Cover Image

An Extensible Software Transport Layer for GPU Networking

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Zhou, Yang, Chen, Zhongjie, Mao, Ziming, Lao, ChonLam, Yang, Shuo, Kannan, Pravein Govindan, Gao, Jiaqi, Zhao, Yilong, Wu, Yongji, You, Kaichao, Ren, Fengyuan, Xu, Zhiying, Raiciu, Costin, Stoica, Ion
Format:	Preprint
Published:	2025
Subjects:	Networking and Internet Architecture
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

_version_	1866911091252002816
author	Zhou, Yang Chen, Zhongjie Mao, Ziming Lao, ChonLam Yang, Shuo Kannan, Pravein Govindan Gao, Jiaqi Zhao, Yilong Wu, Yongji You, Kaichao Ren, Fengyuan Xu, Zhiying Raiciu, Costin Stoica, Ion
author_facet	Zhou, Yang Chen, Zhongjie Mao, Ziming Lao, ChonLam Yang, Shuo Kannan, Pravein Govindan Gao, Jiaqi Zhao, Yilong Wu, Yongji You, Kaichao Ren, Fengyuan Xu, Zhiying Raiciu, Costin Stoica, Ion
contents	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.
format	Preprint
id	arxiv_https___arxiv_org_abs_2504_17307
institution	arXiv
publishDate	2025
record_format	arxiv
spellingShingle	An Extensible Software Transport Layer for GPU Networking Zhou, Yang Chen, Zhongjie Mao, Ziming Lao, ChonLam Yang, Shuo Kannan, Pravein Govindan Gao, Jiaqi Zhao, Yilong Wu, Yongji You, Kaichao Ren, Fengyuan Xu, Zhiying Raiciu, Costin Stoica, Ion Networking and Internet Architecture 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.
title	An Extensible Software Transport Layer for GPU Networking
topic	Networking and Internet Architecture
url	https://arxiv.org/abs/2504.17307