Throttling: New Developments in Application Performance with CPU Limits
Talkat KubeCon + CloudNativeCon North America 2019 · on Wednesday, 2019-11-20· at 16:25· Room 17AB - San Diego Convention Center, Mezzanine Level·
summary
Explains that CFS quota slices and per-CPU expiration cause unexpected throttling on high-core nodes: a leftover 1ms slice per CPU can cost up to 0.87 CPU loss on an 88-core host. Reproducer fib_test isolates the bug; a kernel bisect points to commit 512ac999 and subsequent patches that remove per‑CPU expiration are landed and backported (5.4+). Real results: p95 drops from >200ms to 13ms when quota is adjusted, and an 88‑core case goes from 137ms to 482ms CPU usage after the fix. Recommendations: monitor throttle percentage, upgrade kernels, use whole‑CPU quotas or raise limits.
AI-written summary · about
chapters (5)
- 0:00Start
- 0:49Throttling: New Developments in Application Performance with CPU Limits
- 2:15Who should care about CPU limits?
- 14:47Multithreaded calculation of the Fibonacci sequence
- 29:55Questions?
related talks
38mA Tale of Tail Latency: Understanding k18n CPU Requests and Limits for Sust. and Profit
24mCPU Burst: Getting Rid of Unnecessary Throttling, Achieving High CP...
37m73,000 Pods a Day, Lessons From Misadventures In Multi-Tenant
35mCPU Utilization: The Hidden Cost of Running Hot
13mHarvesting Spare CPU Resources in Container Systems
23mThis is The Way: A Crash Course on the Intricacies of Managing CPUs...
26mStrategies for Mitigating Performance Interference in Cloud-Native Systems
42mWhen the OS gets in the way