Measuring Performance Variability in the Clouds

Abstract

ABSTRACTCloud computing is attractive because of its ubiquity and economy of scales. However, there is a wide variability of performance among servers in a public cloud due to virtualization and multiple users sharing the same physical server. Any excessive resource usage by a user will impact other users on the same serve. Performance variability of production cloud services has been reported elsewhere, however, we document hardware resource level performance variations in public clouds. The focus of our study is on the computational capability of a server in a public cloud, and its overall loading level, instead of an end-user application's performance. We will share a study conducted over six different public clouds, with 24 users over a period of three months showing results that present flavours and extent of randomness in the public clouds. Our focus is not limited to any selected workload or user application type, but on the server hardware being deployed and its loading levels by various cloud service providers.