PCC vivace: Online-learning congestion control

TCP's congestion control architecture suffers from notoriously bad performance. Consequently, recent years have witnessed a surge of interest in both academia and industry in novel approaches to congestion control. We show, however, that past approaches fall short of attaining ideal performance. We leverage ideas from the rich literature on online (convex) optimization in machine learning to design Vivace, a novel rate-control protocol, designed within the recently proposed PCC framework. Our theoretical and experimental analyses establish that Vivace significantly outperforms traditional TCP variants, the previous realization of the PCC framework, and BBR in terms of performance (throughput, latency, loss), convergence speed, alleviating bufferbloat, reactivity to changing network conditions, and friendliness towards legacy TCP in a range of scenarios. Vivace requires only sender-side changes and is thus readily deployable.