[jira] [Commented] (MAPREDUCE-3512) Batch jobHistory disk flushes

["Siddharth Seth (Commented) (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MAPREDUCE-3512?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13183737#comment-13183737 ] 

Siddharth Seth commented on MAPREDUCE-3512:
-------------------------------------------

Had run some crude benchmarks - writing out about 40MB of data (iirc). With a flush after every 0.5KB - this took roughly 35 seconds. With one flush at the end - around 6 seconds. Flush every 500KB - about 12 seconds.

Also, a benchmark like AMScalability (100K map tasks) currently takes over 5 minutes after job complete to finish the history file.
                
> Batch jobHistory disk flushes
> -----------------------------
>
>                 Key: MAPREDUCE-3512
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-3512
>             Project: Hadoop Map/Reduce
>          Issue Type: Sub-task
>          Components: mr-am, mrv2
>    Affects Versions: 0.23.0
>            Reporter: Siddharth Seth
>            Assignee: Siddharth Seth
>            Priority: Blocker
>         Attachments: MR3512.txt, MR3512.txt
>
>
> The mr-am flushes each individual job history event to disk for AM recovery. The history even handler ends up with a significant backlog for tests like MAPREDUCE-3402. 
> History events could be batched up based on num records / time / TaskFinishedEvents to reduce the number of DFS writes - with the potential drawback of having to rerun some tasks during AM recovery.

--
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