Is BBR better than Cubic?
For example, under 200ms RTT and 500Mbps bandwidth, BBR improves 115% goodput compared to Cubic. This is because BBR uses the bandwidth and delay estimates as the congestion signal rather than packet losses. We will discuss it more next. Rather than using packet loss as the congestion signal, BBR regulates its traffic based on observed bandwidth and delay values. Specifically, BBR limits its number of packets in flight to be a multiple of the bandwidth-delay product (BDP). Furthermore, BBR also uses pacing to control the inter-packet spacing.Google developed BBR to overcome the limitations of loss-based algorithms. BBR is designed to optimize network utilization and throughput by continuously probing for available bandwidth.In the STARTUP phase, BBR tries to quickly approximate the bottleneck bandwidth. It does so by increasing the sending rate until the estimated bottleneck bandwidth stops growing. Bottleneck bandwidth is estimated from the amount of data ACKed over a given period, filtered through a windowed max-filter.The difference in goodput between BBR and BBRv2 is greater under the simulated high bandwidth condition; for all three buffer sizes, BBR outperforms BBRv2 by at least 2×. This is because BBR does not react to losses and so continues to aggressively use the available bandwidth.
How does BBR work?
BBR flows spend the vast majority of their time in ProbeBW state, probing for bandwidth using an approach called gain cycling, which helps BBR flows reach high throughput, low queuing delay, and convergence to a fair share of bandwidth. With gain cycling, BBR cycles through a sequence of values for the pacing_gain. BBRv3 is a minor revision of BBRv2, and it addresses two key performance issues of BBRv2. First, an implementation change caused BBRv2 to prematurely exit the ProbeBW phase that in turn prevented it from sharing bandwidth equitably with BBRv1 or loss-based CCAs such as Cubic [8].