Barre: Empowering Simplified and Versatile Programmable Congestion Control in High-Speed AI Clusters

Network interface cards (NICs) and switches have entered the 400 Gbps era. RoCEv2 networks face significant challenges in congestion management, particularly under high-throughput workloads. While advanced congestion control algorithms have been proposed, their deployment in large-scale data centers remains hindered by complex parameter tuning and dependency on sophisticated hardware features. In this paper, we present Barre, a simple yet highly effective congestion control scheme designed for modern AI/HPC clusters operating at 400 Gbps. By leveraging commodity hardware and standard network functionalities, Barre achieves near-optimal performance in fairness, congestion responsiveness, and scalability with minimal overhead. Deployed in our 400 Gbps RoCE cluster for over a year and supporting up to 10,000 GPUs, Barre improves AI training task throughput by an average of 9.6%. Furthermore, we demonstrate that Barre’s core principles can be seamlessly applied to enhance DCQCN, a widely deployed congestion control algorithm, underscoring its practicality and versatility.