Resource provisioning of web applications in heterogeneous clouds

Abstract

Cloud computing platforms provide very little guarantees regarding the performance of seemingly identical virtual machine instances. Such instances have been shown to exhibit significantly different performance from each other. This heterogeneity creates two challenges when hosting multi-tier Web applications in the Cloud. First, different machine instances have different processing capacity so balancing equal amounts of load to different instances leads to poor performance. Second, when an application must be reprovisioned, depending on the performance characteristics of the new machine instance it may be more beneficial to add the instance to one tier or another. This paper shows how we can efficiently benchmark the individual performance profile of each individual virtual machine instance when we obtain it from the Cloud. These performance profiles allow us to balance the request load more efficiently than standard load balancers, leading to better performance at lower costs. The performance profiles also allow us to predict the performance that the overall application would have if the new machine instance would be added to any of the application tiers, and therefore to decide how to make best use of newly acquired machine instances. We demonstrate the effectiveness of our techniques by provisioning the TPC-W e-commerce benchmark in the Amazon EC2 platform. 1