We address the problem of dynamic resource placement in general networking and cloud computing applications. We consider a large-scale system faced by time varying and regionally distributed demands for various resources. The system operator aims at placing the resources across regions to maximize revenues, and thus needs to address the problem of how to dynamically reposition the resources in reaction to the time varying demand. The challenge posed by this setting is to deal with arbitrary multi-dimensional stochastic demands which vary over time. Under such settings one should provide a tradeoff between optimizing the resource placement as to meet its demand, and minimizing the number of added and removed resources to the placement. Our analysis and simulations reveal that optimizing the resource placement may inflict huge resource repositioning costs, even if the demand has small fluctuations. We therefore propose an algorithmic framework that overcomes this difficulty and yields very efficient dynamic placements with bounded repositioning costs. Our solution is developed under a very wide cost model, and thus allows accommodation of many systems. Our solutions are based on new analytic techniques utilizing graph theory methodologies that can be applied to other optimization/combinatorial problems.
- Graph algorithms