[MAPREDUCE-3612] Task.TaskReporter.done method blocked for some time when task is finishing

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Target Version/s: 0.20.205.0, 1.0.0
• Assignee: Binglin Chang
• Reporter: Binglin Chang
• Created: 03/Jan/12 12:11
• Updated: 07/Jul/12 00:31

Description

We recently have done some tests to evaluate performances of different Hadoop versions(1.0, 0.23, Baidu internal version), and found some weird results. One of them is in 1.0 Task.TaskReporter.done() takes too much time, about 2s, this is bad for small tasks. After reviewing source code and add some log, the following code block Task.TaskReporter.done

src/mapred/org/apache/hadoop/mapred/Task.java

 658           try {
 659             Thread.sleep(PROGRESS_INTERVAL);
 660           }


 723     public void stopCommunicationThread() throws InterruptedException {
 724       // Updating resources specified in ResourceCalculatorPlugin
 725       if (pingThread != null) {
 726         synchronized(lock) {
 727           while(!done) {
 728             lock.wait();
 729           }
 730         }
 731         pingThread.interrupt();
 732         pingThread.join();
 733       }
 734     }

Originally line 724-730 don't exists, I don't know why it is added. If it is needed, we can replace Thread.sleep with Object.wait(timeout) and Object.notify instead, so it won't block.

Issue Links

• is superceded by: MAPREDUCE-4400 Fix performance regression for small jobs/workflows