3 ms·

Note that that is about BBRv1 (as available in recent linux kernels), BBRv2 is going to address the fairness issues https://datatracker.ietf.org/meeting/105/ma

• justine 7y ago
  Hi! Thanks for your interest in BBR! 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!)