Evaluating BBRv2 on the Dropbox Edge Network

Abstract

Nowadays, loss-based TCP congestion controls in general and CUBIC specifically became the de facto standard for the Internet. BBR congestion control challenges the loss-based approach by modeling the network based on estimated bandwidth and round-trip time. At Dropbox, we've been using BBRv1 since 2017 and are accustomed to its pros and cons. BBRv2 introduces a set of improvements to network modeling (explicit loss targets and inflight limits) and fairness (differential probing and headroom for new flows.) In this paper, we go over experimental data gathered on the Dropbox Edge Network. We compare BBRv2 to BBRv1 and CUBIC showing that BBRv2 is a definite improvement over both of them. We also show that BBRv2 experimental results match its theoretical design principles.