【作者】 钱峰；

【导师】 胡光岷；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2009， 博士

【摘要】 随着互联网的飞速发展，网络的服务和网络的架构也在发生深刻变化，要成功设计、控制和管理网络，就需要了解和掌握网络的内部特性(包括各种网络特征参数和拓扑结构)。如何在日益复杂的互联网网络环境中实时、精确测量网络性能已成为国内外科学界和工业界共同关注的前沿科学问题之一，该问题是网络行为学的基本研究内容，也是网络优化设计、异常检测与分析、网络监测与评估等工作的基础。传统的网络测量方法需要通过网络节点间的协作来求解，但是随着网络朝着分布化、非协作、异构化管理和基于边缘控制的方向演变，我们发现传统网络测量方法不能完全满足各种应用对网络测量的要求，针对某些特定情况，我们还需要研究能够不依赖网络中间节点协作的网络测量方法。网络层析成像将医学和地震学等领域成功应用的层析成像理论和方法应用于通信网络的测量问题，能够在没有中间节点协作的条件下，将网络测量问题转化为反问题的求解，估计得到大尺度网络内部的各种性能统计信息，是目前备受国际学术界关注的网络测量新技术之一。本文系统地回顾了网络层析成像的起源与各个阶段所取得的重要成果，系统地对网络层析方法进行了分类，指出了在当前网络环境下进行网络层析成像所面临的一些主要问题；从网络层析成像解的稳定性、唯一性、求解实时性和非线性网络层析成像问题出发，针对网络层析成像中所面临的几个实际问题进行了深入研究，在以下几个方面取得了一些有价值的研究成果：1．网络链路时延估计的层析成像方法研究提出了一种基于递归多感知器的网络链路延迟追踪算法，该算法能够在没有任何先验信息条件下追踪非平稳网络链路平均延迟的时变行为，估计链路延迟的概率密度分布，克服了以往方法对网络链路先验模型的依赖，很大程度提高了解的精度和求解的稳定性、实时性。2．网络链路丢包率估计的层析成像方法研究提出了基于包群的单播链路丢包率估计层析成像方法，在根节点一次性向所有叶节点发送探测包群，包群中相邻节点或相近节点可以自由组合成包对、三包组、四包组等，获得路径的丢包率和条件丢包率，构成超定方程组；为了精度和计算复杂度考虑，利用基于奇异值分解的广义逆方法求解超定系统方程得到最后链路的丢包率。该方法与现有的三包组方法相比，可以在探测包数量相同的情况下，通过自由组合，构成包含更多方程的超定方程组，提高了解的精度和稳定性。3．大尺度网络流量矩阵估计方法研究提出了一种基于递归多感知器大尺度网络流量矩阵的追踪算法，首先引入了黑盒子模型对流量矩阵估计问题进行无先验建模；其次，利用递归神经网络对模型进行求解；最后得到整个网络流量矩阵的追踪算法。该方法克服了对先验模型的依赖，提高了估计精度、稳定性和实时性，还便于采用并行算法或网络分布式算法进一步提高求解的实时性。4．网络拓扑结构估计方法研究提出了一种快速的网络拓扑结构的分层聚类估计算法，通过引入了基于高斯混合模型的递归算法，改进了网络拓扑结构估计中相似性量度聚类过程，可以在不降低估计精度的前提下，提高了网络拓扑结构估计的实时性。5．网络流量分类方法研究提出了一种基于混合高斯的网络流量半监督分类方法。与有监督分类方法相比，半监督方法只需要利用少部分的标识数据(大约10％)进行分类，由于标识的数据获得经常是困难的、昂贵的或者花费大量的时间，半监督方法非常适合具体的应用，具有更好的实时性；与无监督分类方法相比，由于使用标识数据进行聚类团的类别确定，具有更高的分类精度。同时，分析了影响半监督分类方法性能的几大因素，优化了半监督分类方法的参数配置，进一步提高了半监督分类方法的分类性能。更多还原

【Abstract】 With the rapid development of the internet, network services and architecture are also undergoing profound changes. To design, control and manage the network successfully, we have to understand and grasp its internal characteristics (including all kinds of network parameters and topology). How to accurate and timely estimate network performance characteristics in more and more complex internet environment have become one of the forefront scientific problems which are focused by the academic community and industry community all over the world; Meantime, this problem is not only the basic research content of network behavior, but also the foundation of network optimize design, anomaly detection and analysis, network monitoring and evaluation and so on. In past, the traditional internet measure method usually resolves such a problem through the cooperation of node, however, as internet toward the evolution of distributive, no corporative, heterogeneous management and edge control, the traditional internet measure method do not meet fully the requirements of various applications on network measurement, for certain specific circumstances, we still need to study the network measurement method without the cooperation of individual servers and routers.By applying tomography theory which is successfully applied in fields such as medicine and seismology to the problems of network measurement in communication network, network tomography can infer the internal performance without the cooperation of internal node and transform network measurement problem into getting the solution of inversion problem in large-scale networks, which has become one of the focused new technologies. In the thesis, we review systematically the past progresses and important achievements at different phases in network tomography research area. The classification of network tomography methodologies are summarized systematically. We also point out some key issues for large scale network tomography. This thesis study profoundly some practical issues of network tomography from nonlinear, stable, unique and real-time solution point of view and results in following innovative achievements: 1. Research on network tomography method for link delay estimationThis thesis propose a novel link delay tracking algorithm based on the recurrent multilayer perceptron (RMLP) network, this algorithm is capable of tracking time-varying average delay behavior in non-stationary network and estimating the probability density distribution of internal delay characteristics without any prior information. Compare with existing delay tomography, RMLP method don’t depend on the queue delay prior model and improve the stability, real-time of solution, obviously.2. Research on network tomography method for link loss estimationThis thesis propose a method of estimating the packet loss rate on unicast network which is based on multi-packet stripe probe. The multi-packet stripe probe is firstly sent to all receivers to cover all paths in the same time. Then, multi-packet stripe probe to the adjacent nodes or similar nodes are free to form two-packet stripe, three-packet stripe or four-packet stripe and we calculate absolute packet loss rate and conditional packet loss rate of the path to construct an over-determined equations. In order to having better accuracy and reducing computational complexity, we choose singular value decomposition (SVD) to calculate the Moore-Penrose generalized inverse matrix of the large-scale over-determined system equations, and finally get the link loss rate estimation. Compare with existing three-packet stripe method, our method can construct over-determined equations with more equations under the same number of probe packet circumstances, and that obtain more stability, flexibility and accurate estimation.3. Research on large scale traffic matrix estimation methodThis thesis propose a novel method of large scale traffic matrix estimation based on recurrent multilayer perceptron (RMLP) network. First of all we introduce the black-box models to model tracking traffic matrix without any prior information, then we use RMLP to resolve the solution of black-box models, finally, we get the entire based-RMLP traffic matrix tracking algorithm. Compare with existing delay tomography, RMLP method don’t depend on the traffic prior model and improve the accuracy of estimates, stability and real-time demand obviously. Meanwhile, by using parallel algorithms and network distributed algorithms, we will further improve the real-time of solution. 4. Research on network topology discoveryBy introducing recursive unsupervised learning of Gaussian Mixture Models (GMM) to topology discovery approach, this thesis proposed a Fast Hierarchical Topology Estimation (FHTE) approach. By improving similarity metrics clustering process in network topology estimation, our proposed method evidently reducing the computation time under the same estimation accuracy circumstances compare to the Hierarchical Topology Estimation (HTE) approach.5. Research on network traffic classificationThis thesis propose a GMM-based semi-supervised classification method. Compare with supervised classification method, our approach takes advantage of only a few labeled data (about 10%) to identify different internet applications. Since obtaining the labeled instances is often difficult, expensive, or time consuming, the approach is particularly well suite for practical application and is more real-time. Semi-supervised classification method obtain more accurate classification compare to unsupervised classification method since it use a small amount of labeled instances to help us identify which Gaussian belongs to some class. Meanwhile, by analysis of semi-supervised classification method factors and their impact on classification performance, an optimum configuration is achieved for the GMM-based semi-supervised classification system, which further improve the classification performance of our approach.更多还原