BBR
2017-01-07
The linked paper describes a new TCP congestion control algorithm. The key insight is an old assumption about networks is no longer true. It used to be the case that networks only dropped packets when they were unusable or when they were congested. Thus TCP stacks could work out how rapidly to send packets by ramping up until they saw packet loss.
In modern networks, packet loss is common long before a TCP connection reaches capacity. Our old congestion control algorithms were unnecessarily slowing connections.
The most important part of the paper is the use of field experiments. It goes beyond the usual prototype implementation, and deployed the algorithm to a large network and measured benefits.
More computer science needs to happen in the field.
queue.acm.org/detail.cfm?id=3022184