[
https://issues.apache.org/jira/browse/YARN-6670?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16039597#comment-16039597
]
Haibo Chen commented on YARN-6670:
----------------------------------
Thanks [~kkaranasos] for your review!
bq. The parameter names that contain both "overallocation" and "allocation"
seem a bit redundant. I'd suggest to use either the simpler
"overallocation.memory.threshold" or the more precise
"overallocation.memory-utilization-threshold".
Good point. Will address it in the new patch
bq. Using 0 as a threshold to disable overallocation is somehow
counter-intuitive (as in "allow containers to use 0% of the resources"). How
about we use -1 to make clear it is disabled?
NM overallocation threshold x can be interpreted as something like NM saying,
I'd like to oversubscribe myself, i.e. accept opportunistic containers to
consume allocated-but-unused resources, if my resource utilization is under x%
and my resources have been fully allocated. I'll update the default to -1 to
make it consistent with other configs.
bq. The preemption threshold is expected to always be higher than the
allocation threshold, right? Shall we add a check for that?
Yes. will address this in the new patch.
> 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
>
>
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
