Congestion Control Research Of High Speed Networks

With the development of the network transmission service and science researching application (for example, telnet, satellite transmission, large scale database), there are more and more networks (High Speed Network), whose bandwidth is larger than 1Gbps, and RTT is longer than 100ms. The development of computer, communications and technology of storage provides enough capacities and fast-effective hardware for calculating and scientific research. The challenge of developing the new generation networks is that current congestion control algorithms and resource sharing algorithm can't be expanded to the new generation High Speed Networks. It is mainly exhibited as in the high bandwidth network environment, the current protocol algorithms can't guarantee nicer service quality of low packet loss rate, RTT and dithering of RTT. Currently, it is in the originative phases of high speed networks researching. There are some representative source algorithms, for example, HSTCP, FAST TCP, XCP etc. Because of simplicity and scalability of the implementation of HSTCP, it is accepted by the IETF. Congestion control algorithms are composed of TCP source algorithm and IP link active queue management algorithm. According to the control theory, the source algorithm and IP link algorithm compose a close-loop feedback control system. Current researching indicates that long RTT will lead to unstable state of the whole network.First, we design the DBC active queue management algorithm using the Dead-Beat theory, which is adaptive to the long time-delay networks according to the performance of delay to the congestion control system. It can make the queue length in the router track the reference queue length with non-error, and also has fast response performance both in large delay networks and dynamic traffic networks.Secondly, the current research indicates that HSTCP also has some serious deficiencies. There exists large number of packet loss upon congestion epoch and serious RTT unfairness with DropTail gateway. To overcome the above deficiency, we propose an improved algorithm, called EHSTCP, which switches between two congestion avoidance modes in congestion avoidance phase. To locate the switch point, we present a new predictive method for end-to-end available bandwidth using history congestion window dynamics. Simultaneously, we introduce the RTT fairness factor to eliminate the serious RTT unfairness in HSTCP.Thirdly, we analyze the performance of the HSTCP/AQM close-loop system, it indicates that the active queue management algorithm can overcome the deficiency of HSTCP in the DropTail gateway, but the existing active can't be brought to the high speed networks directly.Finally, we propose a new active queue management algorithm called SPI which is effective in the high speed networks. SPI can guarantee high link utilization, and also can overcome the above deficiencies of HSTCP managed by the DropTail algorithm.