[
https://issues.apache.org/jira/browse/YARN-9235?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16769357#comment-16769357
]
Antal Bálint Steinbach commented on YARN-9235:
----------------------------------------------
Hi [~pbacsko], [~sunilg] ,
Yeah, that would be great but unfortunately, it is not so easy. This class is
not prepared to be testable.
_GpuDiscoverer.getInstance().getGpuDeviceInformation()_ will throw exception
before we reach the code we would like to test. Me and [~snemeth] has some
patches available to address this issue. I would not do the same change in the
3rd patch for this. (add GpuDiscoverer as a dependency for the class)
YARN-9217 for example has a test which is testing this method. Unfortunately,
this issue is blocked by [~snemeth]'s other pending commits (YARN-9118,
YARN-9213), because they are conflicting badly.
I would recommend submitting those first then I merge my issues or submit this
without test and I resolve the problem with the other mentioned issues.
> If linux container executor is not set for a GPU cluster
> GpuResourceHandlerImpl is not initialized and NPE is thrown
> --------------------------------------------------------------------------------------------------------------------
>
> Key: YARN-9235
> URL: https://issues.apache.org/jira/browse/YARN-9235
> Project: Hadoop YARN
> Issue Type: Bug
> Components: yarn
> Affects Versions: 3.0.0, 3.1.0
> Reporter: Antal Bálint Steinbach
> Assignee: Antal Bálint Steinbach
> Priority: Major
> Attachments: YARN-9235.001.patch
>
>
> If GPU plugin is enabled for the NodeManager, it is possible to run jobs with
> GPU.
> However, if LinuxContainerExecutor is not configured, an NPE is thrown when
> calling
> {code:java}
> GpuResourcePlugin.getNMResourceInfo{code}
> Also, there are no warns in the log if GPU is misconfigured like this.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
