Storage Resource Management Based On Delamination In Distributed Service Systems

Multimedia service becomes the core service on the Internet, as the development of communication technology and computer science. Content delivery networks (CDNs) and peer-to-peer (P2P) networks are widely used in video-on-demand (VoD) systems, for boosting quality of service, expanding throughput and lowering running cost. Al-though these distributed systems, such as CDN and P2P, are extensible and strong, the unwound nodes in the networks require appropriate policies to manage and optimize systems resource. Random VCR operations of clients and strict performance require-ments of streaming media service in VoD systems make it become more complicated. In this dissertation we study storage resource management of distributed VoD systems for efficient caching performance. With the expansion of clients number, storage manage-ment becomes the key problem of the services on the Internet. The caching policies we proposed are useful for the design of distributed systems, like VoD systems and cloud computing.In the dissertation, the storage resource in VoD systems is delaminated into three layers, proxy cache, server random access memory cache and client cache. We discuss the caching algorithms with the characters of each layer respectively.Firstly, a windows-based popularity estimate algorithm is proposed in this thesis. Popularity is a chief parameter in the research of cache management for distributed VoD systems and it impacts the performance of caching algorithms directly. However, the traditional popularity estimate algorithms based on a fixed calculating period are not good at tracking clients’behavior. We study a flexible popularity estimate algorithm which can track the evolution in time. The parameter is calculated by estimating request rate of the corresponding segment.Secondly, we derive the expectation of the demanding delay and study the caching policy of proxy cache. The transmission delay between arbitrary two proxy nodes is always long, so there is no good experience with the traditional caching algorithms if most of the clients’operations are randomized. We define the demanding transfer probability between two segments, and study start delay and jitter delay with the method of prefetching. A near-optimal policy of proxy cache management is given, and the solution can approximate to the optimal one by optimizing online. The prefetching algorithm is given under the condition that the cache state is known. The algorithm can reduce the delay of demands from clients with cooperation of caching and prefetching, and the efficiency of cache is improved.Thirdly, RAM caching algorithm of streaming media server is discussed for reduc-ing disk I/O, and a partly observable Markov decision process (POMDP) model of the problem is constructed. In contrast to other RAM caching algorithms, popularity can be taken advantage of to predict clients’behavior and adjust the caching state. We define the average disk I/O, and give a greedy solution. Additionally, the optimal stochastic policy based on the POMDP model is presented. In the model, the disk I/O, as the long-run average performance measure of a Markov chain, is optimized via applying the policy-gradient algorithm.Lastly, a data replace algorithm based on quadratic programming is designed for client nodes which are unstable in the network. For improving the efficiency of client cache, we define the cache distribution factor, analyze the parameters affecting caching performance, and derive caching weight of the segments. The available resource distri-bution factor is also defined for selecting source nodes and allocating requiring band-width. With optimizing the two factors, the proportion of available resource can be almost in accordance with the proportion of clients’requests, and the client cache can perform better.