《互联网拥塞控制算法研究》由会员分享，可在线阅读，更多相关《互联网拥塞控制算法研究（124页珍藏版）》请在金锄头文库上搜索。
1、华中科技大学博士学位论文互联网拥塞控制算法研究姓名：陈虎申请学位级别：博士专业：控制理论与控制工程指导教师：关治洪20080301华 中 科 技 大 学 博 士 学 位 论 文 华 中 科 技 大 学 博 士 学 位 论 文 II执行。进一步，我们指出了 VCP 这类 TCP+AQM 风格协议的公平性远逊于 XCP 的原因是公平控制器执行的次数过少。通过增加公平控制器的执行次数，我们提出了“公平 VCP” (FVCP)算法，有效提高了 VCP 的公平性。 本文提出了几种“自适应的 VCP”(AVCP) 算法。通过将针对高速链路的 VCP扩展至无线链路，AVCP 可以同时适应于高速、无线链路以及
2、这两种链路混杂的情况。AVCP 是一种显式拥塞算法，其核心思想是通过网络反馈的“路径负载因子”来推测数据包的丢失是拥塞导致的还是的传输错误导致的，然后决定是否减小拥塞窗口的大小。显式拥塞算法中，网络需要链路的带宽来计算反馈信息。但是对于无线链路、竞争型的以太网等，其可用带宽难以准确获得。从控制理论的角度，我们指出带宽估计错误可能导致显式拥塞算法不能取得预期的性能。 本文基于优化理论深入研究了拥塞控制的建模与稳定性。我们从物理意义上纠正了 Kelly 模型中的几个错误，并在 Zhang 等的工作基础上提出了一个通用的拥塞控制模型：GKJZ（General Kelly-Johari-Zhang）模
3、型，可适应多种拥塞控制算法。GKJZ 采用时延各不相同的 N 维差分方程来描绘拥塞控制的动力学行为。在 GKJZ模型的框架下，我们对 VCP 进行了建模，并基于模型给出了 VCP 局部稳定性的充分条件。与大多数现有的稳定性条件不同，我们的稳定性判据与时延无关且具有非常简单的形式，易于验证。基于这个稳定性条件，拥塞控制算法只需静态的设置其控制参数，不必根据时延动态调整，具有很强的鲁棒性。 关键词：关键词：互联网，拥塞控制，TCP 友好性，公平性，显式拥塞算法， 稳定性，时延 华 中 科 技 大 学 博 士 学 位 论 文 华 中 科 技 大 学 博 士 学 位 论 文 IIIAbstract D
4、ue to the unbalance distributions of network resources and traffics, the occurrence of congestion is an intrinsic characteristic of the Internet. Congestion control is necessary to keep the stability of the Internet. Congestion control algorithms can be broadly classified into two categories: source
5、 algorithms and link algorithms. Link algorithm, such as Active Queue Management (AQM), is executed in network devices to detect congestion and generate feedback information. Source algorithm, such as TCP, is executed in end hosts or edge devices to adjust sending rate in response to feedback. Accor
6、ding to the end-to-end argument, TCP is traditionally employed to execute most congestion control. With the development of Internet, many limitations of the traditional congestion contril mechnisams emerged. For example, the window decrease policy of TCP is not suitable for streaming media and TCP h
7、as significant performance degradation in high speed networks. To address these problems, several issues about congestion control is studied in this thesis, including source algorithms, decoupling efficiency and fairness control, explicit algorithms over hybrid links which consists of high speed and
8、 wireless links, modeling and stability analysis. The main research works are as follows. A gentle slow start algorithm using bandwidth measurements, called gentle slow start (GSS), is proposed. The slow start of TCP has two shortcomings: firstly, when a connection is established TCP sets a random v
9、alue for slow start threshold (ssthresh); secondly, TCPs exponential window increase policy is too aggressive. GSS addresses these two problems. For a new connection, GSS estimates its attainable bandwidth and uses the bandwidth to set the initial value of ssthresh. During the slow start procedure c
10、aused by timeout, GSS holds an adaptive parameter. When the size of congestion window exceeds the value of the parameter, GSS gently increase the window size and can smoothly convert to congestion avoidance procedure. Under the assumption of synchronous feedback, several fairness and TCP-friendlines
11、s theories of unicast congestion control has been proved, based on the works of R. Sastry and S. Lam et al. These theories are in terms of window 华 中 科 技 大 学 博 士 学 位 论 文 华 中 科 技 大 学 博 士 学 位 论 文 IVincrease/decrease functions. By choosing different window change function, congestion control algorithms
12、 can be customized to suit different application. We demonstrate this by a stream media algorithm. Motivated by the idea of “cross-layer design”, we present the concept of “customizing congestion control strategy” where the application layer can choose its congestion control algorithm and also discu
13、ss some implementation issues. The decoupling of fairness and efficiency control of XCP and VCP are fully analyzed. XCP executes fairness control (FC) and efficiency control (EC) simultaneously in the network devices and conveys much information between end systems and network. In each control inter
14、val, both EC and FC are executed. VCP executes fairness and efficiency control in the end systems and conveys little information between end systems and network. In each control interval, only one controller is executed. Furthermore, we point out that the few executing times of FC results in signifi
15、cant fairness gap between XCP and TCP+AQM style protocol, such as VCP. By increasing the executing times of FC, we propose Fair VCP (FVCP) algorithm which outperforms VCP in fairness. Several adaptive VCP (AVCP) algorithms are presented. By extending VCP, a protocol aimed at high speed links, to wir
16、eless links, AVCP can be adapted to high speed links, wireless link and hybrid of the two links. AVCP is an explicit algorithm. Its key idea is that the sender conjectures the reason of each packet loss, congestion induced or corruption induced, with the help of “path load factor” provided by network and then decides to reduce the congestion window size or not. For explicit algorithms, network needs link capacity to calculate the feedback information. However, it is difficult to estimate accura
