TechPrufer
Get started
Cloudflare
Cloudflare logoCloudflare · Feature trail

LuaJIT Performance Benchmarking and Optimization

2 posts · Feb 2017 – Dec 2018

What this trail covers

Cloudflare is actively involved in improving LuaJIT performance through a research project at King's College London. This project focuses on developing a comprehensive benchmark suite for Lua implementations, analyzing the performance of different LuaJIT forks (original LuaJIT and RaptorJIT), and identifying areas for optimization. The work aims to improve VM warmup behavior and performance characteristics across various benchmarks, leading to more predictable and faster execution for larger pro

All Cloudflare trails
  1. More consistent LuaJIT performanceThis post details the progress of a year-long project to improve LuaJIT performance. Key contributions include the creation of a larger, more robust Lua benchmark suite, the development of a simple benchmark runner (simplerunner.lua) with statistical analysis (mean, confidence intervals), and a comparative performance analysis of the original LuaJIT repository and RaptorJIT. The post presents detailed benchmark results for both implementations and a comparison between them, highlighting performance differences across various benchmarks and providing classifications for their behavior (e.g., warmup, flat, slowdown).
  2. Helping to make LuaJIT fasterThis post details a research project funded by Cloudflare at King's College London to improve LuaJIT performance. It highlights the complexity of JIT compilers, the importance of rigorous benchmarking, and the hypothesis that heuristics in VMs interact in complex ways. The project leverages a new benchmark suite (Krun and warmup_stats) and the expertise of a LuaJIT specialist to address performance inconsistencies and long-term speed issues in larger LuaJIT programs. All improvements will be open-sourced.
  3. LuaJIT Hacking: Getting next() out of the NYI listThis post details the development of 'loom', a custom HTML-formatted dump tool for LuaJIT's JIT compiler output. It explains the internal representations of code in LuaJIT (source, bytecode, trace, IR, mcode) and demonstrates how 'loom' aids in understanding code execution and optimization by providing cross-references and improved formatting. The post also addresses the performance implications of 'NYI' (Not Yet Implemented) operations in LuaJIT, specifically focusing on the `next()` function used for table iteration, and outlines the effort to get `next()` out of the NYI list.