[ 
https://issues.apache.org/jira/browse/YARN-389?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13579562#comment-13579562
 ] 

Zhijie Shen commented on YARN-389:
----------------------------------

I've checked reason why the requested AM size is larger than 
yarn.scheduler.maximum-allocation-mb. In fact, the AM size is not checked 
before requesting a container. It is because MR jobs are submitted through 
ResourceMgrDelegate and managed by MRAppMaster, while checking AM size against 
the maximum resource capability is only implemented in Client and 
ApplicationMaster of the distributed shell.

In addition, since the requested AM size will be rounded up, it is at the risk 
of the roundup value goes beyond yarn.scheduler.maximum-allocation-mb.

Anyway, the aforementioned situation is a separate issue, which I've filed in 
YARN-407. Here, the focused issue is that if the requested resource is 
impossible to be supplied, there should be some mechanism to stop the infinite 
loop of assignment. This issue seems to be similar to YARN-56 and YARN-394.




                
> Infinitely assigning containers when the required resource exceeds the 
> cluster's absolute capacity
> --------------------------------------------------------------------------------------------------
>
>                 Key: YARN-389
>                 URL: https://issues.apache.org/jira/browse/YARN-389
>             Project: Hadoop YARN
>          Issue Type: Bug
>            Reporter: Zhijie Shen
>            Assignee: Zhijie Shen
>
> I've run wordcount example on branch-2 and trunk. I've set 
> yarn.nodemanager.resource.memory-mb to 1G and 
> yarn.app.mapreduce.am.resource.mb to 1.5G. Therefore, resourcemanager is to 
> assign a 2G AM container for AM. However, the nodemanager doesn't have enough 
> memory to assign the container. The problem is that the assignment operation 
> will be repeated infinitely, if the assignment cannot be accomplished. Logs 
> follow.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to