The Research Of Congestion Control Based On Machine Learning

With the rapid development of Internet and network equipments,the drawbacks of traditional transport protocols have gradually emerged.Recently,the QUIC aimed at improving the loading time of web pages has attracted much attention,but it needs further research on the related mechanism and performance.In all reliable transport protocols,it is regarded that all packet losses are caused by network congestion,even in the new QUIC protocol.However,in many lossy networks,such as wireless networks,non-congested packet loss is common,which will greatly impair network performance.In this paper,we focus on the above problems.1.We study the typical reliable transmission protocol,analyze the QUIC and TCP protocol.and carry out the comparison experiments in the test and real networks under different network environments.First,we research the transport aspect of QUIC,with the packet pacing mechanism for congestion control.By decoupling,we can exploit the potential of QUIC in other application scenarios besides HTTP/2.It is also useful for the improvement of QUIC itself.Our main findings indicate that,over Cubic,QUIC performs better in a network with high loss rate,or with small buffer size,or with small propagation delay;the main benefits of QUIC lie in the multi-stream-based multiplexing mechanism and the lack of head-of-line problem;with tradeoff,packet pacing is regarded as an effective machanism.2.We design a packet loss classifier based on supervised learning,and random forest is used as the training algorithm to realize the packet loss classifier over QUIC.The features of the loss classifier including the density of packet loss and the variant of RTT are independent of the environment.Overall consideration,Random Forest is selected as the training algorithm,and the accuracy of our model reaches 94.4%.We have implemented it over the emerging transport protocol,QUIC.Our proposal only deploys on the QUIC server without any modification on the client.The experiment results indicate that,with the classifier,the page load time is improved by over 11% when the loss rate is 0.5%.And the larger the network delay is,the better the effect is.