Minimizing Cost in IaaS Clouds Via Scheduled Instance Reservation

Regular diurnal patterns are often seen in the workloads of cloud-based online applications. This kind of non-stationary workloads changes the processing demands over time. To run application services with minimum costs, the number of cloud instances can be dynamically adjusted according to the workload variations. Recently, a new type of scheduled instances has emerged in the Infrastructure-as-a-Service market to facilitate such configurations. Scheduled instances can be reserved based on a recurring schedule and they offer price discounts. Meanwhile, cloud vendors require minimum scheduled durations to avoid the overhead of frequently launching and terminating cloud instances. Coupled with traditional on-demand and reserved instances, it becomes more complicated for users to find the optimal combination of these three pricing options to minimize their monetary costs. For the new scheduled instances, not only the number of instances but also their start and stop times have to be decided. In this paper, we develop a fast and effective strategy to solve this problem. Based on the hourly workload distributions, we first compute the optimal number of instances to acquire for each pricing option. Then, we design a scheduling algorithm to arrange the scheduled instances in compliance with the restriction of their scheduled durations. Using the workloads of the LOL online game and the Wikipedia Mobile service as two case studies, the efficacy of our strategy is demonstrated.