Panel Discussion: State of Art for Enabling Performance Sensitive Workloads and What We Need Next
summary
Kubernetes lacks per-NUMA visibility so pods often land where CPUs, memory and PCI devices are misaligned, causing topology manager to reject pods and triggering runaway pod creation. Work items include a node-feature-discovery (NFD) topology updater that exposes per-NUMA allocatable resources as CRs, a node resource topology CRD and a topology-aware scheduler plugin that simulates topology-manager alignment. Runtime extensions (NRI) and kubelet pieces—CPU manager, device manager, memory manager and topology manager—are improved but many APIs and proposals remain in review; some custom schedulers were built and open sourced.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:39Panel Introductions
- 4:50Technical Details
- 10:54Complex Setup
- 15:30Scheduling Gaps
- 20:27Component List
- 26:29Lower Layers
- 32:36Resource Management
- 37:20Whats Next
related talks
27mWhat’s in the Box? Resource Management in Kubernetes
42mEnhance Workload QoS with Pluggable and Customizable Smarter Runtimes
25mTune It Up! Enabling Low Latency in Kubernetes Clusters
34mDemystifying Kubernetes Resource Management
32mOptimizing Application Performance on Kubernetes
36mMitigating Noisy Neighbours: Advanced Container Resource Management
34mKubernetes DRA Architecture: Scheduling, Status, and Topology at Scale
38mDynamic Pod Resource Boundary Adjustment in Web Scale Clusters