Forecasting network traffic using FARIMA models with heavy tailed innovations

This paper presents a data traffic model capable of describing the long-range as well as short-range dependence structure of packet data traffic. Specifically, we use the fractionally integrated autoregressive-moving average (FARIMA) process with non-Gaussian white driving sequence to describe packet arrival rate in a unit time. We introduce a procedure to estimate the fractional differencing parameter and ARMA coefficients: this procedure uses a cepstrum approach and does not require any prior knowledge about the driving noise distribution and the type of ARMA system. Since the main purpose of workload modeling is to aid in network performance evaluations, we are particularly interested in using the FARIMA model to predict bandwidth requirements for network traffic. We propose a dynamic bandwidth allocation strategy by employing linear predictors designed based on the estimated FARIMA parameters