Transport Architectures for an Evolving Internet

In the Internet architecture, transport protocols are the glue between an application’s needs and the network’s abilities. But as the Internet has evolved over the last 30 years, the implicit assumptions of these protocols have held less and less well. This can cause poor performance on newer networks—cellular networks, datacenters—and makes it challenging to roll out networking technologies that break markedly with the past. Working with collaborators at MIT, I have built two systems that explore an objective-driven, computer-generated approach to protocol design. My thesis is that making protocols a function of stated assumptions and objectives can improve application performance and free network technologies to evolve. Sprout, a transport protocol designed for videoconferencing over cellular networks, uses probabilistic inference to forecast network congestion in advance. On commercial cellular networks, Sprout gives 2-to-4 times the throughput and 7-to-9 times less delay than Skype, Apple Facetime, and Google Hangouts. This work led to Remy, a tool that programmatically generates protocols for an uncertain multi-agent network. Remy’s computer-generated algorithms can achieve higher performance and greater fairness than some sophisticated human-designed schemes, including ones that put intelligence inside the network. The Remy tool can then be used to probe the difficulty of the congestion control problem itself—how easy is it to “learn” a network protocol to achieve desired goals, given a necessarily imperfect model of the networks where it ultimately will be deployed? We found weak evidence of a tradeoff between the breadth of the operating range of a computer-generated protocol and its performance, but also that a single computer-generated protocol was able to outperform existing schemes over a thousand-fold range of link rates