[
https://issues.apache.org/jira/browse/TWILL-73?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13972126#comment-13972126
]
Terence Yim commented on TWILL-73:
----------------------------------
Lets say there are two runnables inside an application, one needs 1GB of
memory, the other one needs 500MB of memory. When constructed to starts in the
order of the larger one followed by the smaller one, AM only able to acquire a
container for the larger one, but not the smaller one afterwards, even the
cluster has enough capacity.
The underlying reason is because of YARN-314, the RM scheduler doesn't
differentiate different resource requests with the same priority.
When happened was once a 1GB container was acquired, AM told AMRMClient to
remove the container request (which is necessary, otherwise when requesting
more containers of the same size, you'll get extra ones), AMRMClient would send
a container request of size = 0 on the next allocate call. If in the same
allocate call there is another container request of size != 0 and combining
with the effect YARN-314, the schedule may not be able to see it.
> Unable to allocate small container after large container has been acquired
> --------------------------------------------------------------------------
>
> Key: TWILL-73
> URL: https://issues.apache.org/jira/browse/TWILL-73
> Project: Apache Twill
> Issue Type: Bug
> Components: yarn
> Affects Versions: 0.2.0-incubating
> Reporter: Terence Yim
> Assignee: Terence Yim
>
--
This message was sent by Atlassian JIRA
(v6.2#6252)
