Towards Performance Prediction for Public Infrastructure Clouds: An EC2 Case Study

Abstract

The increasing number of Public Clouds, the large and varied range of VMs they offer, and the provider specific terminology used for describing performance characteristics, makes price/performance comparisons difficult. Large performance variation can lead to Clouds being described as 'unreliable' and 'unpredictable'. The aim of this paper is to offer a basis for making probability-based performance predictions in Public (Infrastructure) Clouds, with Amazon's EC2 as our focus. We demonstrate how CPU model determines instance performance, show associations between instance classes and sets of CPU models, and determine class-to-model performance characteristics. We suggest that by knowing the proportion of CPU models backing specific instances, and in absence of provider knowledge or ability to specify model or performance, we can estimate the likelihood of a user obtaining particular models in respect to a request, and that this can be used to gauge likely price/performance.