数据中心网络中基于层析重力空间的流量矩阵估计

发布时间：2018-04-16 13:32

本文选题：数据中心网络 + 流量矩阵估计　；参考：《西南大学》2017年硕士论文

【摘要】：随着云计算、电子商务、网络游戏等Internet应用领域的不断延伸和扩展,目前越来越多的应用需要进行大规模的数据存储和应用处理,网络中的数据已然出现爆炸式的增长。数据中心网络就是随着人们对海量数据的高效存储和处理要求应运而生的。然而,扩展的网络规模和种类繁多的应用服务类型加重了网络操作员对数据中心网络管理的负担。流量矩阵可以完整描述网络中的全部流量状态信息,它不仅可以为学者们研究网络中流量问题提供基本的网络参数,还是多个重要领域的关键输入。但是因为数据中心网络中网络规模较大、结构复杂,网络中的流的行为不稳定,流交互非常频繁,所以直接测量数据中心网络中端到端的流量是非常困难的,并且需要花费较大的开销。网络层析成像技术是近年来提出的一种新的推断网络端到端的测量技术,它是通过易取得的链路数据推断端到端的流量,目前在传统的计算机网络中已有大量的研究成果,然而由于数据中心网络与传统网络在流量特征、交换机扮演角色、大量冗余路径等方面的不同,该技术不能直接应用在当前的数据中心网络。目前针对树型的数据中心网络结构中独特的分层特点,采用分解网络的方式可以降低估计整个网络流量矩阵的复杂性。然而,树形结构的对称性又容易使得收集链路数据过程中存在数据的不完整和不准确性,链路测量误差会对估计误差造成一定的影响。因此,本文主要将层析成像技术和流量矩阵估计作为核心研究问题,提出了拓扑分解下的基于层析重力空间的数据中心网络流量矩阵估计算法。本文的主要研究内容如下:首先,为了降低数据中心网络流量矩阵估计的复杂性,提出将整个网络分解为多个相对独立的网络单元,称之为簇,从而将估计整个网络的流量矩阵降解为估计多个小的网络单元的流量矩阵。其次,结合链路信息和重力模型结合得到数据中心网络的粗粒度流量特征和简单的流量矩阵估计,通过加入附加的链路信息和采用类马氏距离衡量估计误差,提出基于流量特征的层析重力空间的迭代算法(ICGA)。此外,考虑到树形数据中心网络结构具有的对称性和收集得到的链路数据存在适量数据丢失和错误的情况,提出未使用数据中心网络先验流量特征的简单层析重力空间流量矩阵估计算法(SAWP)。最后,搭建了Network Simulator2(NS-2)仿真平台模拟整个实验环境。结果表明:通过对比分析算法的时间复杂度,表明在适量数据丢失下第二种方法比第一种方法更加简单。其次,仿真表明所提的算法在实际测量数据中比其他算法估计更加准确;在少量数据丢失的情景下,提出的两种算法在簇间的流量矩阵估计下性能更相似;当对获得的测量数据加入不同层次的噪声之后,可以发现估计的误差随着噪声层次的增加而增加,但是因为分解之后的网络流量相对更稳定,簇内的误差增加更缓慢。
[Abstract]:With cloud computing, e-commerce, online games and other Internet applications continue to extend and expand, more and more applications at present the need for data storage and processing of large-scale application, the data in the network has emerged. The explosive growth of data center network is people with efficient storage and processing of massive data requirements however came into being. The expansion of network scale, and a wide variety of application service types increased the network operator to the data center network management burden. Traffic matrix can be a complete description of all of the traffic state information, it can not only provide the basic parameters of the network flow problem for scholars in the network, or a number of important areas. But the key input because the size of the network data center network in large, complex structure, network flow behavior of unstable flow interaction is very frequent, So the direct measurement data center network end-to-end flow is very difficult, and takes a large overhead. Network tomography is a new inference of network end to end measurement technology, it is easy to get through the link data from end to end flow, at present in the traditional the computer network has a large number of research results, however, because the data center network and traditional network traffic characteristics play a role in the exchange, a large number of redundant paths, and so different, the technology can not be directly applied in the data center network at present. The unique characteristics of the hierarchical network structure of data center according to the type of tree, the decomposition of the network the method can reduce the complexity of the whole network traffic matrix estimation. However, the symmetry of the structure of the tree and easily makes the data are collected in the process of data link Incomplete and inaccurate link, the measurement error will cause a certain impact on the estimation error. Therefore, in this paper the tomography technology and traffic matrix estimation is the core issue, put forward the topological decomposition algorithm under gravity space chromatography data center network based on traffic matrix estimation. The main contents of this paper are as follows: firstly, in order to to reduce the complexity of the data center network traffic matrix estimation, the whole network is divided into a plurality of independent network unit, called clusters, so as to estimate traffic matrix of the whole network solution for reducing the traffic matrix estimation of multiple small network elements. Secondly, combined with the link information and the gravity model with coarse grain flow characteristics get the data center network and traffic matrix simple estimation, by adding additional link information and using the Mahalanobis distance measure estimation error is proposed Gravity flow chromatography iterative algorithm based on feature space (ICGA). In addition, taking into account the amount of data loss and error of the existing link data with tree network structure of data center symmetry and collected, this simple chromatography gravity space traffic matrix without the use of data center network traffic characteristics a priori estimation algorithm (SAWP) finally, set up the Network Simulator2 (NS-2) in the experimental environment simulation platform. The results show that: through the comparative analysis of the time complexity of the algorithm, that is more simple than the first method of the two methods in the amount of lost data. Secondly, the simulation shows that the proposed algorithm in the actual measurement data in the estimation is more accurate than other algorithms in a small amount; data loss situation, traffic matrix, the two algorithms in the inter cluster estimation performance is more similar to the measured data obtained; when the addition of After the same level of noise, it is found that the estimation error increases with the increase of noise level. However, because the network traffic after decomposition is relatively stable, the error in the cluster increases more slowly.

【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.06

【参考文献】