Modeling BBR’s Interactions with Loss-Based Congestion Control [pdf]

BBRv2 aims to solve a different issue -- which is that in shallow-buffered networks, BBR consumes the entire link and starves out Cubic and Reno completely.

This work shows something different. Even in deep-buffered networks (like over most Internet access links) BBR(v1) has a funny property that it takes a fixed fraction fo the network link -- no matter how many other competing TCP connections there are. The deeper the buffer, the closer this fraction edges towards 50%.

Given that there are a lot of changes from BBRv1 to BBRv2, it's not at all clear whether this newer problem will exist in BBRv2 (we're trying to run some experiments to find out soon!)