Research On Congestion Control Algorithm In High Bandwidth-Delay Product Networks

The TCP congestion control has been prevalently used in the Internet, since its proposal in 1988. As the Internet evolves, high bandwidth-delay product (BDP) networks are emerging where the bandwidth is larger than 1Gbps and the round trip time (RTT) is more than 100ms. Unfortunately, the TCP congestion control can't effectively utilize the network bandwidth in high BDP networks. A great deal of high-speed algorithms has been proposed for this problem, including HSTCP, FAST TCP, XCP, and VCP, and so on. However, each of the high-speed algorithms has their own limitations. Thus, the need for a perfect high-speed algorithm has become increasingly important.In this thesis, we first review the existing research efforts on congestion control algorithms where the performances of high-speed algorithms are analyzed in detail. Next, we proposed the following work based on the current high-speed algorithms:First, TCP-Africa is a typical hybrid algorithm, which has slow convergence performance in high BDP networks. We propose Fast-Convergence Africa (FC-Africa) algorithm for this problem. FC-Africa adopts a novel congestion window dynamics observation mechanism that allows for identifying whether the flow is to underutilize the fair bandwidth or not. When the flow is underutilized, the congestion window increasing rule is still according to HSTCP without switching to TCP Reno even congestion is imminent; otherwise, the flow resorts to TCP-Africa. Thus, the underutilized flow can rapidly achieve fair bandwidth. Extensive NS2 simulations results indicate that FC-Africa not only significantly reduces the convergence time of TCP-Africa, but also achieves better intra-protocol fairness and RTT fairness properties than TCP-Africa, while at the same time keeping the advantages of TCP-Africa, i.e., high bottleneck link utilization rate and low packet loss rate. Second, existing high-speed algorithms have significant limitations in achieving high efficiency and reasonable fairness while maintaining fast convergence properties. We propose the General Congestion Control Protocol (GCCP) to address this limitation. GCCP adopts aggressive behavior in large underutilized network links to achieve high efficiency, and leverages only one ECN bit for network utilization feedback. Once the link is sensed to be highly utilized, the protocol dedicates to fair and rapid bandwidth allocation by requiring congestion window increment is conservative and monotone decreasing with congestion window increasing. Extensive NS2 simulations results show that the proposed algorithm achieves a pretty good tradeoff between high efficiency and reasonable fairness while exhibiting fast convergence property.