UCCL-Tran: An Extensible Software Transport Layer for GPU Networking
Abstract
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-Tran, an extensible software transport layer to evolve GPU networking. UCCL-Tran 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-Tran achieve up to 4.5× higher performance compared to existing RDMA NICs.