Traffic modeling and bandwidth estimation for variable bit-rate (VBR) video transmission

Abstract

The provision of constant-quality variable bit-rate (VBR) video communications remains to be a very challenging problem for network designers. The main theme of this dissertation is to study the transportation of VBR MPEG-encoded video over high-speed networks. In particular, we focus on the aspects of bandwidth allocation and modeling of video traffic. VBR MPEG video traffic exhibits bit-rate variations at both the frame level and the group-of-pictures (GOP) level. When several VBR MPEG streams are multiplexed at the frame level, these two levels of variations give rise to some uncertainties in terms of aggregate bandwidth requirement. Based on Zhang's three-state five-parameter model [Zhang97b], we developed a two-level modeling approach to accurately capture the two-level bit-rate characteristics exhibited by VBR MPEG video. The term level of service robustness is also redefined by means of a simple sliding window approach. We observed from simulations that our model outperformed Zhang's one in terms of scalability. Accurate traffic modeling is very crucial for effective bandwidth allocation. Traffic modeling, or characterization, can be mainly divided into two main streams: stochastic and deterministic. For stochastic characterization, we investigated the possibility of modeling VBR video using normal, gamma and lognonnal mixtures. We employed the M/G/∞ model to generate synthetic traces and the simulation results suggested that a two-component lognonnal mixture can accurately model the long-tailed behavior of VBR video. Beside the above stochastic model, we also investigated the modeling of VBR MPEG video using deterministic characterization. In this direction, we proposed a characterization method called z-characterization which separately characterizes the I-, P-and B-frame traces to prevent the "I-frame domination" effect. Furthermore, the periodic structure of MPEG video is exploited to set reference points for extrapolation purpose, which greatly reduces the computation time. Our simulations showed that the z-characterization yields a very close zigzag bound to the empirical envelope. Finally, we proposed a video transmission service based on the concepts of deterministic characterization and bandwidth renegotiation. In this system, a video stream is characterized at both the micro-segment and segment levels and all bandwidth renegotiations are taken place at micro-segment boundaries. When allocating bandwidth for a micro-segment, service robustness is provided by a parameter called degree of aggressiveness (DoA), and also the short-term and long-term deterministic characterizations form the lower and upper bounds to the resulting bandwidth respectively. We successfully demonstrated in the simulations how different levels of service robustness could be provided to the incoming streams, while ensuring the quality of existing connections not being compromised by new connections with high-DoA requests.