| Internet has evolved from a simple small-scale network to the complex large-scale network with multi-platforms and multi-terminals. But Internet lacks of unified management and control, so that it is unable to describe the performance of network accurately, such as link loss rate, latency and network topology. The traditional method accesses the network performance from interior, which may cause some difficulties such as the cost of computation and communication. In order to overcome these limitations, improve network management and control , then optimize the performance of network, in recent years, a new estimation approach named "Network Tomography" was proposed, and has been concerned by the academic community.The early works of Network loss tomography concern about the link loss of stationary network used mainly some algorithm of statistics. Obviously this approach is unable to track the real time-varying characteristics of the network. In this paper, we propose a link loss model in nonstationary data network, which we threat as a non-stationary random signal, thus the link loss can be tracked time by time.Our estimation method to solve the link loss problem of nonstationary data network is based on recurrent neural network and optimization inversion method. Recurrent neural network is mainly trained by prior information, then establish the map between the end-to-end measurements and the internal link loss rate.Optimization inversion method mainly uses the back-to-back prob packets to collect more statistics information, which can convert the under-constrained problem into over-determined one. Then we use the optimization method to solve the over-determined equations and estimate the non-stationary network link loss rate.Simulations are carried out using NS2 to demonstrate the accuracy of our estimation procedure, which also show the ability of tracking the loss rate of links in real time. We also show the further comparison of these two estimation methods. |
PDF Full Text Request