Monitoring packet traffic levels

Traffic levels in packet networks are currently monitored by measuring rates over time scales of the order of 15-30 minutes. Given the burstiness of packet traffic, such time scales are too coarse to accurately indicate resource usage. We relate the traffic monitoring problem to measurement results indicating the fractal nature of packet traffic. We consider the use of fractal statistics in developing accurate yet efficient descriptions of packet traffic. We consider several measurement alternatives, and on the basis of extensive simulation experiments with actual packet traces, we make the following conclusions: (i) the single parameter descriptions of burstiness considered, including the fractal correlation dimension, are inadequate and (ii) the peak-to-mean rates over carefully selected intervals can provide information on capacity exhaust. Useful engineering guidelines can be developed on the basis of a three parameter description, and the self-similarity of packet traffic can in principle be exploited to reduce measurement overhead.