[
https://issues.apache.org/jira/browse/MAPREDUCE-4671?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13460115#comment-13460115
]
Bikas Saha commented on MAPREDUCE-4671:
---------------------------------------
Saying I need x out of y is done by saying I need x because the protocol is
built around setting absolute ask values. Since the protocol is lossy wrt
relationships between container requests, in general it would be hard (perhaps
impossible in some cases) for the RM to figure out what the other requests were
that it now needs to cancel unless that relationship information is passed in
by the AM.
>From what I see, communication overhead would be similar in both approaches (1
>allocate message to send the information). I am not sure how RM side work
>would be less. Approach 1 seems simpler to me given that RM code does not need
>to be changed :) Though I might be wrong :P
Perhaps I might be missing something in your comments. A few examples would
help.
> AM does not tell the RM about container requests that are no longer needed
> --------------------------------------------------------------------------
>
> Key: MAPREDUCE-4671
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-4671
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Reporter: Bikas Saha
> Assignee: Bikas Saha
> Attachments: MAPREDUCE-4671.1.patch
>
>
> Say the AM wanted a container at hosts h1, h2, h3. After getting a container
> at h1 it should tell RM that it no longer needs containers at h2, h3.
> Otherwise on the RM h2, h3 remain valid allocation locations.
> The AM RMContainerAllocator does remove these resource requests internally.
> When the resource request container count drops to 0 then it drops the
> resource request from its tables but forgets to send the 0 sized request to
> the RM.
--
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
