Improvement of Cyclic Performance Variation between TCP BBR and CUBIC

Abstract

Google proposed a new congestion control algorithm called bottleneck bandwidth round-trip propagation time (BBR) in 2016. CUBIC TCP tends to fill the bottleneck buffer, whereas TCP BBR tries to send data without creating excess queue. BBR periodically measures the maximum available bandwidth and the minimum round-trip time to estimate the bandwidth delay product (bdp). Then, BBR determines its sending rate. However, when BBR competes with CUBIC, it has some problems caused by BBR's incorrect estimation. These problems are as follows: most of the bandwidth is occupied by a certain flow, and excessive packet loss occurs. In this paper, we propose a new BBR loss recovery mechanism so that BBR can coexist with other loss-based congestion control algorithms without excessive packet losses, as we focused on searching for the appropriate minimum RTT to allow BBR and CUBIC to coexist fairly on the bottleneck link. Furthermore, we constructed a simple testbed topology so as to evaluate the performance of the proposed BBR loss recovery mechanism. We showed that proposed BBR got a fair share with CUBIC on the bottleneck link as well as that the packet loss caused by BBR's bdp overestimation were reduced.