I'll bet you are hitting MR-2998.
From the changelog:
MAPREDUCE-2998. Fixed a bug in TaskAttemptImpl which caused it to fork
bin/mapred too many times. Contributed by Vinod K V.
Arun
On Sep 21, 2011, at 9:52 AM, Chris Riccomini wrote:
> Hey Guys,
>
> My ApplicationMaster is being killed by the NodeManager because of memory
> consumption, and I don't understand why. I'm using -Xmx512M, and setting my
> resource request to 2048.
>
>
> .addCommand("java -Xmx512M -cp './package/*' kafka.yarn.ApplicationMaster
> " ...
>
> ...
>
> private var memory = 2048
>
> resource.setMemory(memory)
> containerCtx.setResource(resource)
> containerCtx.setCommands(cmds.toList)
> containerCtx.setLocalResources(Collections.singletonMap("package",
> packageResource))
> appCtx.setApplicationId(appId)
> appCtx.setUser(user.getShortUserName)
> appCtx.setAMContainerSpec(containerCtx)
> request.setApplicationSubmissionContext(appCtx)
> applicationsManager.submitApplication(request)
>
> When this runs, I see (in my NodeManager's logs):
>
>
> 2011-09-21 09:35:19,112 INFO monitor.ContainersMonitorImpl
> (ContainersMonitorImpl.java:run(402)) - Memory usage of ProcessTree 28134 for
> container-id container_1316559026783_0003_01_000001 : Virtual 2260938752
> bytes, limit : 2147483648 bytes; Physical 71540736 bytes, limit -1 bytes
> 2011-09-21 09:35:19,112 WARN monitor.ContainersMonitorImpl
> (ContainersMonitorImpl.java:isProcessTreeOverLimit(289)) - Process tree for
> container: container_1316559026783_0003_01_000001 has processes older than 1
> iteration running over the configured limit. Limit=2147483648, current usage
> = 2260938752
> 2011-09-21 09:35:19,113 WARN monitor.ContainersMonitorImpl
> (ContainersMonitorImpl.java:run(453)) - Container
> [pid=28134,containerID=container_1316559026783_0003_01_000001] is running
> beyond memory-limits. Current usage : 2260938752bytes. Limit :
> 2147483648bytes. Killing container.
> Dump of the process-tree for container_1316559026783_0003_01_000001 :
> |- PID PPID PGRPID SESSID CMD_NAME USER_MODE_TIME(MILLIS)
> SYSTEM_TIME(MILLIS) VMEM_USAGE(BYTES) RSSMEM_USAGE(PAGES) FULL_CMD_LINE
> |- 28134 25886 28134 28134 (bash) 0 0 108638208 303 /bin/bash -c java
> -Xmx512M -cp './package/*' kafka.yarn.ApplicationMaster 3 1 1316559026783
> com.linkedin.TODO 1
> 1>/tmp/logs/application_1316559026783_0003/container_1316559026783_0003_01_000001/stdout
>
> 2>/tmp/logs/application_1316559026783_0003/container_1316559026783_0003_01_000001/stderr
>
> |- 28137 28134 28134 28134 (java) 92 3 2152300544 17163 java -Xmx512M
> -cp ./package/* kafka.yarn.ApplicationMaster 3 1 1316559026783
> com.linkedin.TODO 1
>
> 2011-09-21 09:35:19,113 INFO monitor.ContainersMonitorImpl
> (ContainersMonitorImpl.java:run(463)) - Removed ProcessTree with root 28134
>
> It appears that YARN is honoring my 2048 command, yet my process is somehow
> taking 2260938752 bytes. I don't think that I'm using nearly that much in
> permgen, and my heap is limited to 512. I don't have any JNI stuff running
> (that I know of), so it's unclear to me what's going on here. The only thing
> that I can think of is that Java's Runtime exec is forking, and copying its
> entire JVM memory footprint for the fork.
>
> Has anyone seen this? Am I doing something dumb?
>
> Thanks!
> Chris
