Engineering aspects of an architecture for non-real-time VBR video services

Abstract

In this paper, we analyze the statistical multiplexing efficiencies and related engineering implications of a candidate architecture for providing non-real-time VBR video services, such as pay-per-view and video-on-demand. In particular, we employ a closed queueing network model, driven by realistic VBR video traffic, to estimate the probability of starvation within the playout buffers used in the architecture. While actual VBR video traffic has previously been shown to exhibit long-range dependence, its impact on the perceived Quality of Service (QoS) has recently been a subject of wide debate among researchers. The results reported in this paper provide one example of a generic video system where long-range dependence does have a qualitative impact on performance and engineering. More precisely, using our closed queueing network model and a well-studied trace of actual VBR video traffic, we demonstrate that (i) the think times for playout are directly influenced by the presence of long-term correlations in the traffic, thereby affecting the probability of starvation in the playout buffers, and (ii) statistical multiplexing gains in the range of 12 are feasible, when the peak rate of the traffic is defined in terms of the largest frame in the test trace. There is considerable scope for further work in this area to better quantify the effects of long-range dependence on the engineering of non-real-time VBR video services.