Variability-Aware Performance Modeling: A Statistical Learning Approach

Customizable software systems allow users to derive configurations by selecting features. Building a performance model to understand the tradeoff between performance and feature selection is important to be able to derive a desired configuration. A challenge is to predict performance accurately when features interact. Another is that, in practice, we can often measure only few configurations as a sample for prediction, and we cannot select these configurations freely to cover certain feature interactions. We propose an incremental and variability-aware approach to performance modeling based on statistical learning. Our approach incorporates performance-relevant feature interactions and quantifies their influence implicitly during the process of performance modeling. It identifies the most relevant feature selections automatically for performance prediction. Empirical results on six real-world case studies show that our approach achieves an average of 94% prediction accuracy measuring few randomly selected configurations.