STrack: A Reliable Multipath Transport for AI/ML Clusters

TL;DR

STrack is a hardware-offloaded reliable transport protocol designed to enhance distributed AI/ML training by optimizing congestion control, load balancing, and supporting out-of-order delivery in multipath environments, significantly outperforming existing solutions.

Contribution

The paper introduces STrack, a novel hardware-assisted transport protocol that improves reliability and performance for AI/ML workloads in multipath GPU clusters, with innovative congestion and load balancing features.

Findings

STrack outperforms RoCEv2 by up to 6X with synthetic workloads.

STrack achieves 27.4% performance improvement with collective workloads.

Hardware offloading enables faster loss recovery and out-of-order delivery.

Abstract

Emerging artificial intelligence (AI) and machine learning (ML) workloads present new challenges of managing the collective communication used in distributed training across hundreds or even thousands of GPUs. This paper presents STrack, a novel hardware-offloaded reliable transport protocol aimed at improving the performance of AI /ML workloads by rethinking key aspects of the transport layer. STrack optimizes congestion control and load balancing in tandem: it incorporates an adaptive load balancing algorithm leveraging ECN, while adopts RTT as multi-bit congestion indicators for precise congestion window adjustment. Additionally, STrack facilitates out-of-order delivery, selective retransmission, and swift loss recovery in hardware for multipath environment. The extensive simulation comparing STrack and RoCEv2 demonstrates that STrack outperforms RoCEv2 by up to 6X with synthetic workloads and by 27.4% with collective workloads, even with the optimized RoCEv2 system setup.