Generalized Linear Inversion In Network Tomography

With the rapid development of the Internet, the structure of the network has changed thoroughly. It is necessary to get into the inner property of the specific network to design, control and manage it successfully. Traffic Matrix and link-level packet loss rate are the important parameters to indentify the performance of the network. But it is hard to measure them directly as the Internet is becoming massive, distributed and heterogeneous.Network Tomography is a performing technology to infer the network parameters through indirect measurement. MLE is widely used in network tomography field to assure the accuracy of the result, but it brings in the problem of stability and usability which is the main bottleneck to push the network tomography estimation into practical use.The stability and the usabilty of the network tomography techonology is dicussed here. We introduce the GLI algorithm into network tomography field to estimate OD traffic and link-level packet loss rate. Some improvemet are made in GLI to make the algorithm more stable and more practical.OD traffic estimation is one of the important aspects of network tomography. We emphasize on the existing algorithms and propose a new algorithm which is based on traffic covariance. This novel algorithm is proved to be effienct and fast implemented with GLI and the slipe-window scheme. We use GLI to estimate the TM in large scale network. Different constrains based on initial prior and constrains based on partial measurement are discussed. A new method based on state prediction and another method in which history mean is combined with gravity model is proposed. Under the implementation of the different constrains and incorporated with partical measurement, we get appealing result, which is better that general gravity method.Link-level packet loss rate is an important parameter of network performance.lt is valuable in management of the network. GLI is introduced into this field. A new method is proposed to estimate link-level packet loss rate under general SPT while exiting method can only solve this problem under constant binary or three-leaf tree.