An Online Auction for Deadline-Aware Dynamic Cloud Resource Provisioning

Auction mechanisms have recently been studied as an efficient approach for dynamic resource allocation in a cloud market. Existing mechanisms are mostly limited to the offline setting or execute jobs in continuous time slots. This work focuses on a practical case of online auction design, where users bid for future cloud resources for executing their batch processing jobs with hard deadline constraints. We design an online primal-dual auction framework for Virtual Machine (VM) allocation with social welfare maximization, which is truthful, computationally efficient, and guarantees a small competitive ratio. We leverage the framework of post price auctions to design our online primal-dual algorithm, where a bid is accepted if its expected execution cost in future time slots is smaller than its bidding price. We interpret the dual variables as marginal prices per unit of resource, and iteratively update it according to the allocated amount of resource. Theoretical analysis and trace-driven simulation studies validate the efficacy of the online auction framework, including both its computational efficiency and economic efficiency.