An in-depth analysis of prediction performance based on real network traffic

This paper assesses the predictability of network traffic by considering prediction models usually used in literature. It focuses on the design and the empirical evaluation of the behavior of the ARIMA and the A_SNF models for predicting the input rate of a single link. Via experimentation on real network traffic, we study the effect of some parameters on the prediction performance such as the amount of data needed to identify the model, the inputs of the model, the data granularity, and packet size distribution. Our study reveals that the models provide accurate prediction using only one or two lag. The models are well-identified using only a small amount of data history. Experimental results suggest promising traffic prediction, and generally enhanced predictability if a small granularity is used. We also show that count of large packets is sufficient to predict the throughput.