RACC: Rapid and Accurate INT-Based Congestion Control in RDMA Network
Yanzhe Zhao, Shuo Wang, Siyu Han, Dong Bo Zhou, Guoyu Peng, Tao Huang · 2024
With the rapid growth of network speed, datacenter applications have increasing demands on networks for high throughput and ultra-low latency. RDMA is widely deployed due to its high performance. Most existing RDMA congestion control schemes employ end-to-end architecture with inherent feedback delays of at least one Round-Trip Time (RTT).To overcome these limitations, we propose RACC, a rapid and accurate INT-Based RDMA congestion control scheme. The switches directly provide INT feedback, reducing the feedback signal latency. In addition, the adaptive rate update mechanism is adopted at the sender, which enhances the rapid response to network congestion and the stable control of in-flight bytes. We conduct simulation experiments based on the Fat-Tree topology to analyze the requirements for datacenter performance metrics including convergence, fairness, and dynamic queues. Our evaluations show that the peak queue lengths are reduced by up to 80% compared to HPCC and PowerTCP, and the convergence time after congestion is reduced by half.