Research On New Estimation Methods Of Traffic Matrix Based On The Prior Model

With the rapid development and wide application of network communicationtechnology, profound changes are also taking place in network service and networkarchitecture, in order to better design, optimize and manage the network, we need tounderstand and grasp the network internal characteristics. As an important indicator ofnetwork performance parameters, traffic matrix describes the specific distribution of thewhole network traffic and provides an important reference for the internal features forthe researcher of network. Moreover, as internet scale is increasingly large and structureis increasingly complex, diversification of the type of network and network securityissues becoming increasingly prominent, direct measurement of the flow matrix methodis becoming more difficult or even unfeasible in today’s networks. Therefore, how to getthe real-time and accurate network traffic matrix in increasingly sophisticated internetenvironment has become one of the frontier issues of common concern among domesticand international scientific community and industry.Traffic matrix estimation problem itself is a less qualitative inverse problem, it hasmulti-solvability, In order to get real correct solution, we need to according to the priorinformation of OD flow to narrow the solution space, then overcoming themulti-solvability of traffic matrix. Firstly, for the shortcomings of existing independentconnection model that applied in the field of multi-service traffic matrix estimation, thisthesis proposes a diffserv IC model. Traditional independent estimation methods are toestimate the total flow of the mixed business network, for different traffic characteristic,there is not correspondent estimation. The diffserv IC model is built on the basis of thetraditional independent connection model, which distinguishes the business traffic innetwork. According to the characteristic of inconsistent distribution structure of thedifferent business traffic in the whole network, this model makes use of the OD existindicating matrix and the prior information which provided by business traffic edgeaccess link to get different business traffic matrix. Compared with the traditionalindependent connection model, the diffserv IC model can reflect the distribution ofbusiness traffic in the network better, and the estimated accuracy is higher than the simple independent connection model.Secondly, due to the specific distribution of real OD flow of existing traffic matrixestimation methods which uses simple Gaussian, Poisson distribution model, can not bea better reflection of the actual network, the complex mixture of Gaussian distributionmodel, which although is able to describe the real OD flow distribution, but it increasesthe computational complexity of the traffic matrix estimated. For the shortcomings ofOD flow prior distribution model among existing estimation methods, this thesisproposes a traffic matrix estimation method based on the mixed distribution model. Themixed distribution model can well describe the real probability distribution of ID flowin actual network, especially for Non-Gaussian distribution OD flow described morereasonable and obtained the accurate result when estimated the flow between theNon-Gaussian distribution OD pairs. Meanwhile, this thesis combines the advantage ofthe generalized gravity model applied in estimating Gaussian distribution OD flow withthe accuracy of mixed distribution model applied in estimating Non-Gaussiandistribution OD flow to get an optimized traffic matrix estimation method of mixeddistribution model, and successful to simulate and analysis it.Finally, using the real network traffic data, this thesis does some simulations andanalysis for business distinguished IC model and traffic matrix estimation method basedon hybrid distribution model. According to the simulations and comparison with othertraffic matrix estimation methods, we conclude that the proposed business distinguishedIC model and Fourier domain hybrid model based traffic matrix estimation method inthis thesis get a better estimation results, which is closer to the real network value.