最近集群遇到一个问题,就是集群在跑任务的时候,AM会超时10min而被KILL,但任务重跑则成功,问题是随机的出现的,所以初步怀疑是因为AM心跳汇报出现问题或则RM因为繁忙hang住,AM因为某些机制导致等待10min不汇报心跳,所以我们还是先了解,AM是如何向RM汇报心跳的。

在MRAppMaster中,ContainerAllocatorRouter负责向RM申请资源(发送心跳)

RMAM

RMContainerAllocator其最终父类是RMCommunicator,它实现了RMHeartbeatHandler接口

public interface RMHeartbeatHandler {
  long getLastHeartbeatTime(); // 获取上一次心跳的时间
  void runOnNextHeartbeat(Runnable callback); // 回调注册到callback队列的callback函数
}

每一次心跳回来,都会执行一次注册在heartbeatCallbacks中的回调函数:

allocatorThread = new Thread(new Runnable() {
      @Override
      public void run() {
        while (!stopped.get() && !Thread.currentThread().isInterrupted()) {
              ......
              heartbeat();            
              lastHeartbeatTime = context.getClock().getTime();// 记录上一次心跳时间
              executeHeartbeatCallbacks(); // 执行回调函数
              ....
});

RMCommunicator类中:

private void executeHeartbeatCallbacks() {
    Runnable callback = null;
    while ((callback = heartbeatCallbacks.poll()) != null) {
      callback.run();
    }
  }

在RMCommunicator启动时,首先会向RM注册,把自己的host和port告诉RM,然后在启动一条线程(startAllocatorThread)定期的调用RMContainerAllocator中实现的heartbeat方法(向RM申请资源,定期汇报信息,告诉RM自己还活着)。

AM初始化同时也会初始化RMCommunicator:

protected void serviceStart() throws Exception {
  scheduler= createSchedulerProxy(); // 获取RM的代理
  register(); // 注册
  startAllocatorThread(); // 心跳线程
....
}

AM的ContainerAllocatorRouter事件处理流程如下图:

RMALLO

注册流程:

调用RMCommunicator远程调用ApplicationMasterService的registerApplicationMaster方法,设置维护responseId,然后把它加入AMLivelinessMonitor中,并使用map记录时间,用来监控AM是否因为长时间没有心跳而超时,如果AM长时间没有心跳信息更新,RM就会通知NodeManager把AM移除。

心跳线程:

在发送心跳的过程中,即也是获取资源的过程

@Override
  protected synchronized void heartbeat() throws Exception {
    scheduleStats.updateAndLogIfChanged("Before Scheduling: ");
    List<Container> allocatedContainers = getResources();// 重要的方法
    if (allocatedContainers.size() > 0) {
      scheduledRequests.assign(allocatedContainers);
    }
   ......
  }

获取资源的过程:

private List<Container> getResources() throws Exception {
     ...
     response = makeRemoteRequest(); // 和RM进行交互
     ...
     // 优先处理RM发送过来的命令
     if (response.getAMCommand() != null) {
         switch(response.getAMCommand()) {
                case AM_RESYNC:
                case AM_SHUTDOWN:
                     eventHandler.handle(new JobEvent(this.getJob().getID(),
                                     JobEventType.JOB_AM_REBOOT));
                     throw new YarnRuntimeException("Resource Manager doesn't recognize AttemptId: " +
                             this.getContext().getApplicationID());
                default:
                     ....
      }
     // 等等一系列处理
}
}

构建请求:

protected AllocateResponse makeRemoteRequest() throws IOException {
    AllocateRequest allocateRequest =
        AllocateRequest.newInstance(lastResponseID,
          super.getApplicationProgress(), new ArrayList<ResourceRequest>(ask),
          new ArrayList<ContainerId>(release), blacklistRequest);
    AllocateResponse allocateResponse;
    allocateResponse = scheduler.allocate(allocateRequest); // RPC调用ApplicationMasterService的allocate方法
    .....
}

每一次心跳的调用都会刷新AMLivelinessMonitor的时间,代表AM还活着

而且我们通过代码可以看出,资源请求被封装为一个ask,即一个ResourceRequest的ArrayList的资源列表 例如:

priority:20 host:host9 capability:<memory:2048, vCores:1>
priority:20 host:host2 capability:<memory:2048, vCores:1>
priority:20 host:host10 capability:<memory:2048, vCores:1>
priority:20 host:/rack/rack3203 capability:<memory:2048, vCores:1>
priority:20 host:/rack/rack3202 capability:<memory:2048, vCores:1>
priority:20 host:* capability:<memory:2048, vCores:1>

然而,ask是如何被构造的呢?

RMContainerAllocator中的addMap,addReduce,assign方法中对ask的数据内容进行了修改

addContainerReq --> addResourceRequest --> addResourceRequestToAsk;

通过在代码自己添加日志可以看出,资源会被分为local,rack,和any级别去申请资源

最终变为一个ask list发送到RM上:

 ask Capability:<memory:2048, vCores:1> ResourceName:* NumContainers:384 Priority:20 RelaxLocality:true
 ask Capability:<memory:2048, vCores:1> ResourceName:/rack/rack3201 NumContainers:227 Priority:20 RelaxLocality:true
 ask Capability:<memory:2048, vCores:1> ResourceName:/rack/rack3202 NumContainers:231 Priority:20 RelaxLocality:true
 ask Capability:<memory:2048, vCores:1> ResourceName:/rack/rack3203 NumContainers:152 Priority:20 RelaxLocality:true
 ask Capability:<memory:2048, vCores:1> ResourceName:/rack/rack3204 NumContainers:158 Priority:20 RelaxLocality:true
 ask Capability:<memory:2048, vCores:1> ResourceName:host1 NumContainers:46 Priority:20 RelaxLocality:true
 ask Capability:<memory:2048, vCores:1> ResourceName:host5 NumContainers:52 Priority:20 RelaxLocality:true
 ask Capability:<memory:2048, vCores:1> ResourceName:host6 NumContainers:38 Priority:20 RelaxLocality:true

类似日志为:

getResources() for application_1438330253091_0004: ask=29 release= 0 newContainers=0 finishedContainers=0 resourcelimit=<memory:0, vCores:0> knownNMs=24

总结:

除了了解心跳之外,还学习了许多Map和Reduce的分配机制,收获良多。