Performance Matters
summary
Explains why naive benchmarking lies and introduces Stabilizer, which randomizes memory and code layout continuously so execution times become normally distributed and allow valid t tests (showing -O2 vs -O3 is statistically indistinguishable). Describes Coz, a causal profiler that uses progress points and virtual speedups (slowing everything else) to predict latency and throughput gains. Reports guided fixes: Memcached +9%, SQLite +25%, six other apps up to +68%, usually by changing <10 lines in under 30 minutes; Coz ships via apt install coz-profiler.
AI-written summary · about
chapters (24)
- 0:00Intro
- 0:27Luke
- 1:01Prototype
- 1:21Flow
- 2:37Cool Hand Luke
- 3:06ASCII Art
- 3:23Back in the Day
- 4:39Performance Improvement
- 6:34App Store Updates
- 7:14Performance Analysis
- 18:52Null Hypothesis Testing
- 20:16Randomization
- 21:57Stabilizer
- 22:10Benchmarks
- 26:21Profilers
- 27:58Causal Profiles
- 31:45Audit
- 32:43Progress Points
- 34:12Causal Profiler
- 34:52Parallel Applications
- 36:47Deduplication
- 37:21Hash function
- 37:45Hash table
- 41:39Conclusion
related talks
02h 33mMeasure, don't Guess - How to find out if and where to optimize
01h 11m“Benchmarking C++ Code"
50mPerformance optimization: how do I go about it?
05mLightning Talk: Micro C++ Benchmarks & Murphy's Law
29mHow to Performance
47mPerformance optimisation: how do I go about it?
42mPerformance Paradox: Navigating the Complexities of Mobile App Optimisation
39mProfiling and Benchmarking 101