STrack: A Reliable Multipath Transport for AI/ML Clusters
Yanfang Le, Rong Pan, Peter Newman, Jeremias Blendin, Abdul Kabbani, Vipin Jain, Raghava Sivaramu, Francis Matus
TL;DR
STrack tackles the scalability and reliability challenges of AI/ML collective communication over Ethernet by introducing a NIC-offloaded reliable multipath transport. It blends ECN-guided adaptive packet spraying with a RTT-informed window controller and a robust loss-recovery stack that handles out-of-order arrivals in a multipath setting. The system employs lightweight per-connection state (ACK/SACK bitmaps) and multiple loss-detection modes to achieve fast, accurate recovery without heavy NIC overhead. Across synthetic and AI/ML workloads, STrack delivers substantial gains over RoCEv2, including up to 6x improvements on synthetic traffic and about 27.4% on collective workloads, demonstrating its practical impact for large-scale AI/ML clusters.
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.