Design and performance modeling of distributed video on demand systems

Abstract

Among video-on-demand (VoD) is considered as one of the potential on-demand multimedia services for profitability in the near future. Huge aggregated bandwidth requirement is known to be a hinge factor to its popularity and hence require properly engineering its architecture to reduce the requirements. Distributed approach is a commonly used to address this issue. This research tries to advance the art of engineering the distributed architecture two perspectives. First, a novel caching scheme called partial video sequence (PVS) caching is proposed and evaluated. In addition to the huge bandwidth requirement, ability to satisfy client with wide varying QoS requirements is another important issue for VoD service while there was lack of prior work. Analytical results show that PVS Caching Scheme is significantly better in reducing the edge server capacity requirement when compared to conventional caching scheme, particularly when the resolution distribution is highly skewed. Second, request level mathematical models for performance estimates are proposed to facilitate network engineers to design and dimension distributed VoD systems. This thesis emphasize on VoD system employing dynamic request routing. While research results are abundant for circuit-switched scenarios, there has not been much study that applies to VoD system, particularly, large VoD system. The development of efficient computational procedures for good approximate solutions is of interest. The Erlang Fixed Point Approximation (EFPA) for the analysis of dynamic routing in circuit switched networks appears to be applicable for the analysis of the VoD system but a direct application leads to unacceptable errors for loss probability due to invalid assumptions of Poisson arrival and server independence. Therefore, the development of VoD system request level mathematical performance models, which are accurate and computationally efficient, is of interest. A novel method called the Overflow Priority Classification is proposed. Numerical results demonstrate that, when compared with the EFPA, the proposed mathematical performance model gives more accurate results for random request routing in a fully connected network with infinite link capacity. For finite link capacity system, a Unified Performance Model is proposed. Numerical results show that comparing to EFPA, the proposed model gives more accurate results.