IP Traffic Classification via Blind Source Separation Based on Jacobi Algorithm

By distinguishing long and short TCP flows, we address in this paper the problem of efficiently computing the characteristics of long flows. Instead of using time consuming off-line flow classification procedures, we investigate how flow characteristics could directly be inferred from traffic measurements by means of digital signal processing techniques. The proposed approach consists of classifying on the fly packets according to their size in order to construct two signals, one associated with short flows and the other with long flows. Since these two signals have intertwined spectral characteristics, we use a blind source separation technique in order to reconstruct the original spectral densities of short and long flow sources. The method is applied to a real traffic trace captured on a link of the France Telecom IP backbone network and proves efficient to recover the characteristics of long and short flows.