Study Of TCP Performance Over OBS Networks

Optical Burst Switching (OBS) has attracted considerable research attention from both academic and industrial fields because it perfectly combines high-speed optical transmission/switching technology and flexible electronic processing mechanism. Although OBS is a promising solution for future IP/DWDM all-optical network, a lot of key issues are still under studied. For example, how to support the TCP transmission which is today's prevailing transport protocol of Internet.In this dissertation, LOBS network testbed is firstly introduced simpley. Then the influence of OBS on TCP performance is firstly analyzed, such as burst losses and burst assembly. Then three mechanisms are proposed to improve the performance: burst assembly process optimization, ACK-based assembly and scheduling algorithm and burst retransmission-based OBS scheduling algorithm. The first one is to maximize available TCP bandwidth through burst assembly optimization by simulation and experiments. But the scheme can not reduce burst losses, so the improvement is marginal. The second method improve the TCP performance through ACK loss control, but under low burst loss probability, the weakness of FTO can not be overcomed. The third one is to overcome FTO effect with retransmission sheme and multiple queues introduced.Following are details.1. Since in OBS networks data is transmitted and switched in the format of large bursts, which is different from small packets in traditional packet switching networks, TCP will exhibit different characteristics. Therefore, the influence of burst losses on TCP performance is investigated through simulation and experiments. Results show that in order to improve the performance, it is essential to reduce burst loss probability. Then the effects of delay penalty and correlation benefit due to burst assembly on TCP are studied in detail. Finally, taking into consideration all these effects, experimental results show that there exists an optimal assembly period to maximize the TCP bandwidth.2. A simple modified analytic model is introduced to study the real TCP performance over OBS networks, with ACK (ACKnowledgement) losses introduced. Through simulation and analysis, the modified model is proved proper and no or less ACK losses are suggested to improve the performance. Then two novel ACK-based assembly and scheduling algorithms, i.e. TCP/ACK-based assembly algorithm and, are proposed respectively. Simulation results demonstrate that compared with conventional assembly and scheduling mechanisms, the two new algorithms can improve TCP performance over OBS networks to some extent and the ACK priority scheduling mechanism has better performance than the other one. For example, in the case of 0.6% bust loss probability, compared with conventional burst assembly and scheduling algorithm, TCP/ACK-based assembly algorithm can improve TCP throughput by 7%, while ACK priority scheduling algorithm is about 17%. However, under lower loss probability (lower than 0.1%), the improvement is not evident due to FTO and lost bursts have to be retransmitted by TCP congestion control mechanism. Moreover, in this case, burst loss seldom happens, so the ensurance of no or less ACK losses becomes unimportant.3. A burst retransmission is introduced to make lost bursts retransmitted by OBS firstly, avoiding the unnecessary TCP slow start congestion control mechanism. Then a TCP throughput model based on OBS with burst retransmission is introduced. Simulation and analysis results demonstrate that although burst retransmission scheme can improve network performance to some extent, the performance is degenerated under higher burst contention probability and traffic load. Therefore, a Retransmission-based Group-scheduling algorithm with Multiple Queues (RGMQ) is proposed. With the group scheduling windowΔT due to the group scheduling scheme introduced, the case of burst contention is mitigated. And the multiple queues ensure the burst which has been retransmitted more to have more chances to be transmitted successfully. Simulation results demonstrate the proposed RGMQ can improve network performance evidently under lower contention probability (lower than 0.1%). For instance, in comparison with conventional assembly and scheduling algorithm, in the case of 0.1% probability there is 18% improvement in TCP throughput and even 30% when the probability equals to 0.6%.