Let There Be Topology-Awareness in Kube-Scheduler!
summary
Kubernetes scheduler today ignores NUMA-aligned resources so a pod requesting 4 CPUs and 4 SR-IOV VFs can land on a node that looks free (40 CPUs, 8 devices split 20/4 per NUMA) but fails the node’s Topology Manager and enters a topology affinity error loop. The project adds a pod resource API extension (cpu/device topology), a node resource topology CRD exported by Node Feature Discovery, and a topology aware scheduler plugin that scores nodes by NUMA allocatables. KEPs and demo code exist, plugin and NFD integration remain under review.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:43Agenda
- 1:25What is NUMA
- 3:26What is Normal Alignment
- 5:33Whats the Gap
- 7:30How the Default Scheduler Works
- 9:40Proposed Solution
- 14:01Current Status
- 15:32Use Cases
related talks
34mUnlocking Performance: Topology-Aware CPU Scheduling W...
26mResource Orchestration of HPC on Kubernetes: Where We Are Now and the Journey Ahead!
47mNHD - A Topology-Aware Scheduler for K8s for Low-Latency & HPC Applications
23mThis is The Way: A Crash Course on the Intricacies of Managing CPUs...
24mAccelerate Your AI/ML Workloads With Topology-Aware Scheduling in Kueue
36mNetwork-aware Scheduling in Kubernetes
23mOptimizing Storage Assignment via Pod Scheduling Under Disturbance Factors
20mBoosting Training and Inference Performance via Topology-Aware Scheduling of Heterogeneous Resources