KEP-2902: CPUManager Policy Option to Distribute CPUs Across NUMA Nodes Instead of Packing Them

Implementation History
BETA Implementable
Created 2021-08-26
Updated 2025-01-31
Latest v1.33
Milestones
Alpha v1.23
Beta v1.33
Stable v1.35
Related Enhancements
  • keps/sig-node/2625-cpumanager-policies-thread-placement
  • keps/sig-node/3545-improved-multi-numa-alignment/
  • keps/sig-node/4176-cpumanager-spread-cpus-preferred-policy/
  • keps/sig-node/4540-strict-cpu-reservation
  • keps/sig-node/4622-topologymanager-max-allowable-numa-nodes/
  • keps/sig-node/4800-cpumanager-split-uncorecache/
Ownership
Owning SIG
SIG Node
Primary Authors
@klueska @swatisehgal

KEP-2902: Add CPUManager policy option to distribute CPUs across NUMA nodes instead of packing them

Release Signoff Checklist

Items marked with (R) are required prior to targeting to a milestone / release.

  • (R) Enhancement issue in release milestone, which links to KEP dir in kubernetes/enhancements (not the initial KEP PR)
  • (R) KEP approvers have approved the KEP status as implementable
  • (R) Design details are appropriately documented
  • (R) Test plan is in place, giving consideration to SIG Architecture and SIG Testing input (including test refactors)
    • e2e Tests for all Beta API Operations (endpoints)
    • (R) Ensure GA e2e tests for meet requirements for Conformance Tests
    • (R) Minimum Two Week Window for GA e2e tests to prove flake free
  • (R) Graduation criteria is in place
  • (R) Production readiness review completed
  • (R) Production readiness review approved
  • “Implementation History” section is up-to-date for milestone
  • User-facing documentation has been created in kubernetes/website , for publication to kubernetes.io
  • Supporting documentation—e.g., additional design documents, links to mailing list discussions/SIG meetings, relevant PRs/issues, release notes

Summary

Kubernetes 1.22 introduced a new framework for adding CPUManager policy options (#2625). These new policy options allow one to tweak the behaviour of a given CPUManager policy without the need to introduce an entirely new policy. Moreover, these policy options can build on one another, such that multiple tweaks can be made to a policy in an additive fashion. The first option introduced in conjunction with #2625 allows one to ensure that only full CPUs are allocated to a container, rather than handing out individual hyperthreads from each CPU to different containers. This KEP introduces a new CPUManager policy option to ensure that CPU allocations are evenly distributed across NUMA nodes in cases where more than one NUMA node is required to satisfy the allocation.

Motivation

By default, the CPUManager will pack CPUs onto one NUMA node until it is filled, with any remaining CPUs simply spilling over to the next NUMA node. This can cause undesired bottlenecks in parallel code relying on barriers (and similar synchronization primitivies), as this type of code tends to run only as fast as its slowest worker (which is slowed down by the fact that fewer CPUs are available on at least one NUMA node). By distributing CPUs evenly across NUMA nodes, application developers can more easily ensure that no single worker suffers from NUMA effects more than any other, improving the overall performance of these types of applications.

Goals

  • Enable parallel algorithms to run more efficiently when they request more CPUs than can be allocated by a single NUMA node

Non-Goals

  • Provide a general solution for all types of CPU distributions across NUMA nodes

Proposal

We propose to add a new CPUManager policy option called distribute-cpus-across-numa to the static CPUManager policy. When enabled, this will trigger the CPUManager to evenly distribute CPUs across NUMA nodes in cases where more than one NUMA node is required to satisfy the allocation.

Risks and Mitigations

The risks of adding this new feature are quite low. It is isolated to a specific policy option within the CPUManager, and is protected both by the option itself, as well as the CPUManagerPolicyBetaOptions feature gate (which is disabled by default).

RiskImpactMitigation
Bugs in the implementation lead to kubelet crashHighDisable the policy option and restart the kubelet. The workload will run but with CPU packing semantics - like it was before this new policy option was added.

Design Details

When distribute-cpus-across-numa is passed as a policy option, the following algorithm will be run to distribute CPUs across NUMA nodes instead of packing them:

Foreach NUMA node:
   * If all requested CPUs can be allocated from this single NUMA node;
   --> do the allocation

For each pair of NUMA nodes:
   * If the set of requested CPUs (modulo 2) can be evenly split across the 2 NUMA nodes; AND
   * Any remaining CPUs (after the modulo operation) can be striped across some subset of the NUMA nodes;
   --> do the allocation

For each 3-tuple of NUMA nodes:
   * If the set of requested CPUs (modulo 3) can be evenly distributed across the 3 NUMA nodes; AND
   * Any remaining CPUs (after the modulo operation) can be striped across some subset of the NUMA nodes;
   --> do the allocation

...

For the set of all NUMA nodes:
   * If the set of requested CPUs (module NUM_NUMA_NODES) can be evenly distributed across all NUMA nodes; AND
   * Any remaining CPUs (after the modulo operation) can be striped across some subset of the NUMA nodes;
   --> do the allocation

If none of the above conditions can be met, resort back to a best-effort fit of packing CPUs into NUMA nodes wherever they can fit.

NOTE: The striping operation after all CPUs have been evenly distributed will be performed such that the overall disribution of CPUs across those NUMA nodes remains as balanced as possible.

Compatibility with full-pcpus-only policy options

CompatibilityalphabetaGA
full-pcpus-onlyxxx

Test Plan

We will extend both the unit test suite and the E2E test suite to cover the new policy option described in this KEP.

[x] I/we understand the owners of the involved components may require updates to existing tests to make this code solid enough prior to committing the changes necessary to implement this enhancement.

Prerequisite testing updates
Unit tests
  • k8s.io/kubernetes/pkg/kubelet/cm/cpumanager: 20250205 - 85.5% of statements
Integration tests

Not Applicable as Kubelet features don’t have integration tests. We use a mix of e2e_node and e2e tests.

e2e tests

Currently no e2e tests are present for this particular policy option. E2E tests will be added as part of Beta graduation.

The plan is to add e2e tests to cover the basic flows for cases below:

  1. distribute-cpus-across-numa option is enabled: The test will ensure that the allocated CPUs are distributed across NUMA nodes according to the policy.
  2. distribute-cpus-across-numa option is disabled: The test will verify that the allocated CPUs are packed according to the default behavior.
  3. Test how this option interacts with full-pcpus-only policy option (and test for it enabled and disabled).

Graduation Criteria

Alpha

  • Implement the new policy option.
  • Ensure proper unit tests are in place.
  • Ensure proper e2e node tests are in place.

Beta

  • Gather feedback from consumers of the new policy option.
  • Verify no major bugs reported in the previous cycle.

GA

  • Allow time for feedback (1 year).
  • Make sure all risks have been addressed.

Upgrade / Downgrade Strategy

We expect no impact. The new policy option is opt-in and orthogonal to the existing ones.

Version Skew Strategy

No changes needed

Production Readiness Review Questionnaire

Feature Enablement and Rollback

How can this feature be enabled / disabled in a live cluster?
  • Feature gate (also fill in values in kep.yaml)
    • Feature gate name: CPUManagerPolicyOptions
    • Feature gate name: CPUManagerPolicyAlphaOptions
    • Feature gate name: CPUManagerPolicyBetaOptions
    • Components depending on the feature gate: kubelet
  • Change the kubelet configuration to set a CPUManager policy of static and a CPUManager policy option of distribute-cpus-across-numa
    • Will enabling / disabling the feature require downtime of the control plane? No
    • Will enabling / disabling the feature require downtime or reprovisioning of a node? (Do not assume Dynamic Kubelet Config feature is enabled). Yes – a kubelet restart is required.
Does enabling the feature change any default behavior?

No. In order to trigger any of the new logic, three things have to be true:

  1. The CPUManagerPolicyBetaOptions feature gate must be enabled
  2. The static CPUManager policy must be selected
  3. The new distribute-cpus-across-numa policy option must be selected
Can the feature be disabled once it has been enabled (i.e. can we roll back the enablement)?

Yes, the feature can be disabled by either:

  1. Disabling the CPUManagerPolicyBetaOptions feature gate
  2. Switching the CPUManager policy to none
  3. Removing distribute-cpus-across-numa from the list of CPUManager policy options

Existing workloads will continue to run uninterrupted, with any future workloads having their CPUs allocated according to the policy in place after the rollback.

What happens if we reenable the feature if it was previously rolled back?

No changes. Existing container will not see their allocation changed. New containers will.

Are there any tests for feature enablement/disablement?
  • A specific e2e test will demonstrate that the default behaviour is preserved when the feature gate is disabled, or when the feature is not used (2 separate tests)

Rollout, Upgrade and Rollback Planning

How can a rollout or rollback fail? Can it impact already running workloads?
  • A rollout or rollback can fail if the feature gate and the policy option are not configured properly and kubelet fails to start.
What specific metrics should inform a rollback?

As part of graduation of this feature, we plan to add metric cpu_manager_numa_allocation_spread to see how the CPUs are distributed across NUMA nodes. This can be used to see the CPU distribution across NUMA and will provide an indication of a rollback.

Were upgrade and rollback tested? Was the upgrade->downgrade->upgrade path tested?

Not Applicable. This policy option only affects pods that meet certain conditions and are scheduled after the upgrade. Running pods will be unaffected by any change.

Is the rollout accompanied by any deprecations and/or removals of features, APIs, fields of API types, flags, etc.?

No

Monitoring Requirements

How can an operator determine if the feature is in use by workloads?

Inspect the kubelet configuration of a node – check for the presence of the feature gate and usage of the new policy option.

In addition to that, we can check the metric cpu_manager_numa_allocation_spread to determine how allocated CPUs are spread across NUMA node.

How can someone using this feature know that it is working for their instance?

In order to verify this feature is working, one should:

  1. Pick an node with at least 2 NUMA nodes in your cluster
  2. Ensure no other pods with exclusive CPUs are running on that node
  3. Launch a pod with a nodeSelector to that node that has a single container in it
  4. Run a ´sleep infinity` command and request exclusive CPUs for the container in the amount of (1 + NUM_CPUS_PER_NUMA_NODE)
  5. Verify that the list of CPUs allocated to the container are evenly distributed across 2 NUMA nodes instead of packed

To verify the list of CPUs allocated to the container, one can either:

  • exec into uthe container and run taskset -cp 1 (assuming this command is available in the container).
  • Call the GetCPUS() method of the CPUProvider interface in the kubelet’s podresources API .

Also, we can check cpu_manager_numa_allocation_spread metric. We plan to add metric to track how CPUs are distributed across NUMA zones with labels/buckets representing NUMA nodes (numa_node=0, numa_node=1, …, numa_node=N).

With packed allocation (default, option off), the distribution should mostly be in numa_node=1, with a small tail to numa_node=2 (and possibly higher) in cases of severe fragmentation. Users can compare this spread metric with the container_aligned_compute_resources_count metric to determine if they are getting aligned packed allocation or just packed allocation due to implementation details.

For example, if a node has 2 NUMA nodes and a pod requests 8 CPUs (with no other pods requesting exclusive CPUs on the node), the metric would look like this:

cpu_manager_numa_allocation_spread{numa_node=“0”} = 8 cpu_manager_numa_allocation_spread{numa_node=“1”} = 0

When the option is enabled, we would expect a more even distribution of CPUs across NUMA nodes, with no sharp peaks as seen with packed allocation. Users can also check the container_aligned_compute_resources_count metric to assess resource alignment and system behavior.

In this case, the metric would show: cpu_manager_numa_allocation_spread{numa_node=“0”} = 4 cpu_manager_numa_allocation_spread{numa_node=“1”} = 4

Note: This example is simplified to clearly highlight the difference between the two cases. Existing pods may slightly skew the counts, but the general trend of peaks and troughs will still provide a good indication of CPU distribution across NUMA nodes, allowing users to determine if the policy option is enabled or not.

What are the reasonable SLOs (Service Level Objectives) for the enhancement?

There are no specific SLOs for this feature. Parallel workloads will benefit from this feature in application specific ways.

What are the SLIs (Service Level Indicators) an operator can use to determine the health of the service?

None

Are there any missing metrics that would be useful to have to improve observability of this feature?

Yes, as part of graduation of this feature to Beta, we plan to add cpu_manager_numa_allocation_spread metric to provide data on how the CPUs are distributed across NUMA nodes.

Does this feature depend on any specific services running in the cluster?

This feature is linux specific, and requires a version of CRI that includes the LinuxContainerResources.CpusetCpus field. This has been available since v1alpha2.

Dependencies

Does this feature depend on any specific services running in the cluster?

No

Scalability

Will enabling / using this feature result in any new API calls?

No

Will enabling / using this feature result in introducing new API types?

No

Will enabling / using this feature result in any new calls to the cloud provider?

No

Will enabling / using this feature result in increasing size or count of the existing API objects?

No

Will enabling / using this feature result in increasing time taken by any operations covered by existing SLIs/SLOs?

The algorithm required to implement this feature could delay:

  1. Pod admission time
  2. The time it takes to launch each container after pod admission

This delay should be minimal.

Will enabling / using this feature result in non-negligible increase of resource usage (CPU, RAM, disk, IO, …) in any components?

No, the algorithm will run on a single goroutine with minimal memory requirements.

Troubleshooting

How does this feature react if the API server and/or etcd is unavailable?

No impact. The behavior of the feature does not change when API Server and/or etcd is unavailable since the feature is node local.

What are other known failure modes?

Because of existing distribution of CPU resource across, a distributed allocation might not be possible. E.g. If all Available CPUs are present on the same NUMA node.

In that case we resort back to a best-effort fit of packing CPUs into NUMA nodes wherever they can fit.

What steps should be taken if SLOs are not being met to determine the problem?

Implementation History

  • 2021-08-26: Initial KEP created
  • 2021-08-30: Updates to fill out more sections, answer PRR questions
  • 2021-09-08: Change feature gate from CPUManagerPolicyOptions to CPUManagerPolicyExperimentalOptions
  • 2021-10-11: Change feature gate from CPUManagerPolicyExperimentalOptions to CPUManagerPolicyAlphaOptions
  • 2025-01-30: KEP update for Beta graduation of the policy option
  • 2025-02-05: KEP update to the latest template