Research On TCP Over OBS Networks And Hybrid Optical Networks

As a new type of optical switching technology, OBS (Optical Burst Switching) combines the advantages of circuit switching and packet switching while avoiding part of their shortcomings, so it has received a lot of attentions. Hybrid optical networks (HON), which synthesize the advantages of present optical switching technologies and can meet the demands for undertaking sorts of traffics of future networks, also have good prospects for application. On the other hand, since TCP is the most important transport layer protocol and the main part of the end-to-end congestion control in Internet, its performance in different network environments always receives wide attentions. This dissertation investigates the performance of TCP over OBS networks and hybrid optical networks, including four parts: TCP's issue of multiple packet losses, TCP's issue of consecutive multiple losses over OBS networks, high speed TCPs over OBS networks and TCP over hybrid optical networks.TCP's sender is easy to timeout if more than one packet in a congestion window are lost, which is called in this dissertation as TCP's issue of multiple packet losses. TCP's issue of multiple packet losses has been found by early researchers, however, present explanation for this issue is not so accurate, and detailed theoretical derivation of this issue is also lacked. In Chapter 2 the author proposes a window variables based analytical method (WVBAM) to analyze the evolution of TCP's congestion window and the behaviors of TCP sender when multiple packet losses occur. According to results of the analysis, the author gives out the accurate explanation for TCP's issue of multiple packet losses. The author also gets the theoretical conditions for a TO (timeout) when multiple packet losses occur, i.e., the relationships among the three factors including the number of lost packets, the size of congestion window and the distance between the sequence number of the lost packets. Moreover, according to these conditions the author gives out a probability model of TCP's TO after packet loss occurs. The results in Chapter 2 provide necessary theoretical basis for the following researches.According to the results in Chapter 2, the author proposes TCP's issue of consecutive multiple packet losses over OBS networks, i.e., a burst loss inside OBS networks is very easy to cause a TO for TCP, and thus will result in TCP's throughput degradation. The author points out that, consecutive multiple packet losses is one of the key factors of False Timeout (FTO) and can significantly reduce the throughput performance of TCP over OBS networks. Although New-Reno and SACK, the two different TCP variants, can deal well with consecutive multiple packet losses to some extent, they have their own shortcomings when applied for TCP over OBS networks. The author proposes B-Reno (Burst Reno), a new TCP variant designed for TCP over OBS networks. The basic idea of B-Reno is to retransmit more than one consecutive lost packet when packet loss is detected and in each retransmission round during the phase of fast recovery. So, compared with New-Reno, B-Reno can shorten its period of fast recovery obviously. Moreover, B-Reno needs no special support from the receiver's protocol stack. So, compared with SACK, B-Reno has a lower protocol complexity and lower difficulty in deployment The author evaluates B-Reno's performance over OBS networks through comprehensive simulations. Simulation results indicate that over OBS networks B-Reno can achieve throughput performance better than that of New-Reno and similar with that of SACK. The author also presents a mathematic model for B-Reno's throughput, and this model can esmate the throughput of a single B-Reno flow over OBS networks well when the burst loss probability is small. Finally the author evaluates B-Reno's performance in traditional packet switching networks, and simulation results indicate that B-Reno can also work well in packet switching networks.OBS networks are essentially high speed networks which usually have high BDP (bandwidth delay product), so they are suitable for high speed TCPs. In Chapter 4 the author investigates the performance of five high speed TCPs over OBS networks. First, through simulations the author proves high speed TCPs' issue of consecutive multiple packet losses over OBS networks, and this issue will weaken high speed TCPs' advantages of high bandwidth utilizations. Then the author points out and anylizes the performance differences between high speed TCPs with and without SACK: in the same case high speed TCPs with SACK can achieve throughputs obviously higher than those of high speed TCPs without SACK. To overcome B-Reno's deficiency in loss recovery when applied for high speed TCPs over OBS networks, the author proposes aB-Reno (adaptive B-Reno), which is an improvement of B-Reno and a good solution for high speed TCPs' issue of consecutive multiple packet losses over OBS networks instead of SACK. Compared with B-Reno, aB-Reno can adaptively adjust the number of retransmitted packets according to the amount of consecutive multiple lost packets, and thus can remarkably shorten the phase of fast recovery of high speed TCPs. Compared with SACK, aB-Reno needs no special support from the receiver's protocol stack and thus has a lower difficulty in deployment. Through simulations the author proves that, over OBS networks with a low burst loss probability and high access bandwidth, aB-Reno can achieve thoughts higher than those of B-Reno and similar with those of SACK when applied for high speed TCPs. The author also compares and analyzes the throughput performance among the five high speed TCPs over OBS networks at the end of Chapter 4.In hybrid optical networks, the alternation of switching ways will cause routing fluctuations for transported traffics. And routing fluctuations will lead to packet reordering, which will result in TCP's spurious fast retransmission and throughput degradation. In Chapter 5 the author proposes ENDFR (Explicit Notification for Delayed Fast Retransmission) algorithm, a solution for TCP's spurious fast retransmission over hybrid optical networks. ENDFR algorithm uses explicit notifications from the edge node of hybrid optical networks to help the TCP sender predict the forthcoming packet reordering and accordingly adopt the scheme of delayed response to avoid spurious fast retransmission. Compared with most of solutions for packet reordering, ENDFR algorithm avoids complex schemes of packet reordering detection, and is simple to implement. Through simulations the author proves that ENDFR algorithm can effectively avoid spurious fast retransmissions and thus remarkably improve TCP's throughput performance over hybrid optical networks.To validate and evaluate the performance of TCP variants and algorithms proposed in this dissertation, the author uses NS2 (Network Simulator 2) to establish the network simulating platforms. In Chapter 6 the author introduces the simulating platforms of OBS network and hybrid optical network, including key data structures and codes. Finally the author presents the conclusions of this dissertation.