You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@mesos.apache.org by "Jessica J (JIRA)" <ji...@apache.org> on 2012/07/19 16:30:35 UTC
[jira] [Comment Edited] (MESOS-206) Long-running jobs on Hadoop
framework do not run to completion
[ https://issues.apache.org/jira/browse/MESOS-206?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13418324#comment-13418324 ]
Jessica J edited comment on MESOS-206 at 7/19/12 2:30 PM:
----------------------------------------------------------
After more digging, I've discovered the JobTracker is seriously overusing resources. Top output indicates it's using almost 10 GB of virtual memory (and the number steadily grows larger as the JobTracker continues to run) and almost 800% CPU:
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
28868 jessicaj 20 0 9069m 1.3g 15m S 737.6 8.2 257:12.30 java
NameNode and RunJar are also using fairly large amounts of virtual memory at 8 GB and 6 GB respectively.
My guess is a serious memory leak...
Edited to add: The JobTracker memory usage continues to grow significantly even after I kill any jobs I was attempting to run.
was (Author: esohpromatem):
After more digging, I've discovered the JobTracker is seriously overusing resources. Top output indicates it's using almost 10 GB of virtual memory (and the number steadily grows larger as the JobTracker continues to run) and almost 800% CPU:
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
28868 jessicaj 20 0 9069m 1.3g 15m S 737.6 8.2 257:12.30 java
NameNode and RunJar are also using fairly large amounts of virtual memory at 8 GB and 6 GB respectively.
My guess is a serious memory leak...
> Long-running jobs on Hadoop framework do not run to completion
> --------------------------------------------------------------
>
> Key: MESOS-206
> URL: https://issues.apache.org/jira/browse/MESOS-206
> Project: Mesos
> Issue Type: Bug
> Components: framework
> Reporter: Jessica J
> Priority: Blocker
>
> When I run the MPI and Hadoop frameworks simultaneously with long-running jobs, the Hadoop jobs fail to complete. The MPI job, which is shorter, completes normally, and the Hadoop framework continues for a while, but eventually, although it appears to still be running, it stops making progress on the jobs. The jobtracker keeps running, but each line of output indicates no map or reduce tasks are actually being executed:
> 12/06/08 10:55:41 INFO mapred.FrameworkScheduler: Assigning tasks for [slavehost] with 0 map slots and 0 reduce slots
> I've examined the master's log and noticed this:
> I0608 10:40:43.106740 6317 master.cpp:681] Deactivating framework 201206080825-36284608-5050-6311-0000 as requested by scheduler(1)@[my-ip]:59317
> The framework ID is that of the Hadoop framework. This message is followed by messages indicating the slaves "couldn't lookup task [#]" and "couldn't lookup framework 201206080825-36284608-5050-6311-0000."
> I thought the first time that this error was a fluke since it does not happen with shorter running jobs or with the Hadoop framework running independently (i.e., no MPI), but I have now consistently reproduced it 4 times.
> UPDATE: I just had the same issue occur when running Hadoop + Mesos without the MPI framework running simultaneously.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira