Towards High-Performance and Compatible RDMA Networks with Receiver-Based and Fine-Grained Congestion Control

In Remote Direct Memory Access (RDMA) networks, congestion control (CC) is crucial for high-speed and reliable transmission. Heuristic CC schemes, deployed at the sender for easy integration without requiring special network equipment, offer good compatibility by relying on coarse-grained end-to-end signaling. However, this approach hinders accurate detection of the full path status and timely feedback, ultimately degrading system performance. On the other hand, some schemes adopt fine-grained switch measurements for precise congestion status to boost high performance, but always require specialized network equipment. In this paper, we propose RFCC, a receiver-based and fine-grained RDMA CC scheme with high performance and strong compatibility simultaneously. To achieve these, RFCC identifies congestion flows based on congestion marking ratios and distinguishes congestion types according to total receiving rate, enabling a finer-grained control without the need for specialized network equipment. Then, RFCC further enhances system performance by accurately tracking the number of congested flows, measuring bottleneck bandwidth and individual flow receiving rates, and quantifying congestion levels for precise rate adjustments. Additionally, RFCC encapsulates collected congestion information into a Congestion Notification Packet (CNP) for timely feedback. The extensive experiments demonstrate that RFCC significantly outperforms existing CC schemes, reducing the flow completion time (FCT) slowdown and the buffer usage by 56.7% and 61.3% on average.