Maximizing Workload's Performance With Smarter Runtimes
summary
Noisy-neighbor mitigation improves by moving allocation logic into the runtime and adding pluggable algorithms tested in a working project CRI-Resource-Manager that adjusts CPU cores, cache classes, NUMA/memory nodes, memory-bandwidth and Block I/O assignments and shows measurable performance gains on heterogeneous hardware. Proposal extends runtime integration via NRI plugin API over gRPC/ttrpc, requires runtimes to honor plugin replies, and asks CRI to evolve with CRI changes: pass abstract requirements and group ops plus a container-state subscription API.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:38Performance Optimization
- 6:58CPU Architecture
- 8:40Linux Architecture
- 11:54Caches
- 13:53Kubernetes
- 15:47Pluggable Policies
- 17:51Node Resource Interface
- 22:07Summary
- 23:34Outro
related talks
42mEnhance Workload QoS with Pluggable and Customizable Smarter Runtimes
36mMitigating Noisy Neighbours: Advanced Container Resource Management
21mDo Not Disturb Mode: Fine-Tuning Resources for Latency Sensitive Workloads
36mOptimizing Kubernetes for Low-Latency and Real-Time Containers
30mDeep Dive: CRI- RM Based CPU and NUMA Affinity to Achieve AI Task Accele...
34mUnlocking Performance: Topology-Aware CPU Scheduling W...
36mClass Resources: Kubernetes’ Fastest Way Of Shushing Noisy Neighbors
35mResource Management WG Deep Dive