Dynamic selection of virtual machines for application servers in cloud environments

Abstract

Autoscaling is a hallmark of cloud computing as it allows flexible just-in-time allocation and release of computational resources in response to dynamic and often unpredictableworkloads. This is especially important forweb applications, whose workload is time dependent and prone to flash crowds. Most of them follow the 3-tier architectural pattern, and are divided into presentation, application/domain and data layers. In this work, we focus on the application layer. Reactive autoscaling policies of the type "Instantiate a new Virtual Machine (VM) when the average server CPU utilisation reaches X%" have been used successfully since the dawn of cloud computing. But which VM type is the most suitable for the specific application at the moment remains an open question. In this work, we propose an approach for dynamic VM type selection. It uses a combination of online machine learning techniques, works in real time and adapts to changes in the users' workload patterns, application changes as well as middleware upgrades and reconfigurations. We have developed a prototype, which we tested with the CloudStone benchmark deployed on AWS EC2. Results show that our method quickly adapts to workload changes and reduces the total cost compared to the industry standard approach.