[
https://issues.apache.org/jira/browse/YARN-7496?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16252643#comment-16252643
]
Eric Payne commented on YARN-7496:
----------------------------------
Cluster Configuration:
- Cluster Memory: 20GB
- Queue1 capacity and max capacity: 50% : 100%
- Queue2 capacity and max capacity: 50% : 100%
- Queue1: Intra-queue preemption: enabled
- Default container size: 0.5GB
Use Case:
- User1 submits App1 in Queue1 and consumes 12.5GB
- User2 submits App2 in Queue1 and consumes 7.5GB
- User3 submits App3 in Queue1
- Preemption monitor calculates user limit to be {{((total used resources in
Queue1) / (number of all users)) + (1 container) = normalizeup((20GB/3),0.5GB)
+ 0.5GB = 7GB + 0.5GB = 7.5GB}}
- Preemption monitor sees that App1 is the only one that has resources, so it
tries to preempts containers down to 7.5GB from {{App1}}.
- The problem comes here: Capacity Scheduler calculates user limit to be
{{((total used resources in Queue1) / (number of active users)) + (1 container)
= normalizeup((20GB/2),0.5GB) + 0.5GB = 10GB + 0.5GB = 10.5GB}}
- Therefore, once {{App1}} gets to 10.5GB, the preemption monitor will try to
preempt 2.5GB more resources from {{App1}}, but the Capacity Scheduler gives
them back. This creates oscillation.
> CS Intra-queue preemption user-limit calculations are not in line with
> LeafQueue user-limit calculations
> --------------------------------------------------------------------------------------------------------
>
> Key: YARN-7496
> URL: https://issues.apache.org/jira/browse/YARN-7496
> Project: Hadoop YARN
> Issue Type: Bug
> Affects Versions: 2.8.2
> Reporter: Eric Payne
> Assignee: Eric Payne
>
> Only a problem in 2.8.
> Preemption could oscillate due to the difference in how user limit is
> calculated between 2.8 and later releases.
> Basically (ignoring ULF, MULP, and maybe others), the calculation for user
> limit on the Capacity Scheduler side in 2.8 is {{total used resources /
> number of active users}} while the calculation in later releases is {{total
> active resources / number of active users}}. When intra-queue preemption was
> backported to 2.8, it's calculations for user limit were more aligned with
> the latter algorithm, which is in 2.9 and later releases.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
