【作者】 陈伟杰；

【导师】 王万良；

【作者基本信息】 浙江工业大学 ， 控制理论与控制工程， 2011， 博士

【摘要】 拥塞控制机制在维护网络稳定和保障服务质量上起着至关重要的作用,其主要由基于端到端的传输控制协议(Transmission Control Protocol, TCP)和基于路由的主动队列管理(Active Queue Managment, AQM)组成。部署在路由端的AQM机制能够主动的避免和缓解网络拥塞,有着提高链路利用率、减少丢包率以及降低队列排队延时的优势。但研究表明,AQM算法面对动态突变以及混杂流量的网络场景,存在参数配置难、队列性能不稳定、链路利用率低等不足,进而引起了国内外学术界和工业界的关注。本文围绕着这一系列问题展开了对AQM机制的深入研究。首先搭建了一个基于NS2(Network Simulator 2)的算法设计/性能测试平台,并通过此平台对几个典型的AQM算法进行分析和比较,归纳影响算法性能的主要因素。然后在此基础上,分别提出了3种不同类型的AQM算法。最后,使用丢包概率增益代替复杂的AQM算法控制函数,从非线性分岔理论的角度,着重分析了网络时滞?对拥塞控制系统稳定性及动力学行为的影响。本文的主要工作和研究成果如下:1.针对目前国内外尚未有一个统一的AQM算法设计/测试框架,提出了一个基于NS2的AQM算法设计/性能测试平台架构。该平台包括四个模块:AQM算法引擎模块、性能分析模块、网络拓扑模块和流量突发模块。在分析和总结现实网络中各类拥塞突发场景的基础上,设计了一系列网络测试场景,并在该平台上验证了几种代表性的AQM算法性能。仿真结果表明:网络三元组(N ,RTT ,C )、短时Web流、非响应UDP(User Datagram Protocol)流以及AQM算法的拥塞检测尺度和队列控制函数,都将影响拥塞控制系统的性能。2.针对网络流量突变场景,为提高算法的瞬态响应性能,提出一种带加速因子的自适应SABlue(Self-tune Accelerate Blue)算法。该算法将瞬时队列长度作为早期拥塞检测参量,依据队列负载因子控制丢包步长,实现丢包概率幅度的自适应调整,最终将路由队列长度稳定在目标区域内。为了提高网络突变跨度较大情况时算法的性能,在队列警戒区域内引入了加速因子? ,综合考虑算法各性能指标,最终选取? ? [1.4 3.7]区间。仿真结果表明:由于加速因子的引入,SABlue算法的瞬时队列收敛时间相对PI算法和SBlue算法快,且稳态队列控制性能与PI算法相当;同时,在面对各种突变的网络场景时,SABlue算法的瞬时队列收敛时间较短,链路利用率较高,丢包率较小,且具有一定的鲁棒性。3.针对复杂多变的网络场景,为获得更好的平滑过渡非线性控制输出效果,提出一种基于活动流参数估计的自适应模糊NFL(adaptive Fuzzy-Logic algorithm with active-flow-Number estimation)算法。本文在综合权衡各性能指标的基础上,设计了一组能适应一定网络变化的模糊规则,并对该算法进行了运算优化。为捕获网络突发流,引入了一种基于Bloom Filter的无状态维护活动流参数估计策略,并依此提出一个模糊AQM输出增益补偿器。仿真结果表明:面对负载动态变化时,相对同类型的FAFC算法和FEM算法,NFL算法能较好的适应网络变化,具有更快的收敛速度和稳定的稳态队列控制性能。4.为了解决UDP/TCP混杂流场景中AQM算法存在公平性问题,提出了一种速率感知的多虚拟队列RMVQ(Rate-perceptive Multi Virtual Queue)算法。该算法引入了区分服务思想,为UDP/TCP流维护两个逻辑上独立的虚拟队列,并对它们采用不同的拥塞控制策略。同时,根据活动流感知器提供的UDP流和TCP流负载特征,自适应动态的调整虚拟队列大小。仿真结果表明,相对其他算法,在面对混杂流的场景,RMVQ算法不仅提高了链路吞吐量,也保障了流之间的竞争公平性。5.由于不确定的网络时滞将会造成拥塞控制系统性能不稳定,本文将AQM算法进行模型化抽象,使用丢包概率增益k代替复杂的AQM算法控制函数的输出,来着重研究网络时滞?对时滞GAIMD/AQM(Generalized Additive Increase Multiplicative Decrease/Active Queue Management)系统性能所造成的影响。通过分析时滞GAIMD/AQM系统的特征方程根分布情况,给出了关于时滞?的系统稳定性充要条件。同时,将时滞?作为分岔参数,应用中心流形定理和规范型理论,推得系统的Hopf分岔存在条件、周期解的Hopf分岔方向以及稳定性结论。基于上述结论,给出了拥塞控制参数对GAIMD(? , ? )、网络三元组参数( N , ? , C)、丢包概率函数增益k对GAIMD/AQM系统分岔性质的影响规律。理论分析和仿真结果验证了上述分析的合理性和有效性。更多还原

【Abstract】 Network congestion control mechanism plays a key role in maintaining network stability and guaranteeing quality of service. There are two kinds of congestion control, one is end hosts based (TCP) and the other is intermediate router based (AQM). The AQM has advantages in the enhancement of the efficiency of transfers, network congestion preventing and mitigating, the link utilization improvement, and the packet loss rate reduction, etc. But many researchers find that AQM algorithm is very sensitive to network-load variations, which include burst network traffic, delay jitter and bandwidth limitation, and mixed type traffic. Thus, how to avoid network congestion, increase network traffic performance, and optimize the configuration of network resources to ensure network reliability, timeliness, stability, robustness, arouse the academia and industry’s extensive attention.This thesis focuses on the following topics. First, the platform of AQM based on NS2 is made up. In order to induce the influence factor of the performance of AQM, a series of typical AQM are analysed and compared through this platform. Then, based on the above conclusions, three different types of AQM are proposed. Finally, from the perspective of nonlinear theory, using the packet loss probability gain instead of the complex control function of AQM algorithm, the stability and dynamic behavior of congestion control system is analyzed with the network delay? .The main contributions of the work are as follows:1. There has not existed a universal AQM designing and testing frame yet. A designing and testing platform for AQM algorithm based on Network Simulator 2 is introduced. The platform is divided into the following four parts: The AQM algorithm engine module; the algorithm performance analysis module; network topology module and traffic burst module. Seeking to cover all the contingencies of network congestion, the universal test scenarios are designed for AQM testing based on previous experience. Following an exhaustive simulation evaluation through this testing platform, we test and verify the performance of many AQM. The result shows that the network parameters ( N , RTT ,C ), the short-time Web traffic, the non-response UDP traffic, the congestion detection method and the queue control function of AQM, are all effecting the performance of congestion control system.2. We present a self-tune AQM algorithm with acceleration factor by analyzing Blue algorithm and its variants, which is called SABlue (Self-tune Accelerate Blue). In order to make the queue length keep in the aim area, this algorithm adopts instantaneous queue length as the parameter of incipient congestion detection and calculates the step size of packet drop probability by using load factor. Furthermore, for the sake of response speed, we lead in the acceleration factor in alert area when the network traffic is changed suddenly. Parameter ? can affect the stability of the queue and response rate of burst stream, taking into account the performance trade-offs of the link utilization and packet loss ratio, we choose ? between 1.4 and 3.7. Finally, the simulation demonstrates that SABlue algorithm is more robust, carrying lower packet loss and shorter convergence time in the situation of dynamic traffic. The combination property of SABlue is more excellent than other AQM algorithms.3. In order that the AQM algorithm has smooth control effect, especially in the dynamic network, an adaptive Fuzzy-Logic algorithm with active-flow-Number estimation (called NFL) is poposed. It is composed of two main parts: the fuzzy AQM and the active-flow estimation strategy. Taken the tradeoff among the queue performance, link utilization and other indicators, a set of fuzzy rule is built for NFL to adapt to the dynamic network situation. Furthermore, an optimization mothed is raised, which reduces the computational complexity of fuzzy AQM. A stateless active-flow estimation strategy baesd on Bloom Filter is introduced to capture network congestion status. In order to make up for the deficiency of fixed fuzzy rules, which leads fuzzy AQM robustless, an output gain compensator for fuzzy AQM in accordance with active-flow-number parameter is proposed. Simulation results demonstrate that NFL is adpatvie to dynamic network, while having fast convergence rate and stable steady-state queue control performance. The comprehensive performance of NFL is more excellent than other AQM algorithms.4. In order to solve the fair resources competition issues for TCP flows and UDP flows, an RMVQ(Rate-perceptive Multi Virtual Queue) algorithm is introduced. We use the instantaneous queue length as the primary means, and apply the active flow perceive module as the supplementary means. The active flows perceive module collects flows’information online, and calculates the number of activity UDP, the rate of UDP, and the number of activity TCP periodicly. According to load information, virtual queue distribution module divides buffsize into two parts. One for UDP, anothor is for TCP. The idea of differentiated services is introduced to design the two queue strategies. Simulation results show that, in the UDP/TCP mixed flow scenarios, RMVQ algorithm can catch UDP load trends, and allocate virtual queue accurately. Compared with the other AQMs, the comprehensive performance of RMVQ is satisfactory.5. The uncertain communication delay can cause network congestion control system performance degradation, and even instability. Stability of the equilibrium solution of GAIMD/AQM (Generalized Additive Increase Multiplicative Decrease/Active Queue Management) system is investigated based on analyzing the corresponding transcendental characteristic equation. Using the delay as the bifurcation parameter, we demonstrate that when the delay crosses a critical value, Hopf bifurcation occurs and a periodic solution generates from the equilibrium point. Then, the bifurcation direction and stability of periodic solution is analyzed by means of the center manifold theorem and the normal form theory. Moreover, the impact of the parameters configuration on the performance of the GAIMD/AQM system is investigated exhaustively. Finally, numerical simulations are carried out to verify the feasibility of the theoretical results. The results show that there is a key value of the bifurcation ? 0 for the system. This is consistent with the theoretical results.更多还原