# STrack: A Reliable Multipath Transport for AI/ML Clusters

**Authors:** Yanfang Le, Rong Pan, Peter Newman, Jeremias Blendin, Abdul Kabbani, Vipin Jain, Raghava Sivaramu, Francis Matus  
**arXiv:** [2407.15266](https://arxiv.org/abs/2407.15266)  
**Date:** 2024-07-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2407.15266v1/

## Key points

- STrack is a multipath reliable transport protocol designed for distributed training across GPUs in AI/ML workloads
- It optimizes congestion control and load balancing simultaneously with an adaptive algorithm leveraging ECN and RTT
- Key features include out-of-order delivery, selective retransmission, and swift loss recovery in hardware for multipath environments
- Implements window-based congestion control prioritizing path selection choices before reducing window size
- Assumption of lossy Ethernet as link layer technology with reliable error recovery mechanism based on out-of-order packet counts at receiver NICs
- Outperforms RoCEv2 by up to 6X with synthetic workloads and by 27.4% with collective workloads in simulations
- Utilizes oblivious packet spray to distribute packets evenly across different entropy values for enhanced data transmission efficiency

## Summary

, , , , 

STrack is a groundbreaking multipath reliable transport protocol designed to address the challenges posed by emerging artificial intelligence (AI) and machine learning (ML) workloads that require distributed training across hundreds or even thousands of GPUs. This innovative protocol optimizes congestion control and load balancing simultaneously, incorporating an adaptive load balancing algorithm leveraging Explicit Congestion Notification (ECN) and utilizing Round-Trip Time (RTT) as multi-bit congestion indicators for precise congestion window adjustment. One of the key features of STrack is its ability to facilitate out-of-order delivery, selective retransmission, and swift loss recovery in hardware for multipath environments. By employing a novel mechanism that adaptively sprays packets to multiple paths without maintaining complicated per-path state, STrack ensures efficient data transmission. Additionally, STrack implements a window-based congestion control mechanism that prioritizes path selection choices before reducing the window size in response to pending congestion. Furthermore, STrack assumes lossy Ethernet as the link layer technology and incorporates a reliable error recovery mechanism based on out-of-order packet counts at receiver Network Interface Cards (NICs). This approach ensures fast packet recovery with minimal spurious retransmissions, enhancing overall system performance. In extensive simulations comparing STrack with RoCEv2, it was found that STrack outperforms RoCEv2 by up to 6X with synthetic workloads and by 27.4% with collective workloads, even when using an optimized RoCEv2 system setup. The use of oblivious packet spray in distributing packets evenly across different entropy values further enhances the efficiency of data transmission in multipath environments. Overall, STrack represents a significant advancement in reliable transport protocols for AI/ML clusters, offering improved performance and reliability for managing collective communication in distributed training scenarios.