Uploaded image for project: 'Hadoop YARN'
  1. Hadoop YARN
  2. YARN-4398

Yarn recover functionality causes the cluster running slowly and the cluster usage rate is far below 100

    Details

    • Target Version/s:
    • Hadoop Flags:
      Reviewed

      Description

      In my hadoop cluster, the resourceManager recover functionality is enabled with FileSystemRMStateStore.
      I found this cause the yarn cluster running slowly and cluster usage rate is just 50 even there are many pending Apps.

      The scenario is below.
      In thread A, the RMAppImpl$RMAppNewlySavingTransition is calling storeNewApplication method defined in RMStateStore. This storeNewApplication method is synchronized.

      RMAppImpl.java
        private static final class RMAppNewlySavingTransition extends RMAppTransition {
          @Override
          public void transition(RMAppImpl app, RMAppEvent event) {
      
            // If recovery is enabled then store the application information in a
            // non-blocking call so make sure that RM has stored the information
            // needed to restart the AM after RM restart without further client
            // communication
            LOG.info("Storing application with id " + app.applicationId);
            app.rmContext.getStateStore().storeNewApplication(app);
          }
        }
      
      RMStateStore.java
      public synchronized void storeNewApplication(RMApp app) {
          ApplicationSubmissionContext context = app
                                                  .getApplicationSubmissionContext();
          assert context instanceof ApplicationSubmissionContextPBImpl;
          ApplicationStateData appState =
              ApplicationStateData.newInstance(
                  app.getSubmitTime(), app.getStartTime(), context, app.getUser());
          dispatcher.getEventHandler().handle(new RMStateStoreAppEvent(appState));
        }
      

      In thread B, the FileSystemRMStateStore is calling storeApplicationStateInternal method. It's also synchronized.
      This storeApplicationStateInternal method saves an ApplicationStateData into HDFS and it normally costs 90~300 milliseconds in my hadoop cluster.

      FileSystemRMStateStore.java
      public synchronized void storeApplicationStateInternal(ApplicationId appId,
            ApplicationStateData appStateDataPB) throws Exception {
          Path appDirPath = getAppDir(rmAppRoot, appId);
          mkdirsWithRetries(appDirPath);
          Path nodeCreatePath = getNodePath(appDirPath, appId.toString());
      
          LOG.info("Storing info for app: " + appId + " at: " + nodeCreatePath);
          byte[] appStateData = appStateDataPB.getProto().toByteArray();
          try {
            // currently throw all exceptions. May need to respond differently for HA
            // based on whether we have lost the right to write to FS
            writeFileWithRetries(nodeCreatePath, appStateData, true);
          } catch (Exception e) {
            LOG.info("Error storing info for app: " + appId, e);
            throw e;
          }
        }
      

      Think thread B firstly comes into FileSystemRMStateStore.storeApplicationStateInternal method, then thread A will be blocked for a while because of synchronization. In ResourceManager there is only one RMStateStore instance. In my cluster it's FileSystemRMStateStore type.
      Debug the RMAppNewlySavingTransition.transition method, the thread stack shows it's called form AsyncDispatcher.dispatch method. This method code is as below.

      AsyncDispatcher.java
        protected void dispatch(Event event) {
          //all events go thru this loop
          if (LOG.isDebugEnabled()) {
            LOG.debug("Dispatching the event " + event.getClass().getName() + "."
                + event.toString());
          }
      
          Class<? extends Enum> type = event.getType().getDeclaringClass();
      
          try{
            EventHandler handler = eventDispatchers.get(type);
            if(handler != null) {
              handler.handle(event);
            } else {
              throw new Exception("No handler for registered for " + type);
            }
          } catch (Throwable t) {
            //TODO Maybe log the state of the queue
            LOG.fatal("Error in dispatcher thread", t);
            // If serviceStop is called, we should exit this thread gracefully.
            if (exitOnDispatchException
                && (ShutdownHookManager.get().isShutdownInProgress()) == false
                && stopped == false) {
              Thread shutDownThread = new Thread(createShutDownThread());
              shutDownThread.setName("AsyncDispatcher ShutDown handler");
              shutDownThread.start();
            }
          }
        }
      

      Above code shows AsyncDispatcher.dispatch method can process different type events.
      In fact this AsyncDispatcher instance is just ResourceManager.rmDispatcher created in ResourceManager.serviceInit method.
      You can find many eventTypes and handlers are registered in ResourceManager.rmDispatcher.
      In above scenario thread B blocks thread A, then many following events processing are blocked.

      In my testing cluster, there is only one queue and the client submits 1000 applications concurrently, the yarn cluster usage rate is 50. Many apps are pending. If I disable resourceManager recover functionality, the cluster usage can be 100.

      To solve this issue, I removed synchronized modifier on some methods defined in RMStateStore.
      Instead, in these methods I defined a dedicated lock object before calling dispatcher.getEventHandler().handle.
      In this way, the yarn cluster usage rate can be 100 and the whole cluster is good running.
      Please see my attached patch.

        Attachments

        1. YARN-4398.2.patch
          4 kB
          NING DING
        2. YARN-4398.3.patch
          4 kB
          NING DING
        3. YARN-4398.4.patch
          4 kB
          NING DING

          Activity

            People

            • Assignee:
              iceberg565 NING DING
              Reporter:
              iceberg565 NING DING
            • Votes:
              0 Vote for this issue
              Watchers:
              11 Start watching this issue

              Dates

              • Created:
                Updated:
                Resolved: