[
https://issues.apache.org/jira/browse/YARN-6964?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16118587#comment-16118587
]
Daniel Templeton commented on YARN-6964:
----------------------------------------
Previously the code was:{code} Resource desiredShare =
Resources.min(policy.getResourceCalculator(),
scheduler.getClusterResource(), getMinShare(), getDemand());{code} We
were already taking the min. What I changed here is that instead of allowing
the min share to inflate the the desired share for the non-dominant resource,
we take a proper component-wise min. Assume min share is 6GB and 6 vcores, and
my demand is 6GB and 3 vcores. If memory is dominant, then the code before
would boost the desired share to 6GB and 6 vcores. With my change, the desired
share will be 6GB and 3 vcores.
Also note that this calculation is only for the purposes of determining whether
the queue is min share starved. It has nothing to do with what the container
will be assigned.
> Fair scheduler misuses Resources operations
> -------------------------------------------
>
> Key: YARN-6964
> URL: https://issues.apache.org/jira/browse/YARN-6964
> Project: Hadoop YARN
> Issue Type: Bug
> Components: fairscheduler
> Affects Versions: 3.0.0-alpha4
> Reporter: Daniel Templeton
> Assignee: Daniel Templeton
> Attachments: YARN-6964.001.patch
>
>
> There are several places where YARN uses the {{Resources}} class to do
> comparisons of {{Resource}} instances incorrectly. This patch corrects those
> mistakes.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
