[ 
https://issues.apache.org/jira/browse/YARN-9892?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17016151#comment-17016151
 ] 

Wangda Tan commented on YARN-9892:
----------------------------------

[~pbacsko], 

My concern is adding DRF only in application ordering policy is not meaningful,

All the resource requests of apps are normalized differently by resource 
calculator implementations (default or drf): 
org.apache.hadoop.yarn.util.resource.DefaultResourceCalculator#normalize

When a global resource calculator is configured as "default", we will simply 
ignore vcores in all calculation, including queue, nodes, apps, so what does 
DRF mean when in that scenario?

To me, we should map both FairScheduler's fair/drf to fair in 
CapacityScheduler. Because in CapacityScheduler it do app/queue sorting based 
on global resource calculator, which could be fair (mem only, when default 
calculator configured) or drf (all resource types, when DRC configured).

> Capacity scheduler: support DRF ordering policy on queue level
> --------------------------------------------------------------
>
>                 Key: YARN-9892
>                 URL: https://issues.apache.org/jira/browse/YARN-9892
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: capacity scheduler
>            Reporter: Peter Bacsko
>            Assignee: Manikandan R
>            Priority: Major
>         Attachments: YARN-9892-003.patch, YARN-9892.001.patch, 
> YARN-9892.002.patch
>
>
> Capacity scheduler does not support DRF (Dominant Resource Fairness) ordering 
> policy on queue level. Only "fifo" and "fair" are accepted for 
> {{yarn.scheduler.capacity.<queue-path>.ordering-policy}}.
> DRF can only be used globally if 
> {{yarn.scheduler.capacity.resource-calculator}} is set to 
> DominantResourceCalculator.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org

Reply via email to