[
https://issues.apache.org/jira/browse/MAPREDUCE-3289?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13237715#comment-13237715
]
Todd Lipcon commented on MAPREDUCE-3289:
----------------------------------------
Found an issue with this optimization today in the case where some map output
partitions are really large (eg on the order of many GBs). Because the fadvise
is a blocking call, this can block on the order of minutes before sending the
first bytes of response data. This causes the reducer to time out, etc.
Instead, we should fadvise only a "chunk" ahead of the current read position.
> Make use of fadvise in the NM's shuffle handler
> -----------------------------------------------
>
> Key: MAPREDUCE-3289
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-3289
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Components: mrv2, nodemanager, performance
> Affects Versions: 0.23.0
> Reporter: Todd Lipcon
> Assignee: Todd Lipcon
> Attachments: mr-3289.txt
>
>
> Using the new NativeIO fadvise functions, we can make the NodeManager
> prefetch map output before it's send over the socket, and drop it out of the
> fs cache once it's been sent (since it's very rare for an output to have to
> be re-sent). This improves IO efficiency and reduces cache pollution.
--
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
