A Tale of Tail Latency: Understanding k18n CPU Requests and Limits for Sust. and Profit
summary
Kubernetes CPU requests serve scheduling and also translate directly to CFS shares, so under contention containers get CPU time proportionally; limits map to cgroup quota (period=100ms) and cause throttling not eviction. Real customer data shows many containers waste requested CPU while ~10% use >2× their request, driving tail latency and energy waste. Track container CPU usage vs. kube pod/container resource requests and cgroup throttle metrics; use static CPU manager with integer requests==limits for exclusive cores and bench with limits to discover real needs.
AI-written summary · about
chapters (8)
- 0:00Welcome and introduction
- 1:52Why CPU sizing matters
- 5:05CPU requests and scheduling
- 10:11CPU time distribution with CFS
- 14:08Noisy neighbors and right sizing
- 18:21CPU limits and throttling
- 24:30CPU manager and static policy
- 31:20Key takeaways and Q&A
related talks
47mOptimizing Kubernetes Resource Requests/Limits for Cost-Efficiency and Latency
39mThrottling: New Developments in Application Performance with CPU Limits
23mThis is The Way: A Crash Course on the Intricacies of Managing CPUs...
37m73,000 Pods a Day, Lessons From Misadventures In Multi-Tenant
13mHarvesting Spare CPU Resources in Container Systems
47mCPU & Memory, Requests & Limits, QoS & CFS: The JVM in the Kubernetes Pod
24mFreedom of Kubernetes requires Chaos Engineering to shine in production
24mResource Requests and Limits Under the Hood: The Journey of a Pod Spec