Choosing an Accurate Network Model using Domain Analysis

Network link simulation is perhaps the most common method for evaluating application and network protocol designs. In modeling realistic networks, researchers face measurements whose characteristics experience non-stationarity (time variability) and complex patterns due to a number of factors, including both internal network elements and external events. In this paper we introduce a methodology we call domain analysis to quantify the accuracy of different models, and show how to use it to choose the best model for a given set of network characteristics. Our work seeks to aid network and application protocol developers in developing and choosing appropriate models for network simulation. We introduce our data preconditioning methodology for modelling non-stationary datasets, and present the new multiple states MTA model (MMTA). We show that it is better in capturing error burst statistics than classical models and more consistently accurate across different networks that our previous MTA model