【作者】 程双梅；

【导师】 李建华；

【作者基本信息】 上海交通大学 ， 信息与通信工程， 2009， 博士

【摘要】 随着光纤通信等新技术的应用,网络带宽迅速增加,传统TCP协议因为其保守的拥塞控制机制而不能充分利用可用带宽。因此,高速传输协议成为近年来的研究热点,其中较有影响的有HSTCP、STCP、BIC、FAST TCP和XCP等协议。这些协议中,XCP由于其快速收敛、极佳的公平性和很小的排队时延等卓越性能,受到众多研究者的关注。但是,由于XCP的拥塞控制机制与传统TCP的拥塞控制机制存在很大不同,它们之间的互联存在很大困难。针对这一问题,研究者提出了TCP友好的XCP、基于XCP的CSFQ和XCP-i等方案。然而,这些方案对升级过程中混合网络中可能存在的各种流没有作深入研究,它们的TCP兼容性很差,且升级过程的可操作性不强。这些互联方案的研究之所以不够深入,其原因就在于对XCP的实时特性不够了解。针对这一问题,本文对XCP的时域性能进行了深入分析,推导出了XCP流速率的时域表达式,并以此为基础给出了XCP协议的响应函数。在这些知识的基础上,作者进一步提出了基于XCP-TCP网关的互联方案,并分别从理论分析和实验仿真对方案进行了验证。总结起来,全文的创新点如下:通过分析反馈与输入流速率的关系,得到XCP瓶颈链路的输入流速率依指数速率收敛于链路带宽容量的结论。在单瓶颈链路条件下,从分析吞吐量占比入手,首次得到吞吐量占比的时域表达式。该表达式依指数速率收敛于流数量的占比。其次,作者首次给出了XCP流速率的时域表达式。这些结论不仅从理论上证实了以前关于XCP性能的有关结论,例如公平性和带宽利用率,也为XCP的拥塞控制机制提供了更直观的认识。综合以前的研究成果,提出了评价高速传输协议的五个指标。规模性作为其中的一项重要指标,可以用响应函数来描述。在周期丢包模型条件下,作者首次推导出了FAST TCP和XCP协议的响应函数。与其他协议的响应函数不同的是,XCP的平均吞吐量不仅与丢包率有关,还与每流带宽有关。这从理论上证实XCP的规模性优于其它协议。在对现有的XCP与传统TCP互联方案深入分析比较的基础上,提出了基于XCP-TCP网关的方案。该方案首次完整地考虑了升级过程中各种可能的流的工作状态,并对这些流之间的公平性进行了理论分析。结论显示在该方案中各种流之间能够友好共存。其次,文中还对XCP瓶颈链路的带宽利用率和收敛时间、排队时延等进行了理论分析。结果表明在稳态下瓶颈链路的带宽利用率不低于80%,其排队时延优于传统FIFO路由器的排队时延。更多还原

【Abstract】 With the deployment of new technology such as fiber communications, the bandwidth capacity grows rapidly, and the traditional TCP cannot fully utilize the available bandwidth because of its conservative congestion control mechanism. Therefore, high speed transport protocols have been a hot topic in recent years, among them some famous ones are HSTCP, STCP, BIC, FAST TCP and XCP. Because of its prominent performance such as rapid convergence, excellent fairness and very small queueing delay, XCP has attracted much attention from researchers. However, because the congestion control of XCP is quite different from that of traditional TCP, and as a result, it is difficult to connect XCP with traditional TCP. To address this problem, some schemes such as TCP-friendly XCP, XCP-based CSFQ and XCP-i have been proposed by researchers. However, these schemes do not deal in-depth with the possible flows appears in the hybrid networks during the upgrade process, and the TCP compatibility and feasibility of these schemes are poor.Why the proposals cannot study in-depth is because little knowledge is known about the time-scale characteristics of XCP. To address this problem, the author investigates deeply into the time-domain performance of XCP, and derives the time-domain sending rate of XCP, based on this point, the response function of XCP is also provided. Based on this knowledge, an XCP-TCP gateway based scheme is further proposed by the author, analysis and simulations validate the proposal. The original work of this paper can be concluded as follows.By analyzing the relationship between feedback and input traffic rate, the author derived that the input traffic rate of an XCP bottleneck converges to its bandwidth capacity in exponential rate. Under a single bottleneck model, by concentrating on the throughput ratio, its time-domain expression was first derived. The expression showes the throughput ratio converges to the flow number ratio in exponential rate. Secondly, the time-domain sending rate of an XCP flow is also given. These results not only validate the previous conclusion about XCP performance such as fairness and bandwidth utilization, but also provide more intuitive view about XCP congestion control.Based on the previous studies, five indexes are proposed to evaluate the high speed transport protocols. As one of the most important index, scalability can be described by the reponse function. Under a periodic drop event model, the author first derived the response functions of FAST TCP and XCP. Different from other response functions, the average throughput of XCP is related not only to the drop event rate, but also the per-flow bandwidth. Thus it proved that the scalability of XCP is better than other ones.Based on the comparison of the previous schemes about XCP interconnecting with traditional TCP, the author proposed an XCP-TCP gateway based scheme. The scheme first extensively investigated the work status of possible flows appear in the upgrading process, and derived the analytical fairness between these flows. Results show that these flows can coexist well with each other. Secondly, the author also analyzed the bandwidth utilization, converge time and queueing delay of an XCP bottleneck. Results show that the bandwidth utilization of the bottleneck is not less than 80% when in equilibrium, and that the queueing delay is better than that of a traditional FIFO router.更多还原