Plenary: Performance Matters
summary
Shows that CPU layout noise can swamp compiler gains: repeated LLVM layout randomization (Stabilizer) plus ANOVA finds -O2 vs -O3 indistinguishable from noise, so relying on optimizers is brittle. Introduces causal profiling with virtual speedups (Coz) that marks progress points, slows competitors to simulate speeding a target, and predicts real end-to-end impact. Uses this to find tiny edits that quickly improve real apps — Memcached +9%, SQLite +25%, six others up to +68% — often changing <10 lines and taking under 30 minutes. Coz ships in standard Linux distros.
AI-written summary · about
chapters (9)
- 0:00Introduction to performance
- 7:28End of automatic speedups
- 11:17Why performance analysis is hard
- 23:40Eliminating layout bias
- 33:10Testing compiler optimizations
- 40:20Limitations of traditional profilers
- 46:21Introducing causal profiling
- 56:49Unexpected optimization opportunities
- 1:00:20Conclusion and Q&A
related talks
01h 30mTuning C++: Benchmarks, and CPUs, and Compilers! Oh My!
01h 37mAn Overview of Program Optimization Techniques
01h 06m“Improving Performance Through Compiler Switches..."
51mHow Profile-Guided Optimization Makes Your Code Faster Without Any Code Changes
01h 27mDangerous Optimizations! in C++ - Part Deux
57mC++ MythBusters Strike 2
49mAn Engineering Approach to Optimising C++
01h 04mDesign for Performance