Reproduce “Athena: Synergizing Data Prefetching and Off-Chip Prediction via Online Reinforcement Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rahul Bera, Zhenrong Lang, Caroline Hengartner, Konstantinos Kanellopoulos, Rakesh Kumar, Mohammad Sadrosadati, Onur Mutlu

Athena, an online reinforcement learning framework from ETH Zürich, dynamically coordinates data prefetching and off-chip prediction in high-performance processors. This framework yielded an average performance improvement of up to 14.9% over naive combinations in complex single-core configurations and 9.7% in eight-core systems, demonstrating particular strength in bandwidth-constrained environments.

Key insights

  • Data prefetching and off-chip prediction offer complementary performance benefits; OCPs can provide gains even in scenarios where prefetchers degrade performance.
  • A simple concurrent activation of prefetching and OCP ('Naive' combination) often fails to realize full potential and can even degrade performance, underscoring the necessity for intelligent coordination.
  • A state-aware reinforcement learning agent, combined with a carefully designed composite reward, enables robust, dynamic adaptation to varying workload behaviors and system constraints, significantly outperforming static or less informed approaches.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.