[
https://issues.apache.org/jira/browse/YUNIKORN-3115?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18014628#comment-18014628
]
Peter Bacsko commented on YUNIKORN-3115:
----------------------------------------
During our offline meeting with [~wilfreds], we recommend and absolute number
which becomes a queue-level setting. If it's unset, it's inherited from the
parent. The value cannot be higher than the one which propagates from "root".
We need a good default value for the threshold (eg. 100?) and for the backoff
time.
> Skip scheduling application if too many pods are unschedulable
> --------------------------------------------------------------
>
> Key: YUNIKORN-3115
> URL: https://issues.apache.org/jira/browse/YUNIKORN-3115
> Project: Apache YuniKorn
> Issue Type: Improvement
> Components: core - scheduler
> Reporter: Peter Bacsko
> Assignee: Peter Bacsko
> Priority: Major
>
> If an application has too many unschedulable pods, it can drastically affect
> scheduling throughput.
> In every cycle, we attempt to schedule all pods. If the app has hundreds or
> thousands of pods on a large cluster, then it can severly limit the overall
> performance of Yunikorn.
> We can just skip the application for a given amount of time (90..300 seconds)
> if a certain number of pods are pending.
> Possible solutions:
> 1. Percentage of pods vs absolute number (eg 25% vs 100)
> 2. Partition-wide config vs queue-level setting
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
