CONCEPT Cited by 1 source

CPU Manager static policy¶

Definition¶

The CPU Manager is a Kubernetes kubelet feature that governs how pods receive CPU time. It supports two policies:

none (default) — standard CFS bandwidth control: pods get CPU time via cgroups but can be scheduled on any core at any time.
static — pods in the Guaranteed QoS class with integer CPU requests receive exclusive cpusets: the kubelet pins them to specific physical CPUs and no other workload (including system daemons) can preempt them on those CPUs.

The static policy is the Kubernetes-native way to achieve hyperthread-aware placement without hand-rolling taskset commands or sidecars.

By default, the kernel CFS scheduler can move a pod between cores freely, which:

Breaks cache locality — L1/L2 caches warmed on one core are cold after migration.
Can land the pod on a sibling hyperthread of a busy core — the source of the HT softirq contention effect.
Introduces run-queue latency variance unrelated to the workload itself.

The static policy takes these scheduling decisions out of the kernel's hands for Guaranteed pods, granting them exclusive physical CPUs.

Requires pods to request integer whole CPUs — cpu: "1.5" does not qualify. Fractional-CPU workloads still share.
Shrinks the shared CPU pool for the rest of the node — fewer pods per node if many use Guaranteed + integer CPU.
Requires careful reservation of cores for kubelet, container runtime, and kernel threads via --reserved-cpus.
HT-aware allocation: with CPU Manager static policy and the full-pcpus-only option, the kubelet allocates entire physical cores (both sibling HTs together) rather than individual logical CPUs — the direct fix for concepts/hyperthread-softirq-contention.

sources/2020-06-23-zalando-pgbouncer-on-kubernetes-minimal-latency — Zalando cites CPU Manager as the operator-level mitigation for the HT-softirq effect they observed: "It could be beneficial to configure CPU manager in the cluster, so that this would not be an issue." Concrete motivation for enabling static policy: don't let PgBouncer pods land on sibling hyperthreads of already-busy cores.