Resource Management Based On User Behavior Analysis In Streaming Service Systems

With the development of communication technology and the Internet, video-on-demand services (Video on Demand, VoD) has got lots of attention for its convenient access to rich content. VoD has boomed recently and become mainstream applications on the Internet. In response to large-scale users, mass data storage and dynamic user interaction request, the logical overlay network such as content distribution network and Peer-to-Peer network is built on top of the physical network to improve system throughput and scalability. The interactive behavior of users and the dynamics of resources in distributed systems are restricting service quality of streaming media services, the user behavior model and resource management strategies has important practical significance to provide high-quality streaming media Services.Our work in this dissertation is based on the National863Project "the development of integration network scope" and the National Science and Technology Support Program "new generation of television service system supporting cross-regional and multi-service-providers". In the background of multimedia overlay network, in order to improve the service performance of video streaming application, we mainly discuss the resource management, the user behavior during playback of streaming media and high-performance media data prefetching strategies. The main contributions are given as follows:(1) QoS guarantee for streaming media services in a distributed environment is a challenging task. On the basis of the traditional content distribution networks, Using of multimedia overlay network is an effective method for the integration and management of distributed systems resources. Through the analysis of various resources, we present a generic QoS model of service composition, as well as an algorithm to the QoS-aware service problem. The proposed algorithm is designed and implemented based on the concepts of learning automation. It has been rigorously tested and evaluated through extensive simulations. Our results show that the learning automation based approach is scalable and can effectively achieve service composition for QoS requirement satisfaction in polynomial time.(2) User interactive behavior is the foundation and key technology of streaming system. Hidden Markov model was used to model the correlation in interactive behavior. The user viewing behavior is modeled as a hidden Markov model. The strategy estimated the hidden Markov model system parameters by utilizing a maximum likelihood estimation method called Baum-Welch algorithm. The posterior probability of the current user browsing state was updated by Bayesian inference, which was based on hidden Markov model and deduced from the posterior probability of the previous browsing state. Finally, the user browsing state was estimated according to the maximum posteriori criterion. Thanks to the use of the rich VCR operations in popular videos, the prediction accuracy rate can reach more than77.5%and the strategy has obvious modeling capabilities.(3) In this paper, we derive the expectation of the start delays during the visit as well as network jitter delay during playback and propose and study the caching policy of proxy cache. The user experience with the traditional caching algorithms is bad if most of the clients’operations are randomized. We define the initial access probability of media data and the demanding transfer probability between two segments. According to the current model parameters and the user state, we judge the segment which may be accessed the next time and prefetch it to reduce the user access delay. Experimental results show that the algorithm using prefetching techniques can fully utilize the bandwidth of the system and can effectively reduce the probability of access latency, especially in the case of popular video access. Compared with the prefetching algorithms using only simple statistics and LRU cache strategies, the access delay was reduced by6%and20%respectively.