Video traffic modeling and delivery

Video is becoming a major component of the network traffic, and thus there has been a great interest to model video traffic. It is known that video traffic possesses short range dependence (SRD) and long range dependence (LRD) properties, which can drastically affect network performance. By decomposing a video sequence into three parts, according to its motion activity, Markov-modulated self-similar process model is first proposed to capture autocorrelation function (ACF) characteristics of MPEG video traffic. Furthermore, generalized Beta distribution is proposed to model the probability density functions (PDFs) of MPEG video traffic. It is observed that the ACF of MPEG video traffic fluctuates around three envelopes, reflecting the fact that different coding methods reduce the data dependency by different amount. This observation has led to a more accurate model, structurally modulated self-similar process model, which captures the ACF of the traffic, both SRD and LRD, by exploiting the MPEG structure. This model is subsequently simplified by simply modulating three self-similar processes, resulting in a much simpler model having the same accuracy as the structurally modulated self-similar process model. To justify the validity of the proposed models for video transmission, the cell loss ratios (CLRs) of a server with a limited buffer size driven by the empirical trace are compared to those driven by the proposed models. The differences are within one order, which are hardly achievable by other models, even for the case of JPEG video traffic. In the second part of this dissertation, two dynamic bandwidth allocation algorithms are proposed for pre-recorded and real-time video delivery, respectively. One is based on scene change identification, and the other is based on frame differences. The proposed algorithms can increase the bandwidth utilization by a factor of two to five, as compared to the constant bit rate (CBR) service using peak rate assignment