Research And Implementation Of Congestion Control Oriented User-space Custom Transport Protocol

The development of infrastructure and network technology spurs the explosive growth of all kinds of distributed applications.The complex and changing application scenarios make the research of transport protocol become a hot spot in the field of network once again.Congestion control is the core mechanism of transport protocol.In recent years,with the promotion of machine learning and other technologies,new algorithms have been proposed.It is the trend of transport protocol to customize congestion control according to application requirements.However,most of these algorithms stop at theoretical simulation,and the implementation or even deployment in the transport protocol has always been a pain point for researchers.Therefore,this thesis studies and analyzes the implementation scheme of transport protocol since its development,proposes a user-defined transport protocol framework CCUTP for congestion control,and customizes congestion control for the transmission requirements of distributed machine learning applications.The core work of this thesis is as follows:(1)By comparing the advantages and disadvantages of the existing transport protocols,combined with the research requirements of application-oriented congestion control,the user space transport protocol scheme based on UDP is selected,and the framework of CCUTP protocol is designed and implemented,so that all kinds of congestion control algorithms can be customized conveniently and quickly.Firstly,three congestion control protocol models are abstracted based on different congestion sensing modes: packet loss control mechanism,delay control mechanism and hybrid control mechanism.Different control events are designed based on the three models to adapt to various congestion control algorithms.Secondly,CCUTP implements the unified data transmission function,decouples the basic transmission and congestion control,makes the algorithm research shield the underlying communication,so as to achieves the purpose of rapid customization of congestion control.Finally,this thesis introduces the development process of CCUTP custom congestion control.Based on the corresponding abstract model of CCUTP,three congestion control algorithms TCP Reno,Verus and DMLCC are implemented to verify the feasibility of the framework.(2)Distributed machine learning is a typical distributed application.Based on the framework mentioned above,this thesis further studies the user-defined congestion control for distributed machine learning applications.In view of the transmission characteristics of long tail flow and bounded-loss tolerance in distributed machine learning application scenarios,the congestion control algorithm DMLCC based on online learning is designed and implemented.In terms of congestion sensing,DMLCC takes delay gradient,packet loss rate and throughput as congestion information,transforms congestion information into utility value through an utility function,and perceives network state more finely.In terms of rate control,DMLCC directly controls the sending interval of packets according to the change of utility value,which is not limited by the size of the window,and improves throughput as much as possible to reduce the flow completion time.Finally,compared with TCP congestion control algorithm commonly used in distributed machine learning,the results show that DMLCC can improve the bandwidth utilization by at least two times compared with TCP at low packet loss rate,and can complete the transmission of large amount of data faster.