Fluid Analysis of Network Content Dissemination and Cloud Systems

Abstract : The project was concerned with optimizing performance in complex network content dissemination and cloud systems. We employed tools of theory, convex optimization and control theory, to study how to disseminate content in a distributed network, managing tradeoffs inefficiency, energy, and resource allocation. Fluid models provide a tractable path to represent these high dimensional problems, retaining accuracy in key performance questions. A first line of work, initiated in our previous AFOSR/SOARD project, concerns peer-to-peer dissemination in wireless ad-hoc networks. We focus on the necessary tradeoff between an efficient use of the network substrate, and the necessary reciprocity between peers, aspects that may be in conflict in the wireless setting. Our results published in use convex optimization to formulate a relevant tradeoff, and propose decentralized algorithms which involve peer-to-peer interactions, and are shown to converge to the corresponding tradeoff point. A second line of contributions referred to the optimization of cache systems, a widespread method of content dissemination. In we address the question of which files to cache and its impact on performance; we worked in the setting of time-to-live (TTL) caching, where the decision involves a choice of timer for each stored content, and its relative popularity must be considered. We formulate a relevant optimization problem, and solve it in cases of practical interest. Numerous insights on practical caching mechanisms result from this mathematical analysis. Extensions of the method to networks of caches were tackled in. A third direction concerned cloud computing and server systems, where processing resources may be adjusted dynamically in real time. The main question is how to control active service capacity, and how to allot it to current jobs, to pursue relevant performance.