The regression analysis of the data to determine the buffer size when serving a self-similar packets flow

Abstract

Using the methods of regression analysis on the basis of simulation data, a model for predicting the queue size of the input self-similar packet flow, distributed according to the Pareto law when it is transformed into a flow having an exponential distribution, is constructed. Since the amount of losses in the general case does not give any information about the efficiency of using the buffer memory space in the process of transforming a self-similar packet flow, a quality metric (penalty) was introduced to get the quality of the models after training, which is a complex score. This criterion considers both packet loss during functional transformations and ineffective use of the buffer space in switching nodes. The choice of the best model for predicting the queue size when servicing a self-similar packet flow was carried out using the following characteristics: the coefficient of determination; root-mean-square regression error; mean absolute error; the penalty score. The best in terms of the investigated characteristics are the models using the isotonic regression and the support vector regression