Control TCP BBR Throughput in a Programmable Switch

In Internet communication, TCP (Transmission Control Protocol) controls congestion. In 2016, a new congestion control algorithm called TCP BBR was proposed to solve the problems of widely used loss-based TCP algorithms, such as TCP NewReno and CUBIC TCP, and it is now widely used. Previous studies have indicated that the throughput fairness is not achieved when multiple TCP BBR connections with different RTTs share the same bottleneck link. To solve the throughput unfairness among TCP connections in a network element, such as a switch and router, some methods based on active queue management, such as RED and CoDel, have been proposed. These methods have been effective for conventional loss-based congestion control algorithms. However, these methods cannot control the throughput of TCP BBR because TCP BBR does not react directly to packet loss detection and does not change its congestion window size upon detection. In this paper, we propose a method to control the throughput of TCP BBR in a network element by making it to observe a small RTT and reduce its congestion window size. We then evaluate the method and show that it can control the throughput of TCP BBRs and improve the throughput fairness among TCP BBR connections.