Reproduce “Optimizing ML Concurrent Computation and Communication with GPU DMA Engines” 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
Anirudha Agrawal, Shaizeen Aga, Suchita Pati, Mahzabeen Islam
Researchers at AMD systematically characterized interference in concurrent computation and communication (C3) on AMD Instinct MI300X GPUs, finding that offloading communication to GPU DMA engines substantially improves C3 performance. This approach achieved an average of 72% of the ideal C3 speedup, a considerable increase from the baseline 21%.
Key insights
- Mutual interference between concurrent GPU compute and communication kernels significantly limits C3 performance, with all-to-all collectives showing greater interference than all-gather.
- Software-driven strategies like schedule prioritization and resource partitioning can double C3 speedup, improving resource utilization and reducing contention for compute units.
- Offloading communication to dedicated GPU DMA engines is highly effective in mitigating compute and cache interference, demonstrating a superior approach to achieving high C3 efficiency for concurrent computation and communication (C3).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.