Wu Dexin
Compiler Optimization Expert at Kuaishou
Compiler Optimization Expert at Kuaishou with over six years of experience in C++ development and compiler technology. He specializes in compiler middle- and back-end optimization, intelligent resource allocation under complex constraints, and profile-guided optimizations based on runtime feedback, with a deep understanding of compiler theory and practice. He led the development of China’s first commercially deployed P4 network compiler system, overcoming multiple technical bottlenecks and delivering an efficient compilation solution for network data processing. Actively contributing to the LLVM open-source community, he proposed and implemented an innovative solution to address industry pain points in Propeller optimization—specifically its sensitivity to compiler flag changes and outdated profiles—resulting in significant performance gains for large-scale Kuaishou service projects.
Topic
Profile-Guided Compilation Optimization: Theory and Innovative Practice
Profile-guided optimization (PGO) is a key technology for improving application performance. Its core principle lies in leveraging real runtime data from applications—accurately capturing branch execution characteristics, function call frequencies, and other critical information observed in production environments. Compared with traditional static analysis, which relies solely on code structure for speculative optimization, PGO enables decisions that better reflect real-world execution behavior in areas such as function inlining, splitting, and code layout. This approach effectively avoids the inefficiencies caused by blind optimizations. Techniques such as AutoFDO, Bolt, and Propeller have demonstrated performance gains exceeding 10% in production environments. This talk will systematically explain the fundamental principles of PGO and share Kuaishou’s innovative practices in this area. It will also present real-world cases demonstrating how PGO has been deployed at scale and the measurable performance improvements achieved, providing practical insights for the broader industry. Outline: 1. Introduction to Profile-Guided Optimization (PGO) 2. Kuaishou’s Innovative Practices 3. Real-World Deployment and Performance Gains 4. Summary and Future Outlook
