【作者】 陈亮；

【导师】 张宏；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2011， 博士

【摘要】 无线自组网(Ad Hoc网络)是由一组带有无线收发装置的移动节点组成的、多跳、临时性自组织网络系统。Ad Hoc网络中的每个节点既可能作为发送数据流的源主机,也可能作为转发数据流的路由器。无线网络多跳、多对一的通信方式及无线链路质量,都容易引起网络的局部或全局拥塞。由于Ad Hoc网络的带宽资源非常有限,因此拥塞研究显得十分重要。论文分别从拥塞控制的源端算法与链路算法两个方面,重点研究了Ad Hoc网络的TCP拥塞控制机制。本文首先分析了主动队列管理(Active Queue Management, AQM)的比例积分(Proportional Integral, PI)与比例积分微分(Proportional Integral Differential, PID)算法的稳定性与鲁棒性；其次建立了Ad Hoc网络的TCP/AQM微分模型；再次设计了一种比例求和微分(PSD)的神经元PID主动队列管理控制器；最后基于拥塞控制原理,建立了Ad Hoc网络TCP性能模型。论文主要研究内容如下：(1)主动队列管理PID算法的稳定性是实现其拥塞控制的基础。目前的PID设计及整定大多基于经验和试凑,往往得到一些孤立的整定结果,缺乏稳定区域的理论分析。针对Ad Hoc网络的无线与时滞特点,分析了PID算法在时延Ad Hoc网络中的稳定性,在不同的微分系数下,分别给出了时延系统PID-AQM控制的稳定区域。与传统的工程整定比较,稳定区域研究提供了时滞系统的PID稳定理论依据,为整定PID参数带来便利。通过Matlab和NS (Network Simulator)仿真,验证了稳定区域的结论及优越性。(2)由于研究者建立的控制对象模型只能是实际物理系统不精确的表示,在模型不精确或控制对象发生变化的条件下,控制系统仍能保持原有的控制性能,这是鲁棒性控制的目标,因而鲁棒性也是Ad Hoc网络AQM控制算法的重要性能指标。目前AQM的PI算法大多基于经验和试凑来设计和整定控制器系数,而作为控制对象的Ad Hoc网络,其环境参数经常变化,控制器系数能在多大程度上保持系统稳定,还缺乏鲁棒性的理论分析。根据鲁棒控制理论,分析了PI算法在时延Ad Hoc网络中的鲁棒性,推导了PI控制器确定时的链路容量、TCP连接数量和时延之间的关系,给出了某个PI控制参数下的时延R0的变化范围。通过Matlab和NS仿真,验证了时延参数鲁棒性的范围。(3)主动队列管理研究通常关注队列控制器设计,而作为被控对象,传输控制协议(TCP)往往利用NS仿真实现,Ad Hoc网络的TCP机制与AQM的相互关系尚不明确,因此有必要研究Ad Hoc网络TCP及AQM特性。基于TCP窗口加性增、乘性减规则及排队原理,推导了TCP窗口及队列的微分方程,再基于比例积分的AQM控制,推导了拥塞丢弃概率的微分方程,通过联立微分方程组,提出了Ad Hoc网络TCP/AQM微分模型。与NS的对比仿真显示,新模型能较好地估计Ad Hoc网络性能,基于本模型的研究也表明,网络跳数、无线丢失和过小的队列成为AQM性能瓶颈,队列信息则有助于TCP区分Ad Hoc网络的拥塞丢弃与无线丢失。(4)在AQM众多控制算法中,神经元PID算法能较好地控制队列长度,但其神经元增益对被控对象的状态较为敏感,恒定的神经元增益设定往往使控制效果难以保证。基于TCP窗口加性增、乘性减规则及AQM原理,推导了TCP窗口、拥塞丢弃概率及队列长度的微分方程。对该微分方程使用小扰动线性化理论,获得Ad Hoc网络TCP/AQM拥塞控制系统模型。基于该控制系统模型,将递推计算修正功能引入神经元PID,设计了一种神经元自适应PSD (Proportional Summation Differential)的AQM,该算法可以根据网络对象状况在线调整神经元增益。NS仿真表明,在无线分组丢失、突发流及链路容量变化的Ad Hoc网络中,PSD队列管理性能优于神经元PID。(5)由于Ad Hoc网络的多跳和无线信道特性,Padhye提出的有线TCP Reno模型不能准确反映Ad Hoc网络的吞吐量,而目前的Ad Hoc网络TCP性能建模往往利用马尔科夫链。基于802.11协议DCF (Distributed Coordination Function)的RTS/CTS (Request To Send/Clear To Send)通信机制,推导了多跳拓扑的可用链路容量,根据TCP的Tahoe版本及Reno版本拥塞窗口规则,分别建立了TCP窗口、可用链路容量及分组丢弃概率的数学关系,由此获得Ad Hoc网络TCP Tahoe与Reno的性能模型。仿真研究表明,新模型较好地估计了Ad Hoc网络TCP窗口及网络吞吐量,平均误差低于7%,另外,Reno的快速恢复算法无法更正关联丢失的所有分组,最终触发超时重传,因此在Ad Hoc网络中性能劣于Tahoe。更多还原

【Abstract】 Ad Hoc network is a multi-hop, temporary and self-configuring network of mobile devices with wireless receive-send equipment. Each node acts either as a host generating flows, being the destination of flows from other mobile nodes, or as a router forwarding flows directed to other nodes. Its features of multi-hop, one against many communication method and wireless link quality will lead to network local and global congestion. Ad Hoc network bandwidth resource is very limited, so congestion control research is very important.In this dissertation, an emphasis research has been made on the Ad Hoc network TCP congestion control from congestion control source algorithm and link algorithm. This paper first analyzes the stability and robustness of PI (Proportional Integral) AQM (Active Queue Management) and PID (Proportional Integral Differential) AQM in Ad Hoc network. Secondly, it deduces the TCP/AQM differential model. Thirdly, it designs an adaptive neuron PID AQM controller based on proportional summation differential (PSD). Finally, it models the TCP performance in Ad Hoc network based on congestion control principle. The main research content is as follows:(1) Stability of PID AQM algorithm is base of congestion control. Now PID design and tuning often use experience and experimentation. These methods lack theoretic analysis of stability region and often get some isolated tuning results. Stability characterization of PID scheme is analyzed in wireless and delay Ad Hoc network. Finally, it gives stability regions of PID-AQM control under different differential coefficients in delay networks. Comparing with tradition engineering tuning ways, the research also gives stability theory of PID in delay system and helps to determinate controller coefficients. Matlab and NS (Network Simulator) simulations indicate that stability regions and its advantage are proved.(2) The object model which researchers deduce is not accurate expression of the real physical system. Robust control goal is to maintain system stability and performance when a controlled object changes its parameters or modeling is inaccurate. Robustness of the PI AQM algorithm is important performance target. Now PI AQM often uses experience and experimentation to design controller and its algorithm. As controlled object, Ad Hoc network parameters often change. So how far can the controller coefficients maintain system stability? This research lacks theoretic analysis of robustness. According to robustness control theory, it analyzes robustness characterization of PI algorithm in delay Ad Hoc network. It deduces the relation of link capacity, number of TCP connection and delay for certain PI controller. Finally, it gives range of the delay under the PI controller’s parameters. By Matlab and NS simulations, the range of delay parameter robustness is proved.(3) The research on AQM is usually concerned about queue controller design. As a controlled object, transmission control protocol (TCP) is often realized by Network Simulator (NS) simulation. So it is necessary to study the character of TCP and AQM in Ad Hoc network because the relation between TCP and AQM in Ad Hoc network is not clear. TCP windows size and queue length differential equations are deduced based on TCP window additive-increase multiplicative-decrease rule and queuing principle. Then, congestion loss probability differential equation is deduced based on proportional integral AQM control. So the Ad Hoc network TCP/AQM differential model is proposed through building the simultaneous differential equations. The comparison simulations show that the new model can estimate Ad Hoc network performance well. The model research also shows that the number of hops, wireless loss and very small queue become AQM performance bottlenecks. Furthermore, the queue information can help TCP discriminate between congestion loss and wireless loss in Ad Hoc network.(4) Neuron PID algorithm can control queue length successfully in many AQM scheme, but its neuron gain is sensitive for controlled object state. So it is difficult to guarantee the control performance because of the fixed neuron gain in neuron PID algorithm. Congestion window size, loss probability and queue length differential equations are deduced based on TCP window additive-increase and multiplicative-decrease (AIMD) principle and AQM mechanism. TCP/AQM control system model is obtained in Ad Hoc network through the small perturbations and equations linearization. Then it introduces recursion and modification gain to neuron PID based on the model. Finally, a neuron adaptive proportional summation differential (PSD) AQM scheme is proposed. PSD algorithm can modify neuron gain dynamically according to the network situation. NS simulations demonstrate that PSD queue management performance is better than neuron PID under conditions of wireless packet loss, sudden flow and different link capacity.(5) The Reno wired throughput model proposed by Padhye can not estimate accurately Ad Hoc network throughput because Ad Hoc network has characteristics of multi-hop and wireless channel. But Ad Hoc network TCP performance modeling use Markov chain usually. Based on 802.11 protocol DCF (Distributed Coordination Function) RTS/CTS (Request To Send/ Clear To Send) communication mechanism, multi-hop topology available link capacity is deduced. Then TCP windows size, available link capacity and wireless packet loss probability mathematical relation is deduced based on congestion window principle of TCP-Reno version and TCP-Tahoe version. Therefore Ad Hoc network TCP-Tahoe model and TCP-Reno model were proposed. The simulation research shows that the new TCP models can estimate Ad Hoc network TCP window size and throughput. The models average error is below 7%. Reno’s fast-recovery can not correct all packets in correlated loss and leads to time-out retransmit at last. So Reno performance is worse than Tahoe in Ad Hoc network.更多还原