A Cost-Effective Resource Provisioning Framework using Online Learning in IaaS Clouds

Cloud vendors such as Amazon EC2 offer two types of purchase options: on-demand and spot instances. An important problem for all users is to determine the way of utilizing different purchase options so as to minimize the cost of processing all incoming jobs while respecting their response-time targets. To be cost-optimal, the process under which users utilize self-owned and cloud instances to process each job (e.g., the order of utilizing them, when to update the allocation of them) is defined in advance by the ways that users are charged, and, to configure the process, we need to determine the optimal amounts of various instances at each allocation update of a job. In this paper, we uncover what parameters are dominating the minimum cost of utilizing various instances and propose (near-)optimal functions of parameters to determine the amounts of different instances allocated to a job. Although these parameters are unknown due to the cloud market dynamics, we can apply the technique of online learning to learn them. Compared with some existing or intuitive policies to utilize self-owned and cloud instances, simulations are done to show a cost reduction by up to 62.85% when spot and on-demand instances are considered and by up to 44.00% when self-owned instances are considered.