Reproduce “Performance Evaluation of Advanced Features in CUDA Unified Memory” 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
Steven W. D. Chien, Ivy B. Peng, Stefano Markidis
This study from KTH Royal Institute of Technology and Lawrence Livermore National Laboratory empirically evaluates the performance of advanced CUDA Unified Memory features, specifically memory advises and asynchronous prefetch, across diverse heterogeneous architectures. It finds that their effectiveness is highly dependent on the underlying hardware (e.g., PCIe vs. NVLink interconnect) and memory pressure (in-memory vs. oversubscription), revealing both performance gains and unexpected degradations.
Key insights
- The performance impact of advanced CUDA UM features is highly conditional, varying significantly with the specific hardware architecture and whether memory is oversubscribed.
- Memory advises can yield substantial performance improvements (up to 70%) for in-memory workloads, particularly on NVLink-connected Power9 systems, by optimizing data placement.
- Under memory oversubscription, advises can paradoxically degrade performance on NVLink systems, causing severe page thrashing, while offering modest benefits on PCIe systems.
- Asynchronous prefetching generally improves performance by eliminating page faults, often showing stronger benefits on PCIe systems compared to NVLink, especially during in-memory executions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.