Extracting the Last Microseconds of Performance From Co...
summary
Claim: combining a container-specific Linux scheduler with CPU C‑state and frequency controls cuts wake‑up latency up to 99% in their tests. Implementation: a Node Resource Interface plugin set using containerd + OCI runtime spec and topology-aware balloons defined by CRDs to pin CPUs, set CPUfreq/C‑states and IO/scheduling classes. Benchmarks: cyclic sleep/wake tests on idle machines show a typical 15–20% improvement from scheduling class alone and down to ~2% of baseline jitter with full tuning. All plugins and tests are published and work as described.
AI-written summary · about
chapters (16)
- 0:00Real-time performance challenges
- 2:29Jitter reduction limits
- 3:33Linux & Kubernetes tools
- 4:59Node Resource Interface (NRI)
- 6:44NRI Balloons plugin
- 8:49OCI runtime controls
- 11:59Latency benchmarks
- 17:33Jitter reduction analysis
- 18:37Balloon configuration
- 20:26Hardware & kernel tuning
- 21:51NRI plugins & community
- 23:58Q&A testing methods
- 24:39NRI vs isolated cores
- 26:25Bare metal vs containers
- 27:29Q&A performance trade-offs
- 31:24Closing remarks
related talks
34mUnlocking Performance: Topology-Aware CPU Scheduling W...
53mA cloud cost saving journey: Strategies to balance CPU for containerized JAVA workloads in K8s
10mLightning Talk: Millisecond Scale-to-Zero and the Quest to Never Pay for Idle Again
24mBobtail: Avoiding Long Tails in the Cloud
25mFrom Cold Start to Warp Speed: Triton Kernel Caching with OCI Container images
29mWhat's Up With All The Different Container Runtimes? - Ricardo Aravena, Branch Metrics
18mLet There Be Topology-Awareness in Kube-Scheduler!
36mTaming the Beast: Advanced Resource Management With Kubernetes