topologymanager: add CPU-attached NUMA filter option

[fanzhangio]

What type of PR is this?

Bug Fix

/kind bug

What this PR does / why we need it:

This PR adds an opt-in TopologyManager mode for large NUMA systems with many CPU-less NUMA nodes.
On systems such as NVIDIA Grace/Vera-class platforms (typically GB200), firmware can expose a large NUMA ID space even though only a small subset of NUMA nodes actually have CPUs. When kubelet is configured to allow that large NUMA count, topology hint generation can still scale poorly in the device-manager / topology-manager path and block admission before device plugin Allocate() is reached.

This PR adds a new TopologyManager policy option:

• restrict-to-cpu-numa-nodes

When enabled, TopologyManager computes an effective NUMA-node set containing only NUMA nodes with CPUs attached and propagates that same view to all topology-aware hint providers.

Specifically:

• TopologyManager uses the filtered NUMA set for policy creation and NUMA accounting.
• CPUManager generates hints only across that filtered set.
• MemoryManager generates hints only across that filtered set.
• DeviceManager projects device topology onto that same filtered set using NUMA distance information so hint generation and aligned allocation stay consistent.

This is intended to address the large-NUMA root cause in #135541.
This PR is separate from the reserved-memory consistency fix. That is a different bug and should be reviewed independently.

Which issue(s) this PR is related to:

Fixes #135541
KEP: KEP-5726

Special notes for your reviewer:

1. This reworks the earlier approach from Support TopologyManager in Large NUMA System (NVIDIA GB200) #135581 around a TopologyManager option instead of provider-local heuristics.

Behavior is unchanged unless restrict-to-cpu-numa-nodes is explicitly enabled.

Example kubelet config for affected systems:

topologyManagerPolicy: best-effort
topologyManagerScope: pod
topologyManagerPolicyOptions:
  prefer-closest-numa-nodes: "true"
  max-allowable-numa-nodes: "34"
  restrict-to-cpu-numa-nodes: "true"
cpuManagerPolicy: static

2. This PR reuses some of the same plumbing which was carried by a separate PR reserved-memory consistency bug fix fix: exclude fully reserved NUMA nodes consistently from topology hints, not just memory manager decisions #137280
   e.g.,

• topologymanager.Store.GetNUMANodeIDs()
• storing/exporting filtered NUMA IDs from TopologyManager
• CPU/Memory hint paths consuming that filtered NUMA set
• DeviceManager reading the filtered NUMA set from TopologyManager

What is happening now is this PR duplicates some common infrastructure that the reserved-memory PR also introduced. That is expected from an implementation standpoint.
If reserved-memory PR is likely to merge first, then rebase this PR on top of that and drop the duplicated plumbing.

Need to flag out for reviwewer that the overlap is only shared plumbing and that the reserved-memory semantics are intentionally not part of this PR.

Does this PR introduce a user-facing change?

Added an alpha TopologyManager policy option, restrict-to-cpu-numa-nodes, to limit topology hint generation to NUMA nodes with CPUs attached. This helps kubelet handle large CPU-less NUMA topologies such as NVIDIA Grace/Vera-class systems.

Additional documentation e.g., KEPs (Kubernetes Enhancement Proposals), usage docs, etc.:

NONE

[k8s-ci-robot]

Adding the "do-not-merge/release-note-label-needed" label because no release-note block was detected, please follow our release note process to remove it.

Details

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[k8s-ci-robot]

This issue is currently awaiting triage.

If a SIG or subproject determines this is a relevant issue, they will accept it by applying the triage/accepted label and provide further guidance.

The triage/accepted label can be added by org members by writing /triage accepted in a comment.

Details

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[k8s-ci-robot]

Hi @fanzhangio. Thanks for your PR.

I'm waiting for a kubernetes member to verify that this patch is reasonable to test. If it is, they should reply with /ok-to-test on its own line. Until that is done, I will not automatically test new commits in this PR, but the usual testing commands by org members will still work.

Tip

We noticed you've done this a few times! Consider joining the org to skip this step and gain /lgtm and other bot rights. We recommend asking approvers on your previous PRs to sponsor you.

Once the patch is verified, the new status will be reflected by the ok-to-test label.

I understand the commands that are listed here.

Details

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes-sigs/prow repository.

[k8s-ci-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: fanzhangio
Once this PR has been reviewed and has the lgtm label, please assign ffromani for approval. For more information see the Code Review Process.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• pkg/kubelet/cm/OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[fanzhangio]

@dims @klueska I opened this PR for introducing an alpha TopologyManager policy option to help kubelet handle GB/VR systems that expose many CPU-less NUMA nodes. This is the code implmentation of KEP-5726

[dims]

/ok-to-test

[dims]

/assign @ffromani @klueska

[fanzhangio]

/test pull-kubernetes-unit

[fanzhangio]

/retest

[klueska]

This needs a KEP if you are adding a new option to the user facing API.

I see a KEP issue (created by me), but not an actual KEP (unless I'm missing something).

Edit: I see a KEP was opened 5 hours ago. Please update the description with the proper links.

[fanzhangio]

Thanks @klueska , I've updated the KEP-5726 PR link

Also updated the PR description that this PR reuses some common infrastructure that a separate PR reserved-memory PR also introduced. reserved-memory consistency bug fix #137280

The overlap is only shared plumbing and that the reserved-memory semantics are intentionally not part of this PR.
If reserved-memory PR got merged first, we can rebase this PR on top of that and drop the duplicated plumbing.

[ffromani]

Thanks. This is 1.37 material, and the implementation may require rework depending on how the KEP review unfolds.