Reproduce “Purlin: Separating Orchestration from the Datapath of Collectives” 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
Osayamen Jonathan Aimuyo, Swapnil Gandhi, Christos Kozyrakis
Purlin separates distributed LLM inference communication semantics, coordination, and hardware-specific data movement. It describes input-output layouts, derives coordination through SNAC, and delegates copying or reduction to replaceable GPU datapaths. In tests on eight-GPU NVLink systems, Purlin improves end-to-end request latency by 1.23–1.25× in one SGLang setup. The evaluation does not establish performance for larger systems or multi-node RDMA.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.