[ 
https://issues.apache.org/jira/browse/YARN-2113?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15973638#comment-15973638
 ] 

Jason Lowe commented on YARN-2113:
----------------------------------

Is a deadzone the proper way to fix this?  I'm thinking of a case where the 
user has a particularly large container, larger than the dead zone.  It will 
still flap in this case, correct?  Seems like we should preempt not until we 
fall below the user limit but instead until the _next_ container we would 
preempt would put the user at or below their limit.  The scheduler essentially 
entitles a user to one container beyond the user limit.  If we preempt down to 
a point at or below the user's limit then we've gone one container too far, and 
the scheduler could very well turn around and give the container right back.

Preempting down to one container before we meet or dip below the user limit has 
the advantage that there's not yet another config to setup correctly.  However 
it brings up an interesting scenario where killling the youngest container 
would lower the utilization below the user's limit but killing older, smaller 
containers would not.

> Add cross-user preemption within CapacityScheduler's leaf-queue
> ---------------------------------------------------------------
>
>                 Key: YARN-2113
>                 URL: https://issues.apache.org/jira/browse/YARN-2113
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: scheduler
>            Reporter: Vinod Kumar Vavilapalli
>            Assignee: Sunil G
>         Attachments: 
> TestNoIntraQueuePreemptionIfBelowUserLimitAndDifferentPrioritiesWithExtraUsers.txt,
>  YARN-2113.0001.patch, YARN-2113.0002.patch, YARN-2113.0003.patch, 
> YARN-2113.0004.patch, YARN-2113.0005.patch, YARN-2113.0006.patch, 
> YARN-2113.0007.patch, YARN-2113.v0.patch
>
>
> Preemption today only works across queues and moves around resources across 
> queues per demand and usage. We should also have user-level preemption within 
> a queue, to balance capacity across users in a predictable manner.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org

Reply via email to