Traffic matrix is of great help in many network applications. However, it is
very difficult to estimate the traffic matrix for a large-scale network. This
is because the estimation problem from limited link measurements is highly
underconstrained. We propose a simple probability model for a large-scale
practical network. The probability model is then generalized to a general
model by including random traffic data. Traffic matrix estimation is then
conducted under these two models by two minimization methods. It is shown
that the Normalized Root Mean Square Errors of these estimates under our
model assumption are very small. For a large-scale network, the traffic
matrix estimation methods also perform well. The comparison of two
minimization methods shown in the simulation results complies with the
analysis.
Traffic matrix estimation using the Levenberg-Marquardt neural network of a large IP system
