Reproduce “KREX: Concurrent Kernel Benchmarking on Shared GPUs via Region-Granular Exclusivity” 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

Tianyu Feng, Haoxuan Yu, Tianyuan Wu, Lingyun Yang, Daocheng Ying, Yuxiao Wang, Ruibo Fan, Yinghao Yu and 3 more

KREX enables GPU benchmark commands to share devices while protecting author-marked kernel-timing regions. It blocks competing submissions, drains work, pauses sibling processes, and reuses persistent GPU contexts. In saturated H20 replays, it completed 54.8 commands per minute per eight-GPU node versus 16.0 with native exclusivity, with 2.5% p95 timing inflation for kernels over 1 ms; author-supplied boundaries are not validated.

Summary by alphaXiv.

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