Performance Really Matters
summary
Argues that traditional profiling and compiler optimizations no longer suffice and introduces two tools with concrete results: Coz causal profiler uses virtual speedup experiments to predict end-to-end impact and finds fixes that improve real apps up to 68%, often by changing under ten lines in under 30 minutes; Coz ships in standard Linux distros. Scalene is a fast, line-level Python profiler with a novel sampling allocator that reports memory, native vs Python time, and copy volume; available on PyPI.
AI-written summary · about
chapters (19)
- 0:00Start
- 3:52"Metal" Languages Irrational Exube
- 4:0440 Year Performance Ride!
- 7:54The Ride Is Over
- 8:20Performance in the '80s
- 8:35Modern-day performance
- 8:55Typical performance evaluation
- 9:17Which is faster?
- 13:08The Prototype Ogle
- 14:16Software Profilers
- 15:48What would speed up Ogle?
- 17:34Performance Experiments
- 18:36Virtual Speedup by slowing everything else down.
- 19:32Speedup Results
- 20:57Progress Points
- 22:58Coz: a Causal Profiler for Linux ships with Debian/Ubuntu
- 23:28Using Causal Profiling on Ogle
- 23:41Ferret
- 25:45Dedup Compression via deduplication
related talks
37mCausal's 1000x Spreadsheet; Performance beyond profiling
15mDivining Profiler Accuracy: An Approach to Approximate Profiler Accuracy Through(…)
14mVESPA: Static Profiling for Binary Optimization
21mDiagnosing Compiler Performance by Comparing Optimization Decisions
21mGetting the Most Bang out of your Buck with Compilers on POWER
06mProfile-Guided Optimization in the Go Compiler
19mOptimization-Aware Compiler-Level Event Profiling
22mBoosting Compiler Testing by Injecting Real-World Code