[jira] [Commented] (MAPREDUCE-4585) Checkpoint shuffle aggregation as map output

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

    [ https://issues.apache.org/jira/browse/MAPREDUCE-4585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13442638#comment-13442638 ] 

Carlo Curino commented on MAPREDUCE-4585:
-----------------------------------------

The image posted shows an example of a run where checkpointing-restart helps "hiding" part of the shuffling costs, by allowing (two in this example) reducers to take turns over the only available slot, and thus getting to the point in which all maps have completed with most of the shuffling done. The basic behavior would have the reducer one sitting on the slot until all maps complete, and thus reduce 2 paying the full cost of shuffling "after" map5 completed and a new slot is available. We believe this should help for non-ideal schedules (multiple waves of maps/reduces which are not perfectly aligned) and to handle skew.
 
                
> Checkpoint shuffle aggregation as map output
> --------------------------------------------
>
>                 Key: MAPREDUCE-4585
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-4585
>             Project: Hadoop Map/Reduce
>          Issue Type: New Feature
>          Components: task
>            Reporter: Chris Douglas
>         Attachments: shufflecheckpoint.pdf
>
>
> Map output collected during the shuffle can be spilled and written as a composite of map outputs. Particularly if the job employs a combiner, this checkpoint can provide fault tolerance and improve job throughput by aggregating intermediate output. The latter is especially helpful for jobs with multiple waves of reduces.

--
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