Güvenilir içerik dağıtımı için yeni bir ara bellek yönetim modelinin tasarım ve analizi

For supporting reliability in distributed content dissemination services, message loss recovery mechanism achieved via efficient buffer management is an indispensable component. The available approaches for buffer management concentrate on several aspects of the problem such as flow control, reducing the memory usage, providing message stability and the replacement of buffer items. In this thesis study, we consider buffer management problem in support of large-scale bio-inspired peer-to-peer data dissemination services. Bio-inspired epidemic protocols have considerable benefits as they are robust against network failures, scalable and provide probabilistic reliability guarantees. Coupled with an efficient buffering mechanism, system wide buffer usage can be optimized while providing reliability and scalability in such protocols. We propose a novel algorithm, Stepwise Fair-share Buffering, that is shown to provide uniform load distribution in comparison to earlier approaches and reduces the overall buffer usage where every peer has the partial view of the system. A major aim of our approach is to be able to choose bufferers uniformly throughout the system so that the load of buffering will be well balanced among participating peers and the efficiency of content dissemination will be improved as a result. This also reduces the memory usage since only a small subset of the peers is chosen as bufferers for each message. Furthermore, it is applicable to large-scale scenarios, provides reliable delivery and is adaptable to dynamic join and leaves to the system. It adjusts the buffer size to achieve message stability with a high probability. Performance evaluation of the buffering model and extensive comparisons with earlier approaches are performed. The evaluations include scalability, reliability, adaptivity to failures and uniformity analysis. We also derive analytical results for reliability of dissemination as a function of buffer levels. These results are based on a Markov chain analysis and are evaluated numerically. Comparison with simulations shows that they provide a good lower bound for reliability. For high level of reliability values, the bounds are very close to the simulation results.