Online multi-instance acquisition for cost optimization in IaaS Clouds

Abstract

Amazon Ec2 service offers two diverse instance purchasing options. Users can either run instances by using on-demand plan and pay only for the incurred instance-hours, or by renting instances for a long period, while taking advantage of significant reductions (up to 60%). One of the major problems facing these users is cost management. How to dynamically combine between these two options, to serve sporadic workload, without knowledge of future demands? Many strategies in the literature, require either using exact historic workload as a reference or relying on long-term predictions of future workload. Unlike existing works we propose two practical online deterministic algorithms for the multi-slope case, that incur no more than 1+1/1−α and 2/1−α respectively, compared to the cost obtained from an optimal offline algorithm, where α is the maximum saving ratio of a reserved instance offer over on-demand plan.