Modelling H.264/AVC VBR video traffic: comparison of a Markov and a self-similar source model

We compare the adequacy of two models for realistic video sources, namely the fractional Brownian motion (fBm) and Markov modulated fluid flow (Mmff) models. We use the effective bandwidth approach to get the probability that the buffer content exceeds a certain threshold. We use a formula in which the variance function, i.e., the variance of the traffic arriving in an interval of length tau, plays a central role and we model it as variance associated either with the Mmff or fBm model. We measure the variance function for an artificial source we construct, using Variable Bit Rate (VBR) H.264/AVC (Advanced Video Codec) video traces of real movies. There is a good correspondence between the buffer threshold exceeding probability obtained via trace-based simulations and the one predicted theoretically, based on the measured variance function. These we take as benchmark results against which we check both models-fBm and Mmff. First, we try to tune the model parameters such that their variance function matches the measured one over a large range of tau values, but this proves to be difficult. When matching the variance function over a short range of the tau of interest, we conclude that the Mmff model is better suited to model VBR video sources. We conduct an error sensitivity analysis with non-optimal model parameters and conclude that they do not influence considerably the buffer exceeding probability. Most reliable results are achieved with the measured variance function; if by some reason it needs to be modelled, the Mmff is preferred over the fBm model.