[jira] [Commented] (IGNITE-494) Ignite node runs OutOfMemory when running Hadoop tasks with native Hadoop filesystem

["Ivan Veselovsky (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/IGNITE-494?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14382039#comment-14382039 ] 

Ivan Veselovsky commented on IGNITE-494:
----------------------------------------

Problem cause are never stopping threads that start from org.apache.hadoop.hdfs.PeerCache#startExpiryDaemon .
These threads hold hard references to HadoopClassLoader (due to Runnable loaded by it and context class loader set), that prevents all its classes from being freed.

> Ignite node runs OutOfMemory when running Hadoop tasks with native Hadoop filesystem
> ------------------------------------------------------------------------------------
>
>                 Key: IGNITE-494
>                 URL: https://issues.apache.org/jira/browse/IGNITE-494
>             Project: Ignite
>          Issue Type: Bug
>          Components: hadoop
>    Affects Versions: sprint-1
>            Reporter: Ivan Veselovsky
>            Assignee: Vladimir Ozerov
>             Fix For: sprint-2
>
>         Attachments: IGNITE-494--b.patch, PermGen.png, classes.png
>
>
> To reproduce: tun terasort Hadoop sample using Ignite mapred engine with native Hadoop filesystem (original Hadoop core-site.xml, with e.g. fs.defaultFs=hdfs://quickstart.cloudera:8020).
> After ~10 runs of the teragen sample without Ignite node restart the node crashes with  "Caused by: java.lang.OutOfMemoryError: PermGen space"
> Important to note that this does *not* happen if igfs:// filesystem backed by hdfs:// is used: in this case PermGen is cleared okay.
> The below screenshots illustrate the problem. 
> (Heap dumps are also available, but they are too large to be attached.)



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)