SR-DCQCN: Combining SACK and ECN for RDMA Congestion Control

In recent years, the use of RDMA in data center networks has increased significantly, with the largest deployments of the RoCEv2 protocol using the go-back-n scheme for loss recovery. However, it can waste time and bandwidth due to redundant packets and can even exacerbate network congestion levels. In this paper, we propose a combination of selective retransmission and DCQCN congestion control schemes (SR-DCQCN) to improve the loss recovery process in data centers, where the sender retransmits only the truly lost packets based on the information carried by the received NACK and the receiver achieves out-of-order packet reception with the support of bitmap. To improve loss recovery in data centers, we combine such a selective retransmission scheme with DCQCN, the default congestion control scheme for data centers, which includes DCQCN's ECN mechanism and its core congestion control algorithm. Through our experimental evaluation, SR-DCQCN reduces the 99%ile flow completion time by 51.9% compared to DCQCN when the network load utilization is 50%.