Federated Private Clouds via Brokers Marketplace: A Stackelberg-Game Perspective,” http://www.cs.hku.hk/~xjqiu/federation-broker.pdf, The University of Hong Kong,

Abstract-More and more enterprises have set up their own private clouds by applying virtualization to their data centers; the benefit is flexible resource supply to different internal demands. Aiming to meet the peak demand in their resource provisioning, private clouds are often under-utilized. A new paradigm has emerged that advocates leasing the spare resources to external users, when and if adequate rental prices are offered. A broker is typically employed which pools the spare resources of multiple private clouds together and leases them to serve external users' jobs. Good mechanisms have yet to be derived for the broker to set the offered prices to buy spare resources from the private clouds, and to schedule jobs on the available resources, such that the economic benefits of both the broker and the private clouds are maximized. The design of the mechanism is especially challenging when we consider the dynamic arrival of users' jobs and volatile availability of spare resources at the private clouds, while aiming at long-term profit optimality. In this paper, we model the interaction between the broker and the private clouds as a two-stage Stackelberg game. As the leader in the game, the broker decides and offers prices for renting VMs of different types from each private cloud. As a follower, each private cloud responds with the number of VMs of each type that it is willing to lease. By combining with the Stackelberg game model we design online algorithms for the broker to set the prices and schedule jobs on the private clouds, and for the private cloud to decide the numbers of VMs to lease, based on the Lyapunov optimization theory. We prove that the broker achieves a time-averaged profit that is close to the offline optimum with complete information on future job arrivals and resource availability, while each private cloud makes their best earning. The proposed online algorithm is carefully evaluated based on usage traces of Google cluster and Amazon EC2