[jira] Commented: (CASSANDRA-504) Reduce GC overhead

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

    [ https://issues.apache.org/jira/browse/CASSANDRA-504?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12768664#action_12768664 ] 

Hudson commented on CASSANDRA-504:
----------------------------------

Integrated in Cassandra #235 (See [http://hudson.zones.apache.org/hudson/job/Cassandra/235/])
    add -XX:SurvivorRatio=128 -XX:MaxTenuringThreshold=0 to default gc options.  patch by Brandon Williams; reviewed by jbellis for 


> Reduce GC overhead
> ------------------
>
>                 Key: CASSANDRA-504
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-504
>             Project: Cassandra
>          Issue Type: Improvement
>         Environment: debian lenny amd64 Java(TM) SE Runtime Environment (build 1.6.0_12-b04)
>            Reporter: Brandon Williams
>            Assignee: Brandon Williams
>            Priority: Minor
>             Fix For: 0.4, 0.5
>
>         Attachments: better_gc_jvm_opts.patch
>
>
> The Concurrent Mark Sweep GC is using about 30% of the cpu during inserts.  This is not reported as GC time by jconsole, but can be observed by finding the thread with jstack and watching it via top.  This is due to the CMSInitiatingOccupancyFraction=1 setting being passed to the JVM.  Setting this to 1 is nonsensical, the default is 68.
> I tested with some runs of stress.py with 200k inserts, and removing  CMSInitiatingOccupancyFraction causes the 30% cpu usage to disappear.  ParNew GC time was 28-29s.  Adding SurvivorRatio=128 and MaxTenuringThreshold=0 drops the GC time to 18-20s.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.