Reproduce “Compiling Specialized GPU-Initiated Networking” 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
Javid Baydamirli, Ismayil Ismayilov, Kaan Oktay, Didem Unat
OBALETH’s compiler treats network operations as part of GPU kernels, enabling it to batch submissions, reorder payloads before signals, and plan completion around buffer reuse. On a 256-GPU H100 all-to-all at 256 bytes per peer, its round latency is 1.25× faster than NVSHMEM and 2.3× faster than NCCL GIN. This result uses host-forwarded doorbells; advantages narrow for larger payloads.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.