Phase Aware Performance Modeling for Cloud Applications

In this paper we propose a new methodology for performance modeling of applications deployed in the cloud based on automatically discovered phases along with their inputs. Our method is based on lightweight sampling that can predict the performance of applications with up to 95% accuracy for previously unseen input configurations at less than 5% overhead. We show the effectiveness of the performance modeling methodology in case of anomaly detection for a variety of real world workloads. As compared to the state-of-the-art, our method gives significant improvements in reducing both false positives and false negatives for anomalous test cases.