网络流量矩阵估计方法研究

发布时间：2018-05-13 16:34

本文选题：流量矩阵 + 流量工程　；参考：《华中师范大学》2014年硕士论文

【摘要】：网络流量矩阵反映了网络内部每一对源至目的网络节点间的流量大小。许许多多网络工程和网络管理项目如负载均衡、拥塞控制、网络安全等都建立在流量矩阵基础之上。因此,流量矩阵具有非常大的实际意义。但是,在目前的实际网络中,由于不同网络设备厂商所生产的设备对于流量测量功能的不同支持,通过直接测量获得准确的流量矩阵是十分耗时、耗资的。相比之下,通过结合数学方法进行流量矩阵估计的方法变得更为可行。但是,尽管几十年来许多学者对流量矩阵估计问题付出了相当多的努力,但是准确的估计方法依然没有能够产生。本文首先综述了目前国际国内关于流量矩阵估计的现有研究,在比较和分析了目前存在的许多有代表性方法及其存在的缺陷之后,分别提出了从方法原理、数学模型、现代优化理论三个不同研究角度产生的三种流量矩阵估计方法,进而详述了这三个方法的思路由来、模型建立、优缺点分析。最后,采用Abilene网络真实数据的实验分别证实了这三种方法的有效性,误差都减少了一半以上。第一种所提出的方法称为Advanced-Tomogravity。这个方法建立在精确的流量重力特征模型和tomography方法之上。通过引入相关因子向量参数至目前存在的流量重力特征模型,提出了精确的流量重力特征模型,进而可以针对具体的某个网络来设置相应的向量参数值,实现估计准确性的提高。通过数学的分析与公式推导,获得了该相关因子向量参数明确的赋值表达式。这个表达式的推导过程用到了广义逆和最小二乘解等相关的数学基础理论。采用了美国Abilene网络真实网络数据的仿真实验验证了所提出方法的有效性。仿真实验结果证实该方法不仅能够更好的追踪流量大小波动特性的能力,而且能够更准确的逼近流整体的平均值趋势。第二种方法Tomofanout建立在所提出的结合边缘链路负载信息的Fanout模型之上。新的Fanout模型具有原来Fanout模型的性质,并且采用了边缘链路负载信息,因而具有更佳的准确性。另外,通过该模型获得的估计结果进一步由期望最大化迭代进行处理以符合流量矩阵估计的初等模型。第三种方法称为MNETME,通过该方法采用了路由矩阵的广义逆与链路负载向量的乘积作为神经网络的输入来进行训练与预测。并且该方法结合了期望最大化迭代作为对神经网络输出数据的进一步处理。得益于这些,与同类方法相比,该方法用于训练的数据量少,而结果却更准确。
[Abstract]:The network traffic matrix reflects the traffic size of each pair of source to destination network nodes in the network. Many network engineering and network management projects such as load balancing congestion control and network security are based on the flow matrix. Therefore, the flow matrix has great practical significance. However, in the actual network at present, it is very time-consuming and expensive to obtain accurate flow matrix by direct measurement because of the different support of the equipment produced by different network equipment manufacturers for the flow measurement function. By contrast, it is more feasible to estimate the flow matrix with mathematical method. However, although many scholars have made considerable efforts to estimate the flow matrix in recent decades, accurate estimation methods have not been produced. In this paper, the current research on the estimation of flow matrix is summarized. After comparing and analyzing many representative methods and their defects, the paper puts forward the method principle and mathematical model, respectively. Three kinds of flow matrix estimation methods from three different research angles of modern optimization theory are presented, and the origin of the three methods, the establishment of models and the analysis of their advantages and disadvantages are described in detail. Finally, the validity of the three methods is verified by using the real data of Abilene network, and the error is reduced by more than half. The first method proposed is called Advanced-Tomogravity. This method is based on the accurate flow gravity characteristic model and tomography method. By introducing the correlation factor vector parameter to the current flow gravity characteristic model, an accurate flow gravity characteristic model is proposed, and the corresponding vector parameter value can be set for a specific network. Improve the accuracy of estimation. Through mathematical analysis and formula derivation, the explicit assignment expression of the parameters of the correlation factor vector is obtained. The derivation of the expression uses the basic mathematical theories such as the generalized inverse and the least square solution. The effectiveness of the proposed method is verified by the simulation of the real network data of the Abilene network in the United States. The simulation results show that the proposed method can not only track the fluctuation characteristics of the flow, but also approach the average trend of the whole flow more accurately. The second method, Tomofanout, is based on the proposed Fanout model combining edge link load information. The new Fanout model has the property of the original Fanout model and adopts the edge link load information, so it has better accuracy. In addition, the estimated results obtained from the model are further processed by the expectation maximization iteration to fit the primary model estimated by the flow matrix. The third method is called MNETME.The method uses the product of the generalized inverse of the routing matrix and the link load vector as the input of the neural network to train and predict. The method combines the expected maximization iteration as the further processing of the output data of the neural network. Because of these, compared with the similar method, this method has less data for training, but the result is more accurate.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.06

【相似文献】