Probabilistic Analysis of WCETs on RT Systems: When Your Worst Case Isn't Worst
summary
A measurement toolkit applies EVT and the generalized Pareto distribution to PREEMPT_RT latency traces, selects thresholds via a stability plot, and reports converged parameters after 96 hours of sampling. It diagnoses sources of outliers (cache effects and kernel housekeeping), fits an extrapolatable model, and claims with 99.9% confidence that worst-case latency stays below 15 microseconds over extreme horizons, enabling data-backed SLAs rather than guesswork.
AI-written summary · about
related talks
49mShifting Safety Techniques To a Statistical World
14mTurboTest: Learning When Less is Enough through Early Termination of Internet Speed Tests
20mEvaluating Time-to-Event Models is Hard
38mEnough statistics so that Zed won't yell at you
12mCATO: End-to-End Optimization of ML-Based Traffic Analysis Pipelines
31mWhy Your TTFT Lies: Diagnosing PD-Disaggregated LLM Inference With Minimal Cross...
16mUnderstanding Stragglers in Large Model Training Using What-if Analysis
25mAI Reliability in Security: A New Data Science Paradigm