Cloudy with a Spot of Opportunity: Analysis of Spot-Priced VMs for Practical Job Scheduling

Abstract

Public Clouds offer elastic computing resources on- demand using a pay-as-you-go model. While this has opened up access to computing infrastructure, the costs for accessing Cloud resources can be a barrier to adoption in emerging markets. Spot- priced virtual machines (VMs) are offered at deep discounts for the same compute capability as fixed price on-demand VMs. But they can be reclaimed by the Cloud provider at any time, affecting reliability. This paper characterises the behaviour of spot-priced VM from Amazon Web Service for the Asia-Pacific and US East Regions, and analyses their practical impact on running jobs on spot VMs. Our simulation study using jobs of diverse sizes evaluates the trade-offs between cost savings over fixed price VMs and job resilience. Our results show that in most cases, for the workloads studied, we can achieve an effective bottom-line cost savings of 80 using spot VMs, with over 95 reliability.