Tomogravity Space Based Traffic Matrix Estimation In Data Center Networks

With the cloud computing,e-commerce,online games and other Internet applications expanding continuously,more and more business needs to carry out large-scale data storage and data processing,so the network data has been explosive growth.Data center networks(DCNs)are the result of the efficient storage and processing of massive amounts of data.However,the increase in network size and application service types will undoubtedly increase the network operator's burden on management of DCNs.The traffic matrix is a complete description of all traffic status in the network.It can not only provide basic network parameters for scholars to study the flow problems,but also the key input for multiple critical research fields.However,since the behavior of the traffic in the largescale DCNs is unstable and the transmission of the flow between the switches is frequent,it is difficult to estimate traffic matrix directly due to high overhead.Network tomography is a new end-to-end inference technique proposed in recent years which infers end-to-end traffics through easily obtained link data,and there are lots of research results in the traditional computer network.But this technique can not be directly applied to the DCNs due to the differences between the DCNs and the traditional network in terms of traffic characteristics,switch roles,and a large number of redundant paths.However,tree-like DCNs have unique hierarchical features,it is proposed to decompose whole network to reduce the complexity of estimating traffic matrix in DCNs.But the symmetry of tree-like DCNs is easy to cause link data incomplete and inaccurate when collecting observed link data.Therefore,this thesis focuses on the traffic matrix estimation in DCNs based on network tomography theory and presents two tomogravity space based iterative algorithms in different scenarios.The main contents of this thesis are as follows:First of all,we decompose entire network into multiple independent networks unit called cluster and divide a large traffic matrix estimation of entire network into several small traffic matrix estimation to reduce the complexity of estimating the traffic matrix in whole DCNs.In addition,to combine the link information and the gravity model,the coarse-grained traffic characteristics of DCNs and prior traffic matrix can be obtained.By adding additional link information and using similar-Mahalanobis distance to balance the estimation errors,an iterative algorithm(ICGA)of tomogravity space based on traffic characteristics is proposed.Moreover,considering the symmetry of the tree-like DCNs and the moderate link data missing,a simple iterative algorithm(SAWP)is proposed.Finally,network simulator 2(NS-2)platform is set up to generate traffic data.The time complexity of SAWP is simpler than ICGA in case of moderate link data missing.Besides,the simulation results show that the proposed algorithms are more accurate than other algorithms based on the observed link data.In the case of moderate link data missing,it is also found that two proposed algorithms are more similar in terms of the traffic matrix estimation in cluster to cluster pair.With different levels of noise on the observed link measurement,it can be observed that the estimation errors of proposed two algorithms increases but slower in ToR in ToR switch pair than that in cluster to cluster pair because traffics in the cluster are more stable after network decomposition.