Mechanism of Cyclic Performance Fluctuation of TCP BBR and CUBIC TCP Communications

摘要

Traditional TCP congestion control algorithms, such as TCP Tahoe and TCP Reno, do not completely utilize the bandwidth of the recent long-fat networks. To address this issue, many fast TCP algorithms, such as CUBIC TCP, have been proposed. In addition, a novel TCP algorithm called TCP BBR was proposed in 2016. It is causes problems such as performance fairness between TCP algorithms. Existing studies have demonstrated that the performances of TCP BBR and CUBIC TCP severely fluctuate in networks where these TCP algorithms function competitively. However, the mechanism of this fluctuation has not been analyzed. In this paper, we reveal the cause of this performance fluctuation by analyzing insights into the behavior of these TCP algorithms and the queue in the bottlenecked link. When the TCP bottleneck bandwidth and round-trip propagation time (TCP BBR) connections dominate the consumption of the bandwidth, the DRAINs of TCP BBR significantly reduce the queue length and RTT. This causes a very small RTprop and a severe decrease in the congestion window size of TCP BBR. This small RTprop corresponds to an accurate round-trip propagation, but this accurate propagation, unfortunately, degrades the performance of the TCP BBR. Hence, CUBIC TCP becomes dominant. When CUBIC TCP connections dominate the bandwidth, the DRAINs of TCP BBR connections have little effect on the queue size and TCP BBR connections experience a large RTT. Thus, the RTprop and the congestion window sizes of TCP BBR connections increase significantly. Hence, TCP BBR becomes dominant.