An empirical evaluation of short-period prediction performance

Traffic prediction constitutes a 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 linear models for predicting the throughput of a single link. In this work, the autoregressive integrated moving average (ARIMA) model and the linear minimum mean square error (LMMSE) are used for predicting. Via experimentation on real network traffic, we study the effect of some parameters on the prediction performance in terms of error such as 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 multi-step prediction that is the number of steps that could be predicted in the future. Besides, we performed a set of predictions based on packets size. Unexpectedly, we find that using more than two lags as inputs for the prediction model increases the prediction error. We find that using the last observation as the predicted value provides the same 1-step prediction performance as ARIMA or LMMSE model. The ARIMA model provides an acceptable multi-step prediction performance. Experimental results show also that there is a granularity value at which the multi-step prediction is more accurate. We also find that the prediction of classified packets based on their size is possible. Especially, throughput of 1,500-byte packets is the less predictable.