3570-cpumanager

CPU Manager

Release Signoff Checklist

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

Summary

The CPU Manager is a new software component in Kubelet responsible for assigning pod containers to sets of CPUs on the local node. In later phases, the scope will expand to include caches, a critical shared processor resource.

The kuberuntime notifies the CPU manager when containers come and go. The first such notification occurs in between the container runtime interface calls to create and start the container. The second notification occurs after the container is stopped by the container runtime. The CPU Manager writes CPU settings for containers using a new CRI method named UpdateContainerResources. This new method is invoked from two places in the CPU manager: during each call to AddContainer and also periodically from a separate reconciliation loop.

This KEP supersedes and replaces kubernetes/enhancements/keps/sig-node/375-cpumanager/README.md.

Motivation

Poor or unpredictable performance observed compared to virtual machine based orchestration systems. Application latency and lower CPU throughput compared to VMs due to cpu quota being fulfilled across all cores, rather than exclusive cores, which results in fewer context switches and higher cache affinity.
Unacceptable latency attributed to the OS process scheduler, especially for “fast” virtual network functions (want to approach line rate on modern server NICs.)

Goals

Provide an API-driven contract from the system to a user: "if you are a Guaranteed pod with 1 or more cores of cpu, the system will try to make sure that the pod gets its cpu quota primarily from reserved core(s), resulting in fewer context switches and higher cache affinity".
Support the case where in a given pod, one container is latency-critical and another is not (e.g. auxiliary side-car containers responsible for log forwarding, metrics collection and the like.)
Do not cap CPU quota for guaranteed containers that are granted exclusive cores, since that would be antithetical to (1) above.
Take physical processor topology into account in the CPU affinity policy.

Non-Goals

N/A

Proposal

CPU Manager block diagram. Policy, State, and Topology types are factored out of the CPU Manager to promote reuse and to make it easier to build and test new policies. The shared state abstraction allows other Kubelet components to be agnostic of the CPU manager policy for observability and checkpointing extensions.

User Stories (Optional)

Story 1 : High-performance applications

Systems such as real-time trading system or 5G CNFs (User Plane Function, UPF) need to maximize the CPU time; CPU pinning ensure exclusive CPU allocation and allows to avoid performance issues due to core switches, cold caches. NUMA aware allocation of CPUs, provided by CPU manager cooperating with Topology Manager, is also a critical prerequisite for these applications to meet their performance requirement. The alignment of resources on the same NUMA node, CPUs first and foremost, prevents performance degradation due to inter-node (between NUMA nodes) communication overhead.

Story 2 : KubeVirt

KubeVirt leverages the CPU pinning provided by CPU manager to assign full CPU cores to vCPUs inside the VM to enhance performance. NUMA support for VMs is also built on top of the CPU pinning and NUMA-aware CPU allocation.

Notes/Constraints/Caveats (Optional)

N/A

Risks and Mitigations

Bugs in cpumanager can cause the kubelet to crash, or workloads to start with incorrect pinning. This can be mitigated with comprehensive testing and improving the observability of the system (see metrics).

While the cpumanager core policy has seen no changes except for bugfixes since a while, we introduced the cpumanager options policy framework to enable the fine tuning of the static policy. This area is more active, so bugs introduced with policy options can cause the kubelet to crash. To mitigate this risk, we can make sure each policy option can be disabled independently, and is not coupled with others, avoiding cascading failures or unnecessary coupling. Graduation and testing criteria are deferred to the KEPs tracking the implementation of these features.

Design Details

Discovering CPU topology

The CPU Manager must understand basic topology. First of all, it must determine the number of logical CPUs (hardware threads) available for allocation. On architectures that support hyper-threading, sibling threads share a number of hardware resources including the cache hierarchy. On multi-socket systems, logical CPUs co-resident on a socket share L3 cache. Although there may be some programs that benefit from disjoint caches, the policies described in this proposal assume cache affinity will yield better application and overall system performance for most cases. In all scenarios described below, we prefer to acquire logical CPUs topologically. For example, allocating two CPUs on a system that has hyper-threading turned on yields both sibling threads on the same physical core. Likewise, allocating two CPUs on a non-hyper-threaded system yields two cores on the same socket.

Decision: Initially the CPU Manager will re-use the existing discovery mechanism in cAdvisor.

Alternate options considered for discovering topology:

Read and parse the virtual file /proc/cpuinfo and construct a convenient data structure.
Execute a simple program like lscpu -p in a subprocess and construct a convenient data structure based on the output. Here is an example of data structure to represent CPU topology in go. The linked package contains code to build a ThreadSet from the output of lscpu -p.
Execute a mature external topology program like mpi-hwloc -- potentially adding support for the hwloc file format to the Kubelet.

CPU Manager interfaces (sketch)

type State interface {
  GetCPUSet(containerID string) (cpuset.CPUSet, bool)
  GetDefaultCPUSet() cpuset.CPUSet
  GetCPUSetOrDefault(containerID string) cpuset.CPUSet
  SetCPUSet(containerID string, cpuset CPUSet)
  SetDefaultCPUSet(cpuset CPUSet)
  Delete(containerID string)
}

type Manager interface {
  Start(ActivePodsFunc, status.PodStatusProvider, runtimeService)
  AddContainer(p *Pod, c *Container, containerID string) error
  RemoveContainer(containerID string) error
  State() state.Reader
}

type Policy interface {
  Name() string
  Start(s state.State)
  AddContainer(s State, pod *Pod, container *Container, containerID string) error
  RemoveContainer(s State, containerID string) error
}

type CPUSet map[int]struct{} // set operations and parsing/formatting helpers

type CPUTopology // convenient type for querying and filtering CPUs

Configuring the CPU Manager

Kubernetes will ship with CPU manager policies. Only one policy is active at a time on a given node, chosen by the operator via Kubelet configuration. The policies are none and static.

The active CPU manager policy is set through a new Kubelet configuration value --cpu-manager-policy. The default value is none.

The CPU manager periodically writes resource updates through the CRI in order to reconcile in-memory cpuset assignments with cgroupfs. The reconcile frequency is set through a new Kubelet configuration value --cpu-manager-reconcile-period. If not specified, it defaults to the same duration as --node-status-update-frequency (which itself defaults to 10 seconds at time of writing.)

Each policy is described below.

Policy 1: "none" cpuset control [default]

This policy preserves the existing Kubelet behavior of doing nothing with the cgroup cpuset.cpus and cpuset.mems controls. This "none" policy would become the default CPU Manager policy until the effects of the other policies are better understood.

Policy 2: "static" cpuset control

The "static" policy allocates exclusive CPUs for containers if they are included in a pod of "Guaranteed" QoS class and the container's resource limit for the CPU resource is an integer greater than or equal to one. All other containers share a set of CPUs.

When exclusive CPUs are allocated for a container, those CPUs are removed from the allowed CPUs of every other container running on the node. Once allocated at pod admission time, an exclusive CPU remains assigned to a single container for the lifetime of the pod (until it becomes terminal.)

The Kubelet requires the total CPU reservation from --kube-reserved and --system-reserved to be greater than zero when the static policy is enabled. This is because zero CPU reservation would allow the shared pool to become empty. The set of reserved CPUs is taken in order of ascending physical core ID. Operator documentation will be updated to explain how to configure the system to use the low-numbered physical cores for kube-reserved and system-reserved cgroups.

Workloads that need to know their own CPU mask, e.g. for managing thread-level affinity, can read it from the virtual file /proc/self/status:

$ grep -i cpus /proc/self/status
Cpus_allowed:   77
Cpus_allowed_list:      0-2,4-6

Note that containers running in the shared cpuset should not attempt any application-level CPU affinity of their own, as those settings may be overwritten without notice (whenever exclusive cores are allocated or deallocated.)

CPU Manager options

CPUManagerPolicyOptions allow to fine-tune the behavior of the static policy. The details of each option are described in their own KEP. As for kubernetes 1.26, the following options are available:

Implementation sketch

The static policy maintains the following sets of logical CPUs:

SHARED: Burstable, BestEffort, and non-integral Guaranteed containers run here. Initially this contains all CPU IDs on the system. As exclusive allocations are created and destroyed, this CPU set shrinks and grows, accordingly. This is stored in the state as the default CPU set.
RESERVED: A subset of the shared pool which is not exclusively allocatable. The membership of this pool is static for the lifetime of the Kubelet. The size of the reserved pool is the ceiling of the total CPU reservation from --kube-reserved and --system-reserved. Reserved CPUs are taken topologically starting with lowest-indexed physical core, as reported by cAdvisor.
ASSIGNABLE: Equal to SHARED - RESERVED. Exclusive CPUs are allocated from this pool.
EXCLUSIVE ALLOCATIONS: CPU sets assigned exclusively to one container. These are stored as explicit assignments in the state.

When an exclusive allocation is made, the static policy also updates the default cpuset in the state abstraction. The CPU manager's periodic reconcile loop takes care of updating the cpuset in cgroupfs for any containers that may be running in the shared pool. For this reason, applications running within exclusively-allocated containers must tolerate potentially sharing their allocated CPUs for up to the CPU manager reconcile period.

func (p *staticPolicy) Start(s State) {
	fullCpuset := cpuset.NewCPUSet()
	for cpuid := 0; cpuid < p.topology.NumCPUs; cpuid++ {
		fullCpuset.Add(cpuid)
	}
	// Figure out which cores shall not be used in shared pool
	reserved, _ := takeByTopology(p.topology, fullCpuset, p.topology.NumReservedCores)
	s.SetDefaultCPUSet(fullCpuset.Difference(reserved))
}

func (p *staticPolicy) AddContainer(s State, pod *Pod, container *Container, containerID string) error {
  if numCPUs := numGuaranteedCPUs(pod, container); numCPUs != 0 {
    // container should get some exclusively allocated CPUs
    cpuset, err := p.allocateCPUs(s, numCPUs)
    if err != nil {
      return err
    }
    s.SetCPUSet(containerID, cpuset)
  }
  // container belongs in the shared pool (nothing to do; use default cpuset)
  return nil
}

func (p *staticPolicy) RemoveContainer(s State, containerID string) error {
  if toRelease, ok := s.GetCPUSet(containerID); ok {
    s.Delete(containerID)
    s.SetDefaultCPUSet(s.GetDefaultCPUSet().Union(toRelease))
  }
  return nil
}

Example pod specs and interpretation

Pod	Interpretation
Pod [Guaranteed]: A: cpu: 0.5	Container A is assigned to the shared cpuset.
Pod [Guaranteed]: A: cpu: 2.0	Container A is assigned two sibling threads on the same physical core (HT) or two physical cores on the same socket (no HT.) The shared cpuset is shrunk to make room for the exclusively allocated CPUs.
Pod [Guaranteed]: A: cpu: 1.0 B: cpu: 0.5	Container A is assigned one exclusive CPU and container B is assigned to the shared cpuset.
Pod [Guaranteed]: A: cpu: 1.5 B: cpu: 0.5	Both containers A and B are assigned to the shared cpuset.
Pod [Burstable]	All containers are assigned to the shared cpuset.
Pod [BestEffort]	All containers are assigned to the shared cpuset.

Example scenarios and interactions

A container arrives that requires exclusive cores.
1. Kuberuntime calls the CRI delegate to create the container.
2. Kuberuntime adds the container with the CPU manager.
3. CPU manager adds the container to the static policy.
4. Static policy acquires CPUs from the default pool, by topological-best-fit.
5. Static policy updates the state, adding an assignment for the new container and removing those CPUs from the default pool.
6. CPU manager reads container assignment from the state.
7. CPU manager updates the container resources via the CRI.
8. Kuberuntime calls the CRI delegate to start the container.
A container that was assigned exclusive cores terminates.
1. Kuberuntime removes the container with the CPU manager.
2. CPU manager removes the container with the static policy.
3. Static policy adds the container's assigned CPUs back to the default pool.
4. Kuberuntime calls the CRI delegate to remove the container.
5. Asynchronously, the CPU manager's reconcile loop updates the cpuset for all containers running in the shared pool.
The shared pool becomes empty.
1. This cannot happen. The size of the shared pool is greater than the number of exclusively allocatable CPUs. The Kubelet requires the total CPU reservation from --kube-reserved and --system-reserved to be greater than zero when the static policy is enabled. The number of exclusively allocatable CPUs is floor(capacity.cpu - allocatable.cpu) and the shared pool initially contains all CPUs in the system.

Test Plan

[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: 20220929 - 86.2%

Integration tests

N/A

e2e tests

k8s.io/kubernetes/test/e2e_node/cpu_manager_test.go

Graduation Criteria

Alpha

Feature implemented behind a feature flag
Initial e2e tests completed and enabled

Beta

Gather feedback from developers and surveys
Complete features A, B, C
Additional tests are in Testgrid and linked in KEP

GA

N examples of real-world usage
N installs
More rigorous forms of testing—e.g., downgrade tests and scalability tests
Allowing time for feedback

Note: Generally we also wait at least two releases between beta and GA/stable, because there's no opportunity for user feedback, or even bug reports, in back-to-back releases.

For non-optional features moving to GA, the graduation criteria must include conformance tests.

Deprecation

Announce deprecation and support policy of the existing flag
Two versions passed since introducing the functionality that deprecates the flag (to address version skew)
Address feedback on usage/changed behavior, provided on GitHub issues
Deprecate the flag

Upgrade / Downgrade Strategy

No impact. It's always possible to trivially downgrade to the previous kubelet

Version Skew Strategy

Not relevant

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: CPUManager
- Components depending on the feature gate: kubelet

NOTE: in order to enable the feature, the cluster admin needs also to enable the static cpu manager policy.

Does enabling the feature change any default behavior?

No, unless the non-none policy (static) is explicitly configured.

Can the feature be disabled once it has been enabled (i.e. can we roll back the enablement)?

Yes, using the kubelet config.

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

The impact is node-local only. If the state of a node is steady, no changes. If a guaranteed pod is admitted, running non-guaranteed pods will have their CPU cgroup changed while running.

Are there any tests for feature enablement/disablement?

Yes, covered by e2e tests

Rollout, Upgrade and Rollback Planning

How can a rollout or rollback fail? Can it impact already running workloads?

A rollout can fail if a bug in the cpumanager prevents new pods to start, or existing pods to be restarted. Already running workload will not be affected if the node state is steady

What specific metrics should inform a rollback?

"cpu_manager_pinning_errors_total". It must be noted that even in fully healthy system there are known benign condition that can cause CPU allocation failures. Few selected examples are:

requesting odd numbered cores (not a full physical core) when the cpumanager is configured with the full-pcpus-only option
requesting NUMA-aligned cores, with Topology Manager enabled.

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

No to both. Changes in behavior only affects pods meeting the conditions (guaranteed QoS, integral CPU request) scheduler after the upgrade. Running pods will be unaffected by any change. This offers some degree of safety in both upgrade->rollback and upgrade->downgrade->upgrade scenarios.

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

No

Monitoring Requirements

Monitor the metrics

"cpu_manager_pinning_requests_total"
"cpu_manager_pinning_errors_total"

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

In order for pods to request exclusive CPUs allocation and pinning, they need to match all the following criteria:

the pod QoS must be "guaranteed"
the resources request of CPU (pod.spec.containers[].resources.limits.cpu) must be integral.

On top of that, at kubelet level

the cpumanager policy must be static.

If all the criteria are met, then the feature is in use by workloads.

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

Other (treat as last resort)
- Details: check the kubelet metric cpu_manager_pinning_requests_total

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

"cpu_manager_pinning_requests_total" and "cpu_manager_pinning_errors_total" We need to find a careful balance here because we don't want to leak hardware details, or in general informations dependent on the worker node hardware configuration (example, even if arguable extreme, is the processor core layout).

It is possible to infer which pod would trigger a CPU pinning from the pod resources request but adding these two metrics is both very cheap and helping for the observability of the system.

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

Metrics
- Metric name:
  - cpu_manager_pinning_requests_total
  - cpu_manager_pinning_errors_total

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

"cpu_manager_pinning_requests_total"
"cpu_manager_pinning_errors_total"

The addition of these metrics will be done before moving to GA (issue, PR).

Dependencies

None