[MAPREDUCE-4728] Interaction between oob heartbeats and damper can cause TT to heartbeat with zero delay

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 1.0.3
• Assignee: Unassigned
• Reporter: Nathan Roberts
• Created: 17/Oct/12 16:14
• Updated: 20/Feb/13 22:00

Description

When mapreduce.tasktracker.outofband.heartbeat is true and mapreduce.tasktracker.outofband.heartbeat.damper is something largish (like the default of 1000000), the TT doesn't wait for tasks to finish before heartbeating back to the JT. This causes excessive load on the JT which in-turn reduces overall cluster performance.

I believe the problem is that in the following block of code, when getHeartbeatInterval() returns 0, we heartbeat back immediately BUT finishedCount does not get reset. It looks like nothing ever gets us out of this situation so we basically heartbeat without ever sleeping.

        // accelerate to account for multiple finished tasks up-front
        long remaining =
          (lastHeartbeat + getHeartbeatInterval(finishedCount.get())) - now;
        while (remaining > 0) {
          // sleeps for the wait time or
          // until there are *enough* empty slots to schedule tasks
          synchronized (finishedCount) {
            finishedCount.wait(remaining);

            // Recompute
            now = System.currentTimeMillis();
            remaining =
              (lastHeartbeat + getHeartbeatInterval(finishedCount.get())) - now;

            if (remaining <= 0) {
              // Reset count
              finishedCount.set(0);
              break;
            }
          }
        }