Revenue maximization with dynamic auctions in IaaS cloud markets

Abstract

Cloud service pricing plays a pivotal role towards the success of cloud computing. Existing pricing schemes, however, either provide no service guarantees (e.g., Spot Instances in Amazon EC2), or use static on-demand pricing in which the price cannot respond quickly to market dynamics (e.g., On-demand Instances in Amazon EC2). To overcome these problems, in this paper we design dynamic auctions where computing instances are periodically auctioned off to accommodate user demands over time. We address the two main challenges of revenue maximization and auction truthfulness. Our design encompasses a capacity allocation scheme, which determines the amount of instances to be auctioned off in each period, as well as the underlying auction mechanisms, based on dynamic payment schemes corresponding to the proposed capacity allocations over time. We show that our design is two-dimensionally truthful, and it is asymptotically optimal when demand is sufficiently high. Furthermore, by identifying certain optimization structures, we substantially reduce the computational complexity of our solution. Extensive simulations show that our design closely tracks market changes, while generating higher revenues than on-demand pricing.