,基本都是继承自AbstractService这个抽象服务类的。ApplicationMasterLauncher中主要处理2类事件,就是下面的展示的

@Override
  public synchronized void  handle(AMLauncherEvent appEvent) {
    AMLauncherEventType event = appEvent.getType();
    RMAppAttempt application = appEvent.getAppAttempt();
    //处理来自ApplicationMaster获取到的请求,分为启动事件和清洗事件2种
    switch (event) {
    case LAUNCH:
      launch(application);
      break;
    case CLEANUP:
      cleanup(application);
    default:
      break;
    }
  }
然后调用具体的实现方法,以启动事件launch事件为例

//添加应用启动事件
  private void launch(RMAppAttempt application) {
    Runnable launcher = createRunnableLauncher(application, 
        AMLauncherEventType.LAUNCH);
    //将启动事件加入事件队列中
    masterEvents.add(launcher);
  }
这些事件被加入到事件队列之后,是如何被处理的呢,通过消息队列的形式,在一个独立的线程中逐一被执行

//执行线程实现
  private class LauncherThread extends Thread {
    
    public LauncherThread() {
      super("ApplicationMaster Launcher");
    }

    @Override
    public void run() {
      while (!this.isInterrupted()) {
        Runnable toLaunch;
        try {
          //执行方法为从事件队列中逐一取出事件
          toLaunch = masterEvents.take();
          //放入线程池池中进行执行
          launcherPool.execute(toLaunch);
        } catch (InterruptedException e) {
          LOG.warn(this.getClass().getName() + " interrupted. Returning.");
          return;
        }
      }
    }
  }
如果论到事件的具体执行方式,就要看具体AMLauch是如何执行的,AMLauch本身就是一个runnable实例。

/**
 * The launch of the AM itself.
 * Application事件执行器
 */
public class AMLauncher implements Runnable {

  private static final Log LOG = LogFactory.getLog(AMLauncher.class);

  private ContainerManagementProtocol containerMgrProxy;

  private final RMAppAttempt application;
  private final Configuration conf;
  private final AMLauncherEventType eventType;
  private final RMContext rmContext;
  private final Container masterContainer;
在里面主要的run方法如下,就是按照事件类型进行区分操作

@SuppressWarnings("unchecked")
  public void run() {
  	//AMLauncher分2中事件分别处理
    switch (eventType) {
    case LAUNCH:
      try {
        LOG.info("Launching master" + application.getAppAttemptId());
        //调用启动方法
        launch();
        handler.handle(new RMAppAttemptEvent(application.getAppAttemptId(),
            RMAppAttemptEventType.LAUNCHED));
      ...
      break;
    case CLEANUP:
      try {
        LOG.info("Cleaning master " + application.getAppAttemptId());
        //调用作业清洗方法
        cleanup();
      ...
      break;
    default:
      LOG.warn("Received unknown event-type " + eventType + ". Ignoring.");
      break;
    }
  }
后面的launch操作会调用RPC函数与远程的NodeManager通信来启动Container。然后到了ApplicationMaster的run()启动方法,在启动方法中,会进行应用注册的方法,
@SuppressWarnings({ "unchecked" })
  public boolean run() throws YarnException, IOException {
    LOG.info("Starting ApplicationMaster");

    Credentials credentials =
        UserGroupInformation.getCurrentUser().getCredentials();
    DataOutputBuffer dob = new DataOutputBuffer();
    credentials.writeTokenStorageToStream(dob);
    // Now remove the AM->RM token so that containers cannot access it.
    Iterator<Token<?>> iter = credentials.getAllTokens().iterator();
    while (iter.hasNext()) {
      Token<?> token = iter.next();
      if (token.getKind().equals(AMRMTokenIdentifier.KIND_NAME)) {
        iter.remove();
      }
    }
    allTokens = ByteBuffer.wrap(dob.getData(), 0, dob.getLength());

    //与ResourceManager通信,周期性发送心跳信息,包含了应用的最新信息
    AMRMClientAsync.CallbackHandler allocListener = new RMCallbackHandler();
    amRMClient = AMRMClientAsync.createAMRMClientAsync(1000, allocListener);
    amRMClient.init(conf);
    amRMClient.start();
    .....

    // Register self with ResourceManager
    // This will start heartbeating to the RM
    //启动之后进行AM的注册
    appMasterHostname = NetUtils.getHostname();
    RegisterApplicationMasterResponse response = amRMClient
        .registerApplicationMaster(appMasterHostname, appMasterRpcPort,
            appMasterTrackingUrl);
    // Dump out information about cluster capability as seen by the
    // resource manager
    int maxMem = response.getMaximumResourceCapability().getMemory();
    LOG.info("Max mem capabililty of resources in this cluster " + maxMem);

    // A resource ask cannot exceed the max.
    if (containerMemory > maxMem) {
      LOG.info("Container memory specified above max threshold of cluster."
          + " Using max value." + ", specified=" + containerMemory + ", max="
          + maxMem);
      containerMemory = maxMem;
    }
在这个操作中,会将自己注册到AMLivelinessMonitor中,此刻开始启动心跳监控。


AMLiveLinessMonitor监控

在这里把重心从ApplicationMaster转移到AMLivelinessMonitor上,首先这是一个激活状态的监控线程,此类线程都有一个共同的父类

//应用存活状态监控线程
public class AMLivelinessMonitor extends AbstractLivelinessMonitor<ApplicationAttemptId> {
在AbstractlinessMonitor中定义监控类线程的一类特征和方法

//进程存活状态监控类
public abstract class AbstractLivelinessMonitor<O> extends AbstractService {

  private static final Log LOG = LogFactory.getLog(AbstractLivelinessMonitor.class);

  //thread which runs periodically to see the last time since a heartbeat is
  //received.
  //检查线程
  private Thread checkerThread;
  private volatile boolean stopped;
  //默认超时时间5分钟
  public static final int DEFAULT_EXPIRE = 5*60*1000;//5 mins
  //超时时间
  private int expireInterval = DEFAULT_EXPIRE;
  //监控间隔检测时间,为超时时间的1/3
  private int monitorInterval = expireInterval/3;

  private final Clock clock;
  
  //保存了心跳检验的结果记录
  private Map<O, Long> running = new HashMap<O, Long>();
心跳检测本身非常的简单,做一次通信记录检查,然后更新一下,记录时间,当一个新的节点加入监控或解除监控操作

//新的节点注册心跳监控
  public synchronized void register(O ob) {
    running.put(ob, clock.getTime());
  }
  
  //节点移除心跳监控
  public synchronized void unregister(O ob) {
    running.remove(ob);
  }
每次做心跳周期检测的时候,调用下述方法

//更新心跳监控检测最新时间
  public synchronized void receivedPing(O ob) {
    //only put for the registered objects
    if (running.containsKey(ob)) {
      running.put(ob, clock.getTime());
    }
  }
非常简单的更新方法,O ob对象在这里因场景而异,在AM监控中,为ApplicationID应用ID。在后面的AMS和AM的交互中会看到。新的应用加入AMLivelinessMonitor监控中后,后面的主要操作就是AMS与AM之间的交互操作了。


AM与AMS

在ApplicationMaster运行之后,会周期性的向ApplicationMasterService发送心跳信息,心跳信息包含有许多资源描述信息。

//ApplicationMaster心跳信息更新
  @Override
  public AllocateResponse allocate(AllocateRequest request)
      throws YarnException, IOException {

    ApplicationAttemptId appAttemptId = authorizeRequest();
    //进行心跳信息时间的更新
    this.amLivelinessMonitor.receivedPing(appAttemptId);
    ....
每次心跳信息一来,就会更新最新监控时间。在AMS也有对应的注册应用的方法

  
  //ApplicationMaster在ApplicationMasterService上服务上进行应用注册
  @Override
  public RegisterApplicationMasterResponse registerApplicationMaster(
      RegisterApplicationMasterRequest request) throws YarnException,
      IOException {

    ApplicationAttemptId applicationAttemptId = authorizeRequest();

    ApplicationId appID = applicationAttemptId.getApplicationId();
    .....
      
      //在存活监控线程上进行心跳记录,更新检测时间,key为应用ID
      this.amLivelinessMonitor.receivedPing(applicationAttemptId);
      RMApp app = this.rmContext.getRMApps().get(appID);
      
      // Setting the response id to 0 to identify if the
      // application master is register for the respective attemptid
      lastResponse.setResponseId(0);
      responseMap.put(applicationAttemptId, lastResponse);
      LOG.info("AM registration " + applicationAttemptId);
      this.rmContext
如果在心跳监控中出现过期的现象,就会触发一个expire事件,在AMLiveLinessMonitor中,这部分的工作是交给CheckThread执行的

//进程存活状态监控类
public abstract class AbstractLivelinessMonitor<O> extends AbstractService {
  ...
  //thread which runs periodically to see the last time since a heartbeat is
  //received.
  //检查线程
  private Thread checkerThread;
  ....
  //默认超时时间5分钟
  public static final int DEFAULT_EXPIRE = 5*60*1000;//5 mins
  //超时时间
  private int expireInterval = DEFAULT_EXPIRE;
  //监控间隔检测时间,为超时时间的1/3
  private int monitorInterval = expireInterval/3;
  ....
  //保存了心跳检验的结果记录
  private Map<O, Long> running = new HashMap<O, Long>();
  ...

  private class PingChecker implements Runnable {

    @Override
    public void run() {
      while (!stopped && !Thread.currentThread().isInterrupted()) {
        synchronized (AbstractLivelinessMonitor.this) {
          Iterator<Map.Entry<O, Long>> iterator = 
            running.entrySet().iterator();

          //avoid calculating current time everytime in loop
          long currentTime = clock.getTime();

          while (iterator.hasNext()) {
            Map.Entry<O, Long> entry = iterator.next();
            //进行超时检测
            if (currentTime > entry.getValue() + expireInterval) {
              iterator.remove();
              //调用超时处理方法,将处理事件交由调度器处理
              expire(entry.getKey());
              LOG.info("Expired:" + entry.getKey().toString() + 
                      " Timed out after " + expireInterval/1000 + " secs");
            }
          }
        }
check线程主要做的事件就是遍历每个节点的最新心跳更新时间,通过计算差值进行判断是否过期,过期调用expire方法。此方法由其子类实现

//应用存活状态监控线程
public class AMLivelinessMonitor extends AbstractLivelinessMonitor<ApplicationAttemptId> {
  //中央调度处理器
  private EventHandler dispatcher;
  ...

  @Override
  protected void expire(ApplicationAttemptId id) {
  	 //一旦应用过期,处理器处理过期事件处理
    dispatcher.handle(
        new RMAppAttemptEvent(id, RMAppAttemptEventType.EXPIRE));
  }
}
产生应用超期事件,然后发给中央调度器去处理。之所以采用的这样的方式,是因为在RM中,所有的模块设计是以事件驱动的形式工作,最大程度的保证了各个模块间的解耦。不同模块通过不同的事件转变为不同的状态,可以理解为状态机的改变。最后用一张书中的截图简单的展示AM模块相关的调用过程。




全部代码的分析请点击链接https://github.com/linyiqun/hadoop-yarn,后续将会继续更新YARN其他方面的代码分析。


参考文献

《Hadoop技术内部–HDFS结构设计与实现原理》.蔡斌等

posted @ 2020-01-12 19:09  回眸,境界  阅读(317)  评论(0编辑  收藏  举报