[
https://issues.apache.org/jira/browse/MAPREDUCE-4715?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13472432#comment-13472432
]
Vladimir Klimontovich commented on MAPREDUCE-4715:
--------------------------------------------------
FileSystem class keeps a static cache of all FileSystem objects. See
FileSystem#CACHE static field. The cache key is a pair of file system URI
(string) and Configuration object. To fix that I'd suggest to run
FileSystem.closeAll() (which clears the cache) somewhere in the cleanup section
of TaskRunner.java
> Heap memory growing when JVM reuse option enabled
> -------------------------------------------------
>
> Key: MAPREDUCE-4715
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-4715
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: job submission
> Affects Versions: 0.20.2
> Environment: Linux, cdh3u5
> Reporter: Boris Ryakhovskiy
> Priority: Minor
> Attachments: VisualVM_screenshot.png
>
>
> When mapred.job.reuse.jvm.num.tasks option is set to 100 or more tasks, JVM
> fails with "java.lang.OutOfMemoryError: Java heap space" error message. Jobs
> itself are small and job heap size is set to -Xmx200m .
> It looks like the reason of the issue is a FileSystem$Cache object which
> collects JobConf objects from all jobs which were executed in current JVM. GC
> root for the FileSystem$Cache is ApplicationShutdownHooks class (cannot
> attach screenshot from visualvm)
> In my case each of JobConf objects allocates ~500K (50M total in case of 100
> jobs).
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
