Kubernetes v1.37 elevates Pod-Level Resource Managers to beta, allowing the Kubelet to make NUMA-aware allocation decisions based on pod-level resource declarations. This approach improves efficiency by enabling hybrid allocation models where primary containers receive exclusive resources while sidecar containers operate in an isolated shared pool.
Key points
Pod-Level Resource Managers allow NUMA-aware resource allocation at the pod level rather than forcing individual container allocations.
This feature enables sidecars to run in a pod-isolated shared pool, avoiding the waste of dedicating physical cores to non-critical auxiliary containers.
The PodResources gRPC service now supports reporting exclusive pod-level CPU and memory assignments to improve observability.
The feature is gated by PodLevelResourceManagers and remains disabled by default during its beta status.