Modeling and performance evaluation of self-similar behavior of MPEG-4 video traffic generators

Abstract

Variable bit rate (VBR) Moving Pictures Expert Group (MPEG) video is one of the major applications used on networks. Effective design and performance analysis of such networks thus depends on accurate modeling of MPEG video traffic. Recent studies of real traffic data in modern computer networks have shown that traffic exhibits long-range dependence (LRD) properties over a wide range of time scales. The predominant way to quantify the long-range dependence is the value of the Hurst parameter, which shapes the autocorrelations of LRD processes, and it is needed for determining variance of such a process. Thus, correct and efficient estimation of the Hurst parameter is important in traffic analysis. There are several different methods to estimate the Hurst parameter, we have evaluated the most commonly used methods for estimating the self-similarity parameter H using appropriately long sequences of data. Estimators considered include, wavelet-based, R/S-statistic, variance-time, absolute-based and Periodogram-based. Our results have pointed to the wavelet-based estimator and R/S-based estimator as the most efficient estimators of the Hurst parameter. There is still a considerable debate about how to model the VBR video traffic. Many models have been proposed for the modeling of traffic sources. However, further work is required to verify their accuracy for modeling the MPEG video traffic. This thesis evaluates three different traffic source models, namely the Hosking-based model, the RMD-based model and chaotic mapbased model in terms of their statistical characteristics and to compare their outputs with the empirical traces to validate their effectiveness for modeling the MPEG-4 video traces. A comparison of the packet loss rate, queuing delay and throughput performance of RMD-based generator and chaotic-based generator with the performance of the real trace is used in validation of the models. Our simulation results show that the chaotic-based model capture the statistical characteristic of empirical traces better than Hosking-based and RMD-based models