NETWORK TRAFFIC MODELING AND PACKET-LOSS PROBABILITY APPROXIMATION

摘要

In this paper, we analyze network trace data in order to construct adequate mathematical models of traffic. By estimating the spectral density function of a real network traffic, we found it to be very close to the spectral density function of a self-similar process. In this way, we used a Gaussian fractional noise as a first model for network traffic, since a Gaussian fractional noise is the simplest and most natural model for any self-similar process. But, as we found from realizations of this model as well as from the buffer-overflow probability behavior, it is too far from reality. As a second model for network traffic, we considered a Poisson-Pareto model, Poisson arrivals with Pareto distribution of sessions lengths. This model has too many parameters for complete description. Therefore, instead of trying to estimate these parameters, we proposed some more general parameters which can be easily estimated from the network trace. In terms of these parameters, approximations for buffer-overflow probability or packet-loss probability can be computed. Finally, we proposed a way to recompute the buffer-overflow probability by splitting the original process or mixing it with another independent process.