Reproduce “OpenURMA: A Clean-Room Open Implementation of the Unified Bus Protocol” 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
Bojie Li
OpenURMA provides an open-source, clean-room implementation of Huawei's Unified Bus (UB) protocol, demonstrating its architectural advantages over traditional RDMA. It achieves a 4.37x lower latency for 64-byte remote fetches and a 4,855x smaller NIC state at scale compared to conventional RoCE, by integrating the NIC controller onto the CPU's on-chip bus.
Key insights
- The Unified Bus architecture fundamentally re-architects RDMA by moving the NIC controller onto the CPU's on-chip bus, thereby eliminating PCIe bottlenecks and enabling direct CPU load/store access to remote memory.
- UB decouples application-level state from per-host transport state, leading to O(N+M) state scaling that dramatically reduces NIC memory footprint compared to O(N*M) in traditional RDMA.
- The protocol offers a flexible, opt-in ordering surface, allowing applications to pay for consistency guarantees only when explicitly required, improving efficiency for unordered operations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.