IRWIN AND JOAN JACOBS CENTER FOR COMMUNICATION AND INFORMATION TECHNOLOGIES To Cloud or not to Cloud: Optimizing Cloudbursting Costs

The emerging hybrid cloud architectures allow organizations (users) to augment the private infrastructure with practically unlimited public cloud resources in order to cost effectively meet their intermittent peak demands. In such scenarios, users first utilize their already paid private computation infrastructure and offload selected tasks to the public cloud when the private resources become overloaded. Consequently, there is a need to devise efficient on-line task offload algorithms that optimize the overall user cost while maintaining adequate quality of service. Such algorithms should take into account the difference in communication and computation requirements between the different task types. For example, it is clear that between two tasks with the same computational requirements, the task with the lower migration cost is a better candidate to be offloaded to the cloud. In this work, we devise optimal on-line decision algorithms by modeling and solving several associated multi-dimensional Markov Decision Process problems. We address the case in which arriving tasks have multiple communication costs and prove the structural properties of the optimal threshold policy. In addition, we also apply the MDP framework for the complement problem facing cloud providers. If certain cloud resources are not pre-allocated to users, it makes sense for the cloud provider to offer them for opportunistic on-demand usage. We model and provide optimal policies for the buildup of a task backlog by accepting or rejecting tasks that carry user offered price or by dynamically changing advertised prices for tasks. The analytical results are supported by numerical evaluations. We demonstrate the practical advantage of threshold type policies and provide an insight of their dependence on system parameters.