[
https://issues.apache.org/jira/browse/YARN-2113?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15975291#comment-15975291
]
Wangda Tan commented on YARN-2113:
----------------------------------
[~sunilg]/[~jlowe]/[~eepayne],
I agree to keep the logic to as simple as possible, for now we can just make
sure preemption doesn't make used resource of a user dropped lower than
computed UL. If massive changes needed to support this, we need discuss if we
should move this to a separate JIRA.
Actually all jiggly preemption issues cannot be prevented by adding a dead zone
(approach suggested by Jason is one special kind of deadzone, which is not
configurable and has {{value=used-guaranteed}}, and limit number of resource to
preempt). In some cases it won't work such as UL could be changed after
preemption.
I've a raw idea which could possibly solve the problem better, the idea is:
build a directed graph, vertices are entities like queue/users/apps, and values
on edges are resources preempted between each other over a period of time.
Scheduler can find cycles in the graph to figure out and prevent jiggly
preemption happens. I don't have more detailed thoughts. Just throw it here for
more thoughts. [~curino], not sure if you thought about this issue before.
> 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: [email protected]
For additional commands, e-mail: [email protected]
