Impact of Segment Size and Parallel Streams on TCP BBR

Abstract

TCP BBR has been recently proposed as a congestion control algorithm. BBR represents a disruption from the window-based loss-based congestion control used during the last 30 years. While BBR has been tested for trivial applications (e.g., browsing, YouTube), its use for moving big data has not been extensively studied yet. Features that largely impact the efficiency of transporting big flows are the use of parallel streams and the maximum segment size (MSS). This paper studies the impact of these two features on big flows, in the presence of packet losses and latency. Empirical results demonstrate that BBR reacts better than window-based loss-based algorithms (Cubic, Reno, HTCP) to large MSS. Similarly, as the number of parallel streams used in a data transfer increases, the performance gap between BBR and Cubic, Reno, and HTCP increases in favor of BBR. For example, in a 20-millisecond RTT, 10 Gbps network with high corruption rate (0.01%), BBR's average improvement factor from using multiple streams is almost 4. In contrast, HTCP's, Cubic's, and Reno's improvement factors are below 2. Using large MSS and parallel streams permit BBR to sustain high throughput, even in the presence of a significant corruption rate.