DRA - Define UX for using multiple multi-host resources

[geetasg]

Enhancement Description

• One-line enhancement description (can be used as a release note): Provide a recommendation on how to configure a training job to utilize DRA via partitionable devices spanning multiple multi-host resources.
• Kubernetes Enhancement Proposal:
• Discussion Link:
• PRs by stage and milestone:
  • Alpha - v1.xx
    • KEP (k/enhancements) update PR(s):
    • Code (k/k) update PR(s):
    • Docs (k/website) update PR(s):

Please keep this description up to date. This will help the Enhancement Team to track the evolution of the enhancement efficiently.

Issue - kubernetes-sigs/jobset#762 - requests updates to workload constructs to enable the instantiation of resourceclaims at the right level. All workload constructs will need to be updated to understand this paradigm (ref:kubernetes-sigs/jobset#762 (comment)).
This issue is to discuss if there is a way to enable this for all common workload definitions via DRA without requiring the workload definition update.

[kannon92]

kubernetes/kubernetes#132192

@dom4ha @sanposhiho

I think that this work would be in scope for the "workload-aware scheduling".

[johnbelamaric]

/wg device-management
/sig scheduling

[johnbelamaric]

I think that this work would be in scope for the "workload-aware scheduling".

I agree, Kevin. At least, it's related to it. It may also be related to gang scheduling. Effectively, it's "how do we define and utilize resources that are shared amongst one or more groups of pods".

cc @erictune