The Research Of TCP Congestion Control Enhancement

Nowadays,many Web services(e.g.,cloud storage)are deployed inside datacenters.Since most Web service deliver their service via TCP protocols,congestion control is the vital component for the quality of service.Considering the networking environment is becoming more complicated,both industrials and academics have proposed many new congestion control schemes in order to meet the emerging requirement.For example,BBR,which is proposed recently,can maintain high performance even in WAN environment.On the other hand,network operators would find it hard to deploy new congestion control because of the limitation on operating systems.For examples,it is currently impossible for Windows servers to be equipped with BBR while deploying BBR as Linux congestion control needs to upgrade kernel to 4.9 or later.Moreover,upgrading kernel may impact on Web service's function.Furthermore,in practice,each Web server could only choose one congestion control to serve all clients from all over the world,e.g.,Cubic is Linux's default congestion control and CTCP is the only congestion control in Windows.Considering complex networking environment,the default congestion control algorithm may not always be the most suitable one.Thus,how to deploy congestion control transparently in data center and how to switch to to the most suitable congestion control according to clients' network environment are core issues for network operators and service providers.To solve congestion control deployment and dynamic switching problems and challenges,this thesis presents three schemes which are summarised as followed:1.Transparent BBR Congestion Control Deployment Platform TCon.Operators could deploy BBR without modification on server's configuration.With extensive experiments,TCon is demonstrated that could achieve similar performance compared to BBR,while could reduce12.76% delay.2.Fine-grained and Transparent Congestion Control Enforcement scheme Mystique.Mystique can effectively adapt to network conditions and dynamically employ the most suitable algorithm for each connection according to rules specified by administrators.We implement Mystique based on OVS,and conducted extensive experiments to demonstrated that Mystique could outperform Cubic,Reno,BBR,Hybla and Illinois.3.Enabling Better Fairness for BBR Congestion Control Modest BBR.In practice,we deployed BBR as our Web servers' congestion control and found that it can generate large amount retransmission.Hence,we propose Modest BBR,a simple yet effective solution based on BBR,responds to retransmission for less aggressiveness for better performance when it coexists with Cubic.With extensive experiments it is demonstrated that Modest BBR gets similar throughput compared to BBR,while it improves 7.1% of the overall throughput.