Modeling and Simulation of Self Similar Variable Bit Rate Compressed Video
Variable bit rate (VBR) compressed video is expected to become one of the major loading factors in high-speed packet networks such as ATM-based B-ISDN. However, recent measurements based on long empirical traces (complete movies) revealed that VBR video trafic possesses self-similar (or fractal) characteristics, meaning that the dependence in the trafic stream lasts much longer than traditional models can capture. In this paper, we present a unified approach which, in addition to accurately modeling the marginal distribution of empirical video records, also models directly both the short and the long-term empirical autocorrelation structures. We also present simulation results using synthetic data and compare with results based on empirical video traces. Furthermore, we extend the application of eficient estimation techniques based on importance sampling that we had used before only for simple fractal processes. We use importance sampling techniques to eficiently estimate low probabilities of packet losses that occur when a multiplexer is fed with synthetic trafic from our self-similar VBR video model. 1 Introduction An important advantage of packet switched networks (e.g., ATM-based B-ISDN networks), is that such networks support variable bit rate (VBR) connections, thus allowing efi- cient statistical multiplexing of bursty trafic. Video sources (coders) generate inherently VBR trafic, however, in order to transmit video information in circuit-switched networks, the variable content of moving pictures has to be coded in constant bit rate (CBR) form, resulting in ineficient bandwidth utilization and variable picture quality. Due to the advantages of VBR video transmission and the packet-switched nature of ATM, and given the development of highly-sophisticated compression techniques for video sources, VBR compressed video trafic is expected to become one of the main loading components in future BISDN networks. However, the high bandwidth and burstiPresented at ACM SIGCOMM ’95, Cambridge, MA, August, 1995 ness of VBR video trafic, can make network design and management dificult to perform. E ective design and performance analysis depend on accurate modeling of the various trafic types. Among bursty trafic types, VBR video sources are arguably among the most important and demanding to model, due to their bandwidth uctuation and autocorrelation, as well as their complex generation scheme (coding algorithm). Numerous studies have been conducted on issues of video coding, transmission over packet networks, and related modeling and performance analysis topics, see and references within. Traditional models based on Markovian structures (e.g., MMPP, IBP, etc.) have been widely used to statistically approximate VBR video trafic. All these models have in common an asymptotically exponential decay of the autocorrelation function and a rapidly decaying marginal distribution tail. Furthermore they lack a systematic way of simultaneously tting both the empirical marginal distribution and the autocorrelation function.