Performance-Aware Behaviour Models for Feature-Dependent Runtime Attributes in Product Lines

A product line's features describe its configurable functional properties by means of boolean feature toggles or numeric values.Meanwhile, performance models describe the link between features and non-functional product line properties such as binary size, memory usage, or workload latency.So far, these have focused on static features: model formalisms and learning algorithms assume that the product cannot be reconfigured at runtime and that there is no interaction between features, runtime behaviour, and runtime performance attributes.This is inadequate for real-world configurable software systems: compile-time features may be reconfigured at runtime (e.g. by overriding defaults), and non-featured workload attributes may affect performance attributes and thus must be considered within performance models.We propose performanceaware behaviour models to address this challenge.These build upon existing work on behaviour models and dynamic product lines, and link feature models (static configuration), behaviour models (variable workloads), non-featured runtime variability, and performance models in order to allow engineers to determine arbitrary configuration-and workload-dependent performance attributes of product lines.They are compatible with existing modeling methods and support workload-dependent performance attributes that cannot be expressed as a function of product features alone.We demonstrate the benefits of performance-aware behaviour models in three case studies, and show that they reduce performance prediction error by up to 90 %. CCS Concepts Software and its engineering Extra-functional properties; Software verification and validation.