Beacons: An End-to-End Compiler Framework for Predicting and Utilizing Dynamic ...
summary
Compiler and runtime build a pipeline that instruments loops with “beacons” carrying predicted loop trip count, timing, memory footprint and reuse-distance, using LLVM static analysis + lightweight ML and dynamic specialization. Models predict trip counts with 85–95% accuracy and generate loop timing and cache-footprint estimates; a proactive scheduler uses those beacons to co-locate processes and avoid cache thrashing, yielding 1.9x average throughput (up to 3.2x) on Graviton2 for 1k–10k consolidated processes.
AI-written summary · about
chapters (10)
- 0:00Beacons compiler intro
- 3:53HPC scheduling challenges
- 7:45Static analysis overview
- 12:01Loop type categorization
- 13:27Loop trip count ML
- 15:16Nested loop timing
- 17:46Cache reuse scheduling
- 18:28Prediction accuracy
- 19:29Performance gains
- 20:10Conclusion
related talks
52mA bird's eye view of a compiler
14mOn-Stack Replacement for Program Generators and Source-to-Source Compilers
21mTowards Compiler-Guided Static Analysis
27mImproving Vectorization Heuristics in a Dynamic Compiler with Machine Learning Models
01h 16m“What Has My Compiler Done for Me Lately? Unbolting the Compiler's Lid”
20mRefined Input, Degraded Output: The Counterintuitive World of Compiler Behavior
39mFireside chat on JIT compiler
01h 04mJust-in-Time Compilation: The Next Big Thing?