Performance Models of Statistical Multiplexing in GUNNAR KARLSSON.

Abstract-Statistical or asynchronous time division multiplexing of variable bit rate moving image sources is a first step toward packetized video networks. Its goal is to efficiently utilize a common communica-tlons channel while maintaining uniform picture quality at the receiver. In this paper, we present queueing models of such schemes, and use them to assess their probabilistic behavior. First, we analyze the coding hit rate of a single video source as a function of time based on experimental data of a video telephone scene. VBrious statistical quantities are derived from bit rate data collected with a conditional replenishment interframe coding scheme. Two correlated Markov process models (one in discrete t h e and one in continuous time) are shown to fit the experimental data and are used to model the input rates of several independent sources into a statistical multiplexer. The continuous-time model is used with a flow-equivalent approximate queueing analysis to obtain the common buffer length distribution. The analysis is validated with computer simulations that use the dwreie-time source model and take into account the discrete nature of the packet qneue. Numerical results are presented for vadable channel utilization and number of multiplexed video sources. Our analyses demonstrate the statistical smoothing effect of packetized video in a single link. The multiplexer performance, in terms of the probability of buffering or delaying beyond a certain Ilmit, is shown to improve dramatically as the number of multiplexed sources incream beyond one. D I