On Resource Management for Cloud Users: A Generalized Kelly Mechanism Approach

Cloud computing provides network software companies with a platform to develop Software as a Service (SaaS) applications. The pay-as-you-go pricing model unties these SaaS providers from large capital outlays in hardware deployment and maintenance. Depending on the level of virtualization and service abstraction, cloud vendors provide different functionalities to the SaaS providers. For example, with a high-level service abstraction, Google AppEngine can provide automatic scalability for web applications; however, it cannot support general-purpose applications and does not give service controls for the SaaS providers. On the other hand, Amazon EC2 provides a low-level virtualization that allows for general-purpose applications. Nevertheless, it cannot scale the applications automatically for SaaS providers. Moreover, managing the computing resources for SaaS processes can also be challenging for a SaaS provider. The aim of this paper is to provide a theoretical framework for resource management for Saas providers so they can efficiently control the service levels of their users, and as to easily scale their applications under dynamic user arrivals/departures. Our resource bidding and allocation framework can be viewed as a generalization of the Kelly mechanism. Previous work showed that the efficiency loss of the Kelly mechanism is bounded by 25% of the optimal social welfare. By using a built-in penalty differentiation, our mechanism is able to close the efficiency gap. To achieve this efficiency, a feedback control mechanism is proposed to maximize the aggregate valuation of users.