[
https://issues.apache.org/jira/browse/MAPREDUCE-6501?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15435690#comment-15435690
]
Haibo Chen commented on MAPREDUCE-6501:
---------------------------------------
Task 2 is already implemented with taskAttemptProgress, that is, SHUFFLE phase
is expressed in terms of progress, and preemption also tries to preempt
reducers that are of least progress.
Task 3 is inapplicable since reducer preemption never kicks in if there is one
mapper can run.
> Improve reducer preemption
> --------------------------
>
> Key: MAPREDUCE-6501
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-6501
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Components: applicationmaster
> Affects Versions: 2.7.1
> Reporter: Karthik Kambatla
> Assignee: Haibo Chen
>
> As discussed on MAPREDUCE-6302, we could improve the reducer preemption as
> follows:
> # preempt enough reducers so there are enough mappers to match slowstart
> # prioritize preempting reducers that are still in SHUFFLE phase
> # add an option to not preempt reducers that are past SHUFFLE phase
> irrespective of slowstart as long as one mapper can run
> We could do it all in one patch or create subtasks as necessary.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
