An Online Cost Optimization Algorithm for IaaS Instance Releasing in Cloud Environments

Abstract

Benefiting from the development of cloud computing, service trades among IaaS (Infrastructure-as-a-Service) providers, SaaS (Software-as-a-Service) providers, and users are more and more common, where SaaS providers purchase on-demand instances in a pay-per-use way from IaaS providers to execute users' jobs. Given the pay-as-you-go mode of IaaS instances, SaaS providers can acquire and release instances whenever they need. However, considering the preparation time of acquiring new instances and the penalty functions of users, from the perspective of cost optimization, SaaS providers need to decide whether and when to release an idle instance. To make optimal decisions, future information about job arrivals and execution time is generally required, which is actually difficult to accurately predict. To address this problem, we propose an online cost optimization algorithm for SaaS providers to help them make better decisions on whether to release instances. Theoretical analysis shows that our online algorithm achieves a competitive ratio of $2-\alpha$ for different penalty functions, where $\alpha\in(0,1)$ . Extensive experiments on account of realistic Google cluster data demonstrate the effectiveness and efficiency of our online algorithm.