CloudProphet : Predicting Web Application Performance in the Cloud Duke University Technical Report CS-2011-11

As public cloud computing services are gaining popularity, many are considering migrating their applications from on-premise to cloud. However, due to diverse cloud performance, choosing the cloud platform that is the best suited to migrate is a difficult unsolved problem. In this work, we present CloudProphet, a low cost tool to accurately predict the end-to-end response time of an on-premise web application if migrated to a cloud. CloudProphet collects a resource usage trace of the application running on-premise, and then replays it in cloud to predict performance. This approach does not require actual migration, and is agnostic to both the application and the cloud platform. We address two key design challenges in CloudProphet, including how to trace with both high fidelity and low overhead, and how to extract and enforce application dependency. Our evaluation shows that CloudProphet can achieve high prediction accuracy for two real applications on a variety of cloud instances from AWS, Rackspace, and Storm. We further show CloudProphet has low tracing overhead, and is much lighter than migrating the real applications.