Abstract
CUBIC TCP is a congestion control algorithm that enables high-speed communication, and is widely used as the default algorithm in Linux. TCP BBR, which was proposed in 2016, is also a widely used algorithm. It is based on Kleinrock's congestion model. It then estimates the bottleneck bandwidth (BtlBw) and the physical propagation delay (RTprop), and sets the window size as the product of these values, which is called the bandwidth-delay product (BDP). In some existing papers, it has been reported that in a situation where CUBIC TCP and TCP BBR share a bottleneck link, there is a fairness problem in their throughput. In an existing paper, the authors showed a serious problem that the throughput of TCP BBR drops significantly at the beginning of communication in the sharing situation. They then proposed a method to improve fairness where TCP BBR invokes the expire procedure if its estimated RTprop value and the measured RTT value are significantly different. In this paper, we evaluate its performance and show a significant decrease in throughput when TCP BBR communicates without sharing the network, i.e., standalone communication. To address this issue, we propose a method to mitigate this problem by dynamically setting the threshold to invoke the expire procedure and increase the pacing_gain. We then present performance evaluation results showing that the proposed approach achieves TCP fairness equal to or better than the existing method, without degrading the throughput during standalone communication.