近年来，随着计算机和网络技术的迅猛发展，出现了带宽大于1Gbps，甚至10Gbps的高速网络，这样的高速网络环境带来了更为严峻的拥塞控制问题。当前被广泛使用的TCP拥塞控制算法不能满足高速数据传输的需要，因此研究新的适应高速网络的拥塞控制机制具有非常重要的意义。本文分析了影响网络拥塞控制的关键因素，总结了最近的重要研究成果，在此基础上提出了一系列改进机制。本论文的主要内容和结论如下： 第一，介绍了关于网络拥塞控制的基本知识，总结了目前主要的拥塞控制算法，讨论了拥塞控制研究中的主要研究问题，指出了TCP协议在高速网络中的不足及改进思路。阐述了选题背景和论文的主要意义。 第二，针对TCP在高速网络中带宽扩展性不足的问题，对TCP的AIMD机制做了改进，把拥塞间隔信息直接引入拥塞控制中来调节AIMD的参数，提出了一种基于拥塞间隔信息的改进算法TBCC（Time Based Congestiong Control），同时获得了较好带宽扩展性和RTT公平性，改善了TCP的性能。 第三，基于优化理论，根据Kelly的拥塞控制框架提出了一组加速网络系统收敛的算法QKC(Quick Kelly Control)。针对负价格反馈信息带来的公平性受限和稳态超调问题，直接引入链路的容量利用率的作为反馈信号，在实现比例公平性的同时避免了稳态超调。并从理论上证明了原－对偶QKC算法（PDQKC）的全局渐近稳定性。 第四，提出了一种QKC拥塞控制算法的离散化形式，分析了该算法在任意时延情况下的稳定性，给出了与时延无关的渐近稳定性条件。 第五，从实际网络的性能要求出发，提出了A-CQKC（Accelerated Compact QKC）算法，把PDQKC算法中原本由链路端完成的动态调节功能集中到源端主机中实现。链路端只需提供简单的状态信息即可使系统达到较好的性能指标，降低了链路端的负荷和实现难度。 第六，提出了基于二分反馈信息的算法Simple-QKC和Power-QKC。考虑系统反馈不充分的情况，仅仅利用拥塞丢包信息作为控制器输入，将A-CQKC算法推广到缺乏连续反馈信息的高速网络环境中。 最后，对取得的成果进行了总结，并展望了需要进一步研究的工作。
英文摘要
Firstly, the general backgrounds of network congestion control are introduced. And then, the current main network congestion mechanism, that is, TCP protocol is describled. After that, the shortages of the TCP protocol are given and main improvements in recent research are discussed. The main work and backgrounds of this dissertation are described. Secondly, to make TCP more scalable to large bandwidth, the AIMD congestion control mechanism is enhanced. A new algorithm called TBCC (Time Based Congestion Control) is proposed to support bandwidth scalability while alleviating RTT unfairness. In TBCC, the real time intervals of continuous congestion events are adopted to update the parameters of AIMD mechanism. Compared with similar protocols, TBCC can guarantee the bandwidth scalability and achieve better RTT fairness. Thirdly, based on Kelly’s work, we present a new group of algorithms called QKC (Quick Kelly Control) to accelerate the convergence rate. Point to the drawback of negative price, QKC uses the link utilization ratio directly as feedback signal. It is shown that PDQKC has powerful bandwidth scalability and offers fast convergence rate without sacrificing proportional fairness. The global asymptotic stability of a primal-dual form of QKC (PDQKC) is proved without considering delay. Fourthly, a discrete form of QKC is presented. Based on the analysis to this algorithm, the delay-independent stability conditions are given. Fifthly, to satisfy the performance requirements of practical networks, a Compact QKC algorithm called A-CQKC (Accelerated Compact QKC) is proposed. Compared to PDQKC, the link algorithm abandon dynamic update law and only provide their utilization ratios to the sources. All the dynamic update laws are moved into the source algorithm. This simplifies the design of the link algorithm and makes the whole algorithm easer to carry out. Sixthly, in practical networks, packet losses may be the only feedback signals. Based on binary feedback signal，two algorithms called Simple-QKC and Power-QKC are proposed. Using packet losses as the input of the source controller, these algorithms can be applied to the high-speed network system which is lacking continuous feedback.
修改评论