[
https://issues.apache.org/jira/browse/YARN-6612?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16154362#comment-16154362
]
Yufei Gu commented on YARN-6612:
--------------------------------
Thanks for the patch, [~templedf], it looks much better.
- Zero-weight queues are supported by FS. We may only be lucky due to nobody
uses it in DRF scenario. People will tend to use DRF more in case of multiple
resource types. In that sense, it is necessary to handle zero-weight case.
- Let's talk offline about weight in min share calculation and sort by dominant
min share which is minShare/cluste.
- About naming, I think we have used the word 'share' for different concepts.
1. it has been used in concepts like "Fair Share", "Min Share", "Max Share",
which are not necessary but usually absolute resource of a schedulable. 2. In
DRF, 'share' refers to how big a fraction of a resource is comparing to the
cluster resource. In {{calculateShares}} and {{calculateMinShares}}, we don't
only calculate how big a fraction of a used resource is comparing to the
cluster resource, but also calculate how much portion it has in Fair Share and
Min Share. That's why I think names of {{calculateShares}} and
{{calculateMinShares}} are not ideal. In addition, Java doc of those two
functions should mention they calculate how much fair share/min share the
schedulable use.
> Update fair scheduler policies to be aware of resource types
> ------------------------------------------------------------
>
> Key: YARN-6612
> URL: https://issues.apache.org/jira/browse/YARN-6612
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: fairscheduler
> Affects Versions: YARN-3926
> Reporter: Daniel Templeton
> Assignee: Daniel Templeton
> Attachments: YARN-6612.YARN-3926.001.patch,
> YARN-6612.YARN-3926.002.patch, YARN-6612.YARN-3926.004.patch,
> YARN-6612.YARN-3926.005.patch, YARN-6612.YARN-3926.006.patch,
> YARN-6612.YARN-3926.007.patch, YARN-6612.YARN-3926.008.patch
>
>
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
