Sliding Window Kalman Filter-Based DCQCN for RDMA Congestion Control

Remote Direct Memory Access (RDMA) is gradually being widely used in data center networks to achieve low data processing delay. The congestion control algorithm is the key for RDMA to realize packet loss-free, while DCQCN is a frequently-used algorithm. However, DCQCN and its advanced version have performance problems leading to slow rate recovery and unstable when network congestion. In this paper, we propose a Sliding Window Kalman Filter-based DCQCN congestion control algorithm. It consists of there parts: (1) the Sliding Window Kalman Filter predicts the latest congestion rate to maintain the network stability, (2) a novel rate control algorithm makes rate recovery faster, and (3) using feedback enables the sender to respond quickly. We evaluate the performance of SWKF-DCQCN in NS-3. The results show that SWKF-DCQCN can recover quickly and maintain stability after congestion, with a maximum improvement of , and the utilization rate of bottleneck link bandwidth can reach .