CPU Utilization: The Hidden Cost of Running Hot
summary
Claims that CPU utilization predicts latency and capacity with a simple practical formula, validated on production-like load using a dedicated performance environment that routes ~0.1% traffic to isolated nodes. Measurement uses clock_gettime per-request CPU counters plus OS counters from /proc/stat, and experiments ramp CPU with stress-ng. Results show Gen‑5 and Gen‑4 latency climbs sharply past ~40% utilization (up to 50–60% slower), and a data‑driven sweet spot example sets a 61% threshold for a 40% acceptable degradation.
AI-written summary · about
chapters (10)
- 0:00Performance theory
- 2:23Performance at GitHub
- 6:38Growth and CPU trends
- 10:56Experimental methodology
- 14:17Measuring CPU performance
- 18:54CPU utilization and latency
- 23:12Hyperthreading impacts
- 26:12C-states and power management
- 29:00Finding the sweet spot
- 32:56Conclusion and takeaways
related talks
27mNeed for SPEED: Site Performance Efficiency, Evaluation and Decision
42mWhen the OS gets in the way
39mThrottling: New Developments in Application Performance with CPU Limits
15mOptimizing Application Performance on Kubernetes
21mHPC Green Computing: Relationships Between Throughput and Latency
26mStrategies for Mitigating Performance Interference in Cloud-Native Systems
29mThe Cat and Mouse Game of Infra Cost vs. Latency
51mProduction Profiling: What, Why and How