Insights into BBRv3’s Performance and Behavior by Experimental Evaluation

This paper investigates the performance of BBRv3. BBRv3 aims to achieve high throughput with low latency and low packet loss rates. We evaluate BBRv3’s performance in a Linux-based testbed and also identify and explain causes of the observed behavior. In contrast to existing work, we investigate BBRv3’s behavior at bottleneck data rates of 100 Mbit/s–10 Gbit/s and delve deeper into the issue of self-induced queuing delay. We also study the impact of delay jitter on performance. Moreover, we investigate the performance of various concurrent short flows taken from a real-world traffic trace. BBRv3 does not really achieve its low delay goal: while it can limit queuing delay in large buffers, it regularly adds 0.95 RTTmin queuing delay in a single flow scenario and more than 1 RTTmin over 50% of the time with multiple BBR flows (RTTmin: round-trip time without queuing delay). We also observe fairness problems and slow convergence that already existed with BBRv2. BBRv3 flows are not strongly susceptible to delay jitter, although it adversely affects queuing delay and fairness behavior to a certain degree (but not starvation). In our experiments with short flows from a real-world traffic trace, BBRv3 performs only slightly, but consistently better than CUBIC.