[
https://issues.apache.org/jira/browse/YARN-9337?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16788086#comment-16788086
]
Szilard Nemeth commented on YARN-9337:
--------------------------------------
Hi [~adam.antal]!
Also adding [~tangzhankun] so he could chime in with an opinion on this matter.
Adam: I think you are right about this. At least for me, it's looks more usable
like the way you are thinking: nvidia-smi shouldn't have been called ever, if
the GPU devices are specified by an operator. I would also treat the
discovery-executables a non-mandatory parameter if the GPU devices are
specified. I guess the whole point of providing a way for operators to manually
specifiy devices is 2-fold:
# They are not required to install nvidia-smi so they can utilizie GPUs in a
more "standalone" fashion
# They want to restrict the number of GPUs a NM can utilize, which is not yet
possible with auto-discovery, as it will discover all GPU devices.
Does this make sense?
Thanks!
> GPU auto-discovery script runs even when the resource is given by hand
> ----------------------------------------------------------------------
>
> Key: YARN-9337
> URL: https://issues.apache.org/jira/browse/YARN-9337
> Project: Hadoop YARN
> Issue Type: Improvement
> Components: yarn
> Affects Versions: 3.2.0
> Reporter: Adam Antal
> Assignee: Adam Antal
> Priority: Major
> Attachments: YARN-9337.001.patch
>
>
> The nvidia-smi script is called even when the gpu configs are given by hand
> (so there's no need for GPU auto-discovery).
> We should mitigate the call of that script, since it has no effect. (The
> configs written by the user is not overwritten by the result of the
> auto-discovery script.)
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
