Modeling Internet traffic using nonGaussian time series models

Abstract

Internet traffic is usually represented by a time series of number of packets or number of bits received in each time slot. There exists a class of Internet traffic traces that have slowly decreasing autocorrelation, their marginal distributions of the number of packets are fit by negative binomial distributions and the time series of number of bits are fit by Gamma distributions. To model this class of traffic, this paper divides the traffic input stream into several sub-streams by decomposing their autocorrelation functions, and models each substream as a negative binomial time series or a Gamma time series. The proposed models can simultaneously capture the autocorrelation and the marginal distribution. A queue performance criterion is used to validate the models.