Analysis of prediction performance of training-based models using real network traffic

Traffic prediction constitutes a new hot research topic of network metrology. Thus, tuning the prediction model parameters is very crucial to achieve accurate prediction. This work focuses on the design, the empirical evaluation and the analysis of the behavior of training-based models for predicting the throughput of a single link i.e. the incoming input data in Megabit per time interval called granularity. In this work, a neurofuzzy model (alpha_SNF) and the autoregressive moving average (ARMA) model are used for predicting. Via experimentation on real network traffic of different links, we study the effect of some parameters on the prediction performance in terms of error. These parameters are the amount of data needed to identify the model (i.e. training set), the number of last observations of the throughput (i.e. lag) needed as inputs for the model, the data granularity, variance and packet size distribution. We also investigate the use of exogenous variables as inputs for the model. Exogenous variables are variables which are different from the lags such as the number of packets or sampled data. Experimental results show that training-based models, identified with small training set and using only one lag, can provide accurate prediction. We show that counts of packets and especially large packets can be used to efficiently predict the throughput.