| Summary: | 碩士 === 國立交通大學 === 資訊科學與工程研究所 === 96 === The last few years saw the growing popularity of peer-to-peer (P2P) systems that enable collaborative, decentralized sharing of files or other types of resources such as machine cycles, communication bandwidth, storage space, and so on. One important issue of P2P architecture is to ensure fairness – whether all participants can contribute and/or receive their fair shares of resources. While there have been extensive studies on the problem of distributing data in a fair and fully decentralized manner, it remains an open question whether fairness can also be achieved for other types of resources, especially when resource requestors and providers behave highly dynamically and irregularly. In this thesis we propose a more stringent fairness measure than the usual load balancing indicators found in literature. Specifically, the entire P2P network is modeled as a virtual queue where requests for resource consumption and contribution arrive indefinitely. Fairness is judged by the degree of preemption, i.e. the number of times a given request is cut in line by other late-arriving requests. In such a model, existing approaches to load balancing fail to achieve fairness satisfactorily if consumption requests are scheduled based on local information. To address this problem, we first investigate the impact of the proposed fairness model on common P2P networks via simulation, and then propose an alternate scheduling algorithm that routes consumption requests along spanning trees to awaiting providers. The results show reasonable performance in terms of average response time and locality when compared to other decentralized load-balancing algorithms, while keeping the fairness measure low when compared to scheduling via centralized queue
|
|---|
