Congestion and Fairness Control Mechanisms of TCP for the High-Speed Internet

Preface In the Internet, TCP (Transmission Control Protocol) has been widely used as its transport-layer protocol. Many Internet services such as HTTP (Hyper Text Transfer Protocol) for World Wide Web and FTP (File Transfer Protocol) are designed on the basis of TCP. A number of researchers have been extensively working on TCP's congestion control mechanisms which is an essential part to realize the effective network utilization. It has been known that the current TCP does not scale to high-speed networks directly, so that researches on TCP mechanism for high-speed networks is now one of hot topics. However, one problem of those researches is that neither stability nor fairness of TCP is fully considered; that is, most studies on TCP mechanism focus only on its effective throughput. Therefore, it is strongly needed to investigate TCP's essential characteristics in terms of fairness and stability as well as throughput. It is also an important issue how TCP can be applied to very high-speed networks, and how TCP can adapt to new Internet services, such as ADSL (Asymmetric Digital Subscriber Line) and cable modem networks. In this thesis, we first analyze the behavior of the TCP mechanism for better understanding of its dynamical behavior. The primary objective of this part is to investigate stability and fairness of various versions of TCP mechanisms: TCP Tahoe, TCP Reno and TCP Vegas. We analyze behavior of two TCP connections sharing a same bottleneck link, and evaluate system stability and fairness between these connections. Analytic results show that TCP Tahoe/Reno can achieve good fairness among connections, but they cannot to keep stability of their window sizes because of the essential nature of their algorithms. TCP Vegas, on the other hand, can achieve good stability, but its fairness is sometimes degraded. We propose a simple modification on the congestion avoidance mechanism of TCP Vegas for achieving better fairness and stability, and validate its effectiveness through simulation experiments. The simulation results indicate that our proposed mechanism achieves better stability than TCP Tahoe and Reno, and also achieves better fairness than TCP Vegas. We next focus on a problem in the packet retransmission mechanism of TCP. The problem occurs when a fixed amount of bandwidth is assigned to a TCP connection, or when an RTT (Round Trip Times) of a TCP connection changes frequently. In this case, i TCP frequently performs unnecessary packet retransmission (we call mis-retransmission in this thesis), which …