Autonomous Provisioning of Preemptive Instances in Google Cloud for Maximum Performance Per Dollar

Abstract

Cloud computing and its popularity has boomed over the last decade, enabling anyone to rent computing power on demand. Cloud providers such as Amazon and Google rent out surplus computing power for a discounted price according to demand in their data centers, but with the trade off that it is revocable and can only be rented for a short amount of time.This paper investigates the use of surplus computing power in order to reduce the cost of batch computing. We rely on a simple economical principle, the most cost-efficient Virtual Machine (VM) in a public cloud is the one that offers the highest performance per dollar. Therefore by rescheduling the workloads to the most cost-efficient location in terms of performance per dollar our solution dynamically provisions preemptible VMs in Google Cloud while continuously monitoring the performance per dollar of all available resources in every region. The algorithm automatically relocates workloads to a less expensive location if any appears and handles revoked access of the resources. Our algorithm views the cost reduction problem as a linear optimization problem with constraints and solves it using a greedy procedure. In the experiment we spawn Docker containers to mine cryptocurrency. The experimental results show that 67% of the cost is saved compared to renting on-demand VMs. The system can readily be extended to containers processing similar types of workloads and more generally to applications where the performance per dollar is easy to measure.