In modern software systems, configurability has become essential for optimizing performance across varying scenarios and user demands. However, predicting performance in configurable software remains challenging due to the complex interplay between configuration settings and workload characteristics. Existing performance models often lack applicability and ignore the combined influence of configurations and workloads, limiting their applicability in dynamic environments. Additionally, current methods focus on either configurations or workloads in isolation, leaving the interactions between the two insufficiently explored.

This thesis addresses these challenges by proposing a comprehensive framework for performance modeling in configurable software. We conduct an empirical study to understand the complex relationships between configurations, workloads, and performance outcomes. Additionally, we develop a systematic sampling method to enhance the applicability and accuracy of configuration performance models, allowing models to learn from historical data actively. To further improve performance prediction, we propose a hybrid modeling approach that integrates configuration and workload variations, thereby increasing model simplicity and precision. Finally, we explore applying large language models (LLMs) to streamline the modeling process, embedding LLM insights into traditional methods to reduce the cost and complexity of performance modeling. These contributions aim to create robust performance models that better support software configuration and workload management, ultimately enhancing system reliability and efficiency.