TCP Fairness among Modern TCP Congestion Control Algorithms including TCP BBR

For improving communication performance, many fast TCP algorithms, e.g. CUBIC TCP and Compound TCP, have been proposed. These proposals have raised another issue that is performance fairness among TCP congestion control algorithms. Several papers were published for discussing the fairness and some of them revealed their unfairness. In addition, some fairness improving methods based on packet dropping, for example applying CoDel, have been proposed. However, these existing works were based on the existing TCP algorithms, which were loss-based, delay-based, or hybrid type TCP congestion control algorithm. In 2016, another TCP congestion control algorithm, called TCP BBR, was proposed. The algorithm is based on the congestion model by Kleinrock and not loss-based or delay-based. In this paper, we investigate the performance fairness between CUBIC TCP and TCP BBR. We present performance evaluation in conditions wherein connections of TCP BBR and CUBIC TCP are concurrently established. We then demonstrate that the performance fairness between TCP BBR and CUBIC TCP is very low, especially with high latency. In the case of 64 ms RTT, TCP BBR obtained about 45 times of performance of CUBIC TCP. We applied CoDel for improving TCP fairness and evaluated the fairness. Our evaluation showed that it did not work well in case of TCP BBR consumes much bandwidth. On the other hand, it is effective in case of TCP BBR cannot obtain enough throughput.