[
https://issues.apache.org/jira/browse/FLINK-7878?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16243824#comment-16243824
]
ASF GitHub Bot commented on FLINK-7878:
---------------------------------------
Github user tillrohrmann commented on the issue:
https://github.com/apache/flink/pull/4911
I think what you are referring to is the current trunk branch of Hadoop
where they introduce `ResourceInformation`. Currently, the set of supported
resources is `"memory-mb", "vcores", "yarn.io/gpu"`. Given that you have to
exactly set these names for the resource which are exclusive for `YARN`, I
would actually say that we should explicitly add these resources (e.g.
introducing a proper `GPUResource`). That way, this would also work if you let
your job run on Mesos without having to adapt the resource names. Adding more
resources should be fairly easy and therefore we don't need the generic
resource. Does that make sense @shuai-xu.
> Extend the resource type user can define in ResourceSpec
> --------------------------------------------------------
>
> Key: FLINK-7878
> URL: https://issues.apache.org/jira/browse/FLINK-7878
> Project: Flink
> Issue Type: Improvement
> Components: DataSet API, DataStream API
> Reporter: shuai.xu
> Assignee: shuai.xu
> Labels: flip-6
>
> Now, flink only support user define how much CPU and MEM used in an operator,
> but now the resource in a cluster is various. For example, an application for
> image processing may need GPU, some others may need FPGA.
> Only CPU and MEM is not enough, and the resource type is becoming more and
> more, so we need to make the ResourSpec extendible.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
