[jira] [Commented] (YARN-9099) GpuResourceAllocator.getReleasingGpus calculates number of GPUs in a wrong way

["Zhankun Tang (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/YARN-9099?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16714513#comment-16714513 ] 

Zhankun Tang commented on YARN-9099:
------------------------------------

[~snemeth], Thanks for catching up this! The patch looks good to me. And a test case would be better.

> GpuResourceAllocator.getReleasingGpus calculates number of GPUs in a wrong way
> ------------------------------------------------------------------------------
>
>                 Key: YARN-9099
>                 URL: https://issues.apache.org/jira/browse/YARN-9099
>             Project: Hadoop YARN
>          Issue Type: Bug
>            Reporter: Szilard Nemeth
>            Assignee: Szilard Nemeth
>            Priority: Major
>         Attachments: YARN-9099.001.patch
>
>
> getReleasingGpus plays an important role in the calculation which happens when GpuAllocator assign GPUs to a container, see: GpuResourceAllocator#internalAssignGpus.
> If multiple GPUs are assigned to the same container, getReleasingGpus will return an invalid number.
> The iterator goes over on mappings of (GPU device, container ID) and it retrieves the container by its ID the number of times the container ID is mapped to any device.
> Then for every container, the resource value for the GPU resource is added to a running sum.
> Obviously, if a container is mapped to 2 or more devices, then the container's GPU resource counter is added to the running sum as many times as the number of GPU devices the container has.
> Example: 
> Let's suppose {{usedDevices}} contains these mappings: 
> - (GPU1, container1)
> - (GPU2, container1)
> - (GPU3, container2)
> GPU resource value is 2 for container1 and 
> GPU resource value is 1 for container2.
> Then, if container1 is in a running state, getReleasingGpus will return 4 instead of 2.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

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