Unlocking Performance: Topology-Aware CPU Scheduling W...
summary
A Kubernetes DRA-based CPU driver is implemented as a daemonset that publishes per-CPU resource slices with rich attributes and uses a kubelet plugin plus an NRI-based runtime plugin to enforce cgroup pinning via a generated CDI handoff; it requires disabling kubelet static CPU policy. Demo shows a pod requesting 20 CPUs and aligning CPUs, GPUs and RDMA to the same NUMA (previously CPUs and devices landed on different NUMA nodes), addressing the typical ~10% latency hit (up to 30–40% in some apps) while noting scheduler and node overhead tradeoffs.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:22Challenges in resource management
- 5:53DRA for native resources
- 12:12Demo and workflow
- 19:32Future work and use cases
- 22:55Questions and answers
related talks
18mLet There Be Topology-Awareness in Kube-Scheduler!
20mBoosting Training and Inference Performance via Topology-Aware Scheduling of Heterogeneous Resources
23mThis is The Way: A Crash Course on the Intricacies of Managing CPUs...
30mDeep Dive: CRI- RM Based CPU and NUMA Affinity to Achieve AI Task Accele...
47mNHD - A Topology-Aware Scheduler for K8s for Low-Latency & HPC Applications
21mUnlocking How To Efficiently, Flexibly, Manage and Schedule Seven AI Chips in Kubernetes
21mDo Not Disturb Mode: Fine-Tuning Resources for Latency Sensitive Workloads
24mPropagating Programming Paradigms: Lifting High Performance Compute Into K8s