LearningCC: An online learning approach for congestion control

Abstract

AbstractIt is noticed that traditional loss based algorithms are not moving data very well in today's Internet. To get bottleneck bandwidth fully utilized, these algorithms will keep increasing the congestion window and maintain a standing queue of packets in routers, which cause large end to end latency. Such long latency is quite annoying to these ever growing delay sensitive applications. Recently, much effort has been devoted by researchers from both academia and industry to design better congestion control protocols and make a better Internet. In this article, we present an online learning based approach named LearningCC for congestion control. Instead of adjusting the congestion window with fixed rule as these traditional algorithms do, there are several options for an endpoint to choose. Each option is mapped as an arm of a bandit machine. The exploration and exploitation scheme is used to guide the selection on congestion window update rule. Through trial and error, an endpoint improves its performance by choosing action with better reward. Experiments are conducted on ns3 platform to verify the effectiveness of LearningCC. Results indicate it achieves lower transmission delay than loss based algorithms. Especially, we found LearningCC makes significant throughput improvement in link suffering from random loss.