Research On Optimization Of Hierarchical Cache In Video-On-Demand System

As networks become more and more ubiquitous, Internet plays a critical role in the modern economy, the society and the lives of billions of people. Recently, with the fast development of Web2.0 technology, text-based web contents are gradually replaced by more informative multimedia contents including images, sounds and words. As the most typical multimedia contents, online videos are popular all over the world. In order to provide massive network users the needed online videos, Video-On-Demand (VOD) systems emerge. As the size of a video file is typically much larger than that of a text file, the massive demands for online videos will place heavy burdens on the data transmission of Internet, leading to serious network jam and packet dropouts. The performance of Internet could be significantly degraded by the transmission of videos and the user experience of VOD systems may be seriously destroyed.In order to mitigate the data transmission pressure of the Internet and improve the service quality of VOD systems, this thesis focuses on the optimization of the hierarchical caching performance of VOD systems through improving the utilization and reducing the deploying cost of the cache system. The hierarchical cache of VOD systems usually consists of massive cache nodes which may be deployed with heterogeneous storage devices. Such heterogeneous storage devices may significantly increase the difficulty of optimizing the caching performance. Moreover, the time-varying video popularity of VOD systems poses great challenges on the cache optimization. In order to resolve these issues in the hierarchical cache of VOD systems, this thesis intensively investigates the prediction of video popularity, data deployment of hybrid storage devices and cooperation between multiple cache nodes. The main work and contributions of this thesis can be summarized as follows:1) We propose a novel popularity prediction algorithm based on videos’ exponentially weighted history access information. As the interest of users of VOD systems is always changing, videos could show time-varying popularity. Meanwhile, the explosive growth of online videos will yield more time-varying popularity. It is difficult to provide an accurate prediction of video popularity timely based on traditional algorithms. In this thesis, we propose a novel popularity prediction algorithm based on exponentially weighted history access information. The history access information of a video is dynamically weighted according to its time stamp. More specifically, the elder the history access information is, the smaller weight it will get. The random noise in user behavior can be effectively suppressed by the employed history access information, and the latest video popularity can be tracked through reducing the weights of old history information gradually. Our proposed algorithm owns low calculation complexity. With our algorithm, we can quickly and precisely predict the popularity of videos.2) We propose a novel data deployment algorithm of hybrid storage devices based on the write load feedback. The hardware performance of storage devices has critical influence on the read speed and the concurrent capability of cache nodes. Traditional hard disk devices have large capacity and low cost, but poor read performance. New solid state devices have high read speed, but very limited write endurance and high costs prevent them of completely replacing hard disk devices. Hybrid storage devices that integrate both hard disk devices and solid state devices are able to make full use of the advantages of the two complementary types of disks. In this thesis, we propose a novel data deployment algorithm of hybrid storage devices based on the write load feedback. The data deployment of hybrid storage devices is modelled as a data migration problem between different devices, and the write load of the solid state devices serves as a feed-back of the migration algorithm to adjust the migration threshold adaptively. With the data migration and the threshold adjustment, the optimization of the hybrid storage devices under limited write endurance of the solid state devices can be achieved in a simple way.3) We propose a novel distributed cooperation algorithm between multiple nodes in the hierarchical cache system based on the storage benefit defined for each video. In the hierarchical cache system, the service areas of cache nodes in the same layer are next to each other, while the service areas of cache nodes in different layers overlap. Thus, different cache nodes may observe similar video popularity. So cache nodes need to cooperate with each other to reduce caching redundancy. To optimize the cooperation of the hierarchal cache system is a complicated global optimization problem. We first propose the concept of storage benefit of a video in a cache node, and then transform the global optimization problem of the cooperation into a scheduling problem of the storage benefit in each cache node and a data migration problem between different layers of the hierarchical cache system. The complexity of the cooperation is highly reduced and the utilization efficiency of the cache system is greatly improved.Simulation results of our proposed algorithms show that our proposed popularity prediction algorithm has higher byte-hit-ratio than traditional algorithms, our proposed data deployment algorithm of hybrid storage devices meets the strict requirement of the hybrid storage devices’ write life, and our proposed cooperation algorithm provides higher system performance than traditional algorithms.