2025/04/24 by Yang Zhou, Zhou, Yang, Zhongjie Chen +25 · 6 citations
Computer Science · #Cloud Computing and Resource Management #Distributed and Parallel Computing Systems #FOS: Computer and information sciences #Networking and Internet Architecture (cs.NI)
paper · pdf · doi:10.48550/arxiv.2504.17307
openalex publication_date 2025/04/24 · openalex created_date 2025/10/18 · openalex updated_date 2026/07/28
Fast-evolving machine learning (ML) workloads have increasing requirements for networking. However, host network transport on RDMA NICs is hard to evolve, causing problems for ML workloads. For example, single-path RDMA traffic is prone to flow collisions that severely degrade collective communication performance. We present UCCL, an extensible software transport layer to evolve GPU networking. UCCL decouples the data path and control path of existing RDMA NICs and efficiently runs the control-path transport on host CPUs. This software extensibility brings in transport innovations that cannot be achieved in hardware for ML workloads, e.g., a multipath transport to resolve flow collisions. ML collectives atop UCCL achieve up to 4.5x higher performance compared to existing RDMA NICs.