【作者】 钱艳平；

【导师】 李奇；

【作者基本信息】 东南大学 ， 控制理论与控制工程， 2006， 博士

【摘要】 自从互联网(Internet)诞生以来,网络资源和网络流量分布的不均衡使得拥塞问题一直困扰着其发展。伴随着网络规模的日益扩大和应用类型的丰富,网络拥塞也变得越来越严重。虽然实践证明基于源端的TCP拥塞控制机制能够有效防止拥塞崩溃地发生,但是TCP拥塞控制机制仍然面临着许多新的危机。因此,互联网的发展要求网络本身也必须参与到拥塞控制中去。目前,基于源端的TCP拥塞控制机制和基于网络端的拥塞避免机制两者相结合已经成为解决拥塞控制问题的一个主要途径之一,形成了计算机网络、通信与自动控制等几个交叉学科一个新的研究热点。Internet拥塞控制可以看作是一个具有通信时延的非线性动态反馈系统。本文着重讨论基于网络端的拥塞避免机制中的主动队列管理算法设计和网络拥塞控制系统的稳定性分析,主要的研究成果如下:(1)提出了一种鲁棒非线性主动队列管理算法,解决了中小规模网络中存在的传输时延和网络参数时变对系统性能的影响问题。首先,由流体流模型推导出网络模型为参数区间不确定一阶时滞线性系统。其次,在确定使闭环系统稳定的非线性主动队列管理算法控制参数集基础上,利用遗传算法获得了基于改进误差绝对值时间积分指标最优的控制参数,解决了主动队列管理算法参数设置难题。基于扩展到时滞系统的棱边定理,设计了鲁棒非线性主动队列管理算法。仿真结果表明该算法具有良好的控制性能,对参数区间不确定系统有较好鲁棒性。(2)提出了一种简单易用的预测PI拥塞控制算法,解决了大规模网络中存在的大时滞问题。首先,利用Smith预估器补偿时延滞后,按Dahlin算法设计控制器,将控制器参数和预估对象模型参数相结合,设计了预测PI拥塞控制算法。其次,分析了系统鲁棒稳定性和存在链路容量干扰时瓶颈队列的暂态、稳态特性。通过仿真验证了预测PI算法控制性能优于RED、PI算法,能够适用于存在较大时延的网络。(3)提出了一种预测PI拥塞控制算法的参数自适应机制来处理网络参数的较大变化问题。首先,利用预测PI算法控制参数与网络参数的确定关系,通过对网络参数的在线估计来实时调节控制参数,使得控制器能够适应网络参数的变化。其次,结合自适应机制分析了整个系统的稳定性和给出了自适应参数的设定原则。仿真结果显示自适应预测PI算法具有较强的鲁棒性,能够适应网络参数的较大变化。(4)提出了一种基于速率的增强自适应虚拟队列管理算法(EAVQ)。EAVQ引入主从拥塞尺度和期望链路利用比的概念,以输入速率为主要拥塞尺度,保留了AVQ响应速度较快、低队列时延、高链路利用率等优点;同时,以期望链路利用比为辅助拥塞准则,设计了一种基于速率的期望链路利用比自适应机制,解决了AVQ存在着参数设定困难、队列抗干扰能力较弱及存在一定的链路损失等缺点,在改善系统动态性能的同时保证了链路容量的充分利用。在线性化基础上给出了一般网络结构下TCP/EAVQ系统的局部稳定条件。通过仿真验证了EAVQ具有极高的链路利用率,极低的分组丢弃率,受控的队列长度、快速的动态响应和对网络参数具有鲁棒性等优点。(5)对一类原始-对偶拥塞控制算法在无时延和考虑时延两种情形下分析了系统的稳定性。基于优化理论框架,提出了一个带边界限制的原始-对偶统一拥塞控制模型,应用该模型更多还原

【Abstract】 The development of Internet has been encumbered with the congestion problem caused by the unbalance distribution of the network resources and the network flows since its naissance. The network congestion becomes more and more serious because of the increasingly expansion of Internet scale and the rapid growth of applied categories. The success of Internet has already proven that TCP congestion control mechanism based on source nodes is effective in the prevention of congestion collapse, while it also faces many new crises. Therefore, the network itself also has to be participated in congestion control. Currently, the combination of TCP congestion control mechanism based on source nodes and congestion avoidance mechanism based on network nodes has becomes a main approach to resolve the Internet congestion control problem. There has come into being a new investigation for some cross-subjects of computer network, communication, and automatic control and so on.Internet congestion control system can be seen as a nonlinear dynamic feedback system with communication delays. This paper emphasized on designs of active queue management (AQM) algorithms in network nodes for Internet congestion control, and stability analysis of network congestion control system. The main research results are as follows:(1) A robust nonlinear AQM algorithm is proposed to resolve the time-delays and parameters time-varying influence on the performance in middle or small-scale network, which is based on a first-order time-delays system with parametric interval uncertainty derived from fluid-flow model. The region of the nonlinear controller parameters making the closed-loop system stable is ascertained and the optimal parameters of the controller is obtained using genetic algorithm with the improved integrated time absolute error (ITAE). The edge theorem extended to time-delays systems is used to design a robust nonlinear AQM algorithm. Simulation results show the controller can achieve favorable performances and is robust against parametric interval uncertainties.(2) A predictive proportional integral (PPI) algorithm for active queue management is proposed to cope with the large delays in large-scale network. In PPI algorithm, Smith predictor is utilized to deal with large delays and Dahlin principle is employed to design controller to reduce the numbers and interactions of the tuning parameters. At the same time, a classical control method is introduced to analyze the stability of the system and the transients as well as steady-state behaviors of bottleneck queue with link capacity disturbances. The simulation results show PPI algorithm’s advantages through comparisons among random early detection (RED) and proportional integral (PI) algorithms.(3) An adaptive mechanism for PPI algorithm is designed to deal with large-scale network parameters immense varying problems. The control parameters are automatically tuned according to on-line estimation of link capacity and traffic load, which makes PPI perform well for a wide-range of network conditions. The local stability of the overall system and the setting principle of adaptive parameters are analyzed using the method of linearization. The simulation results show the APPI algorithm is very robust against network parameters immense variance.更多还原