Squeezing out the Cloud via Profit-Maximizing Resource Allocation Policies

Abstract: We study the problem of maximizing the average hourly profit earned by a Software-as-a-Service (SaaS) provider who runs a software service on behalf of a customer using servers rented from an Infrastructure-as-a-Service (IaaS) provider. The SaaS provider earns a fee per successful transaction and incurs costs proportional to the number of server-hours it uses. A number of resource allocation policies for this or similar problems have been proposed in previous work. However, to the best of our knowledge, these policies have not been comparatively evaluated in a cloud environment. This paper reports on an empirical evaluation of three policies using a replica of Wikipedia deployed on the Amazon EC2 cloud. Experimental results show that a policy based on a solution to an optimization problem derived from the SaaS provider's utility function outperforms well-known heuristics that have been proposed for similar problems. It is also shown that all three policies outperform a "reactive" allocation approach based on Amazon's auto-scaling feature.