Principles and Methodologies for Serial Performance Optimization
summary
Reviewing 477 OSDI/SOSP papers from the past decade yields a claim that sequential optimizations fall into three principles—remove, replace, reorder—and eight actionable methodologies (batching, caching, precomputing, deferring, relaxation, contextualization, hardware specialization, layering). Two case studies target file/storage systems and kernel synchronization; applying caching and layering to a lock-reordering design boosts throughput by about 23% and 14% respectively. A fine-tuned model, SysGPT, produces more specific, feasible optimization suggestions than GPT‑4.
AI-written summary · about
chapters (11)
- 0:00Serial performance intro
- 1:45Lock contention bottleneck
- 3:57Unified optimization framework
- 5:03Core sequential principles
- 6:11Optimization methodologies
- 7:56Optimization frequency in practice
- 8:20Batching for optimization
- 9:23Deferring for performance
- 10:48Context for smarter decisions
- 11:53Practical optimization example
- 13:40Summary and tools
related talks
42mAdventures in Efficiency and Performance
51mDesigning for Performance
34mCPU-Level Optimizations
01h 50mUnlocking Modern CPU Power - Next-Gen C++ Optimization Techniques
45mThe Joy of Building Large Scale Systems
01h 04mUnlocking Performance Through Reverse Engineering
01h 02mFastware
47mPerformance: What’s in a Name?