FRCS: Congestion Control in Datacenter Networks Based on Fair Rate Calculation by Switches

Remote Direct Memory Access (RDMA) has been widely deployed in data centers to further reduce latency and CPU overhead. The line rate in data center networks (DCNs) is gradually growing from 40Gbps to 100Gbps and beyond, and more and more flows will be completed in a single round-trip time (RTT). Traditional congestion control mechanism cannot accurately regulate different flow transmission rates in such background, and it will produce an increase in the flow completion time (FCT) and a loss in throughput. In this paper, we propose FRCS, a congestion control algorithm based on direct fair rate computation that enables fast convergence and switch queue emptying. FRCS can also address the victim flow problem during multiple congestion points. To achieve this goal, FRCS adopts different control methods according to the lengths of flows, calculates the flow fair rate using programmable switches, and passes the information back to the sender via ACK. Simulation results show that FRCS, compared to HPCC, not only achieves a lower FCTSlowdown but also reduces the average queue length at switches by more than 50%.