Congestion Control with Receiver-Aided Network Status Awareness in RDMA Transmission

Blockchain technology relies on distributed networks, and Remote Direct Memory Access (RDMA) technology, characterized by ultra-low latency, high bandwidth, has the potential to significantly improve the transmission performance of such networks. RDMA requires a underlying lossless network (usually guaranteed by link-layer flow control called PFC) to fully exploit its performance, wherein congestion control emerges as a key technology in RDMA. However, we find that existing congestion control schemes have limitations in rapidly allocating network bandwidth to eliminate congestion, thus even aggravating side effects of PFC (e.g., head-of-line blocking, unfairness, and deadlock). In this paper, we propose an RDMA congestion control scheme based on receiver-aided network state awareness (RRCC). This research introduces the following key innovations: 1) calculating congestion information through the ECN signals in data packets to achieve a more precise network state sensing method; 2) monitoring the throughput in the receiver side in real-time, and in combination with network state information, periodically adjusting the sender’s rate to achieve rapid rate convergence, accurately preventing and controlling congestion, and addressing issues such as increased flow completion time and slow network congestion recovery. We evaluate RRCC using realistic traffic traces under a three-layer Clos network architecture. The results show that RRCC significantly outperforms existing congestion control schemes in terms of throughput and flow completion time while reducing the side effects of PFC.