Study On Access Pattern Of VoD System: Measurement, Analysis, And Modeling

With the continuous improvements in network bandwidth,computing power,and storage capacity,the systems deliverying audio and video with streaming technique increase dramaticly.The design of existing streaming media system and the user behavior generator used to evaluate the performance of media system is based on the measurement,analysis and modeling of user behavior.However,current studies on use access characteristics are comparatively rough and limited.An insightful understanding of user access behavior is essential to guide the design and management of streaming media and the design of user behavior generator.In this Ph.D.dissertation,the measurement,analysis and modeling of user access behavior are carded on the video on demand system of University of Science and technology of China(USTC VOD) in China Education and Research NET (CERNET).We focus on several main characteristics of user interactive behavior,including the session arrival process,the jump distance,play duration and number of interactive requests of all sessions.As current studies on user behavior are inadequate in user interactive behavior,we propose to investigate the distribution of interactive requests in media object.The statistical studies show that the distributions of interactive requests ranks of videos in USTC VOD systems appear to heavy tailed in different time scales.The common power law model may not be suitable to characterize the distribution,while the stretched exponential model well characterized the distribution of number of user interactive requests.The physical meanings of the parameters of stretched exponential model are also given.We then propose the black-box model of users' access to video and conduct an analysis and modeling of video popularity.Three access factors are proposed for characterizing the popularity.One is the common access frequency,one is the ratio between video playback duration and video length,the other is number of jump backwards.The statistical analysis shows that the common Zipf-like model may not be suitable to characterize the popularity distribution,while the stretched exponential model is a better one.The rank correlation analysis also shows that the discrimination of popularity characterized by access frequency could be improved by other two popularity characterizations.Finally,a novel caching strategy,adaptively adjust strategy of streaming media caching length,is proposed based on user access characteristics.Trace-driven simulations demonstrate that the proposed strategy outperforms other caching schemes for interactive requests.