[
https://issues.apache.org/jira/browse/YARN-6670?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16045272#comment-16045272
]
Konstantinos Karanasos commented on YARN-6670:
----------------------------------------------
Thanks for the patch, [~haibochen].
Two nits:
- in the ContainersMonitorImpl, when you are printing the overallocation
preemption threshold, I would print it with the rest of the overallocation
thresholds, to show that it is related to overallocation. Preemption of
opportunistic containers will also happen either way if guaranteed ones arrive
and there are no resources for them to start.
- in the default.xml, there is a typo "is is".
One more comment that does not need to be addressed: we could define the
preemption threshold on top of the overallocation threshold. So in your default
values, it will be 0.01f. This way, it might look less ad hoc than the 0.96f.
But I don't mind it like it is for now -- up to you.
Otherwise, +1 from me.
> Add separate NM overallocation thresholds for cpu and memory
> ------------------------------------------------------------
>
> Key: YARN-6670
> URL: https://issues.apache.org/jira/browse/YARN-6670
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: nodemanager
> Affects Versions: 3.0.0-alpha3
> Reporter: Haibo Chen
> Assignee: Haibo Chen
> Attachments: YARN-6670-YARN-1011.00.patch,
> YARN-6670-YARN-1011.01.patch, YARN-6670-YARN-1011.02.patch,
> YARN-6670-YARN-1011.03.patch
>
>
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
