Time series models for internet traffic

Data traffic sequences from two campus FDDI rings, an Ethernet, two entry/exit points of the NSFNET, and sub-sequences belonging to popular TCP port numbers on one of the FDDI rings indicate that appropriately differenced time-series generated from these traces can be modeled as auto-regressive moving average (ARMA) processes. The variates of the ARMA filter are, however, non-Gaussian. A sequence of steps leading through (i) parameter estimation, (ii) generating the distribution of the variates, (iii) forecasting tail percentiles, and (iv) synthetic generation of non-negative integer sequences is presented. The data indicates that parameter estimates drift slowly with time and may need to be re-computed periodically for accurate forecasts. The forecasting algorithm, has potential application in dynamic resource allocation. The synthetic traffic generation algorithm may be used in simulation studies of resource management algorithms.