Runtime Quantitative Verification of Self-Adaptive Systems

Software systems used in mission- and business-critical applications in domains including defence, healthcare, and finance must comply with strict dependability, performance, and other Quality-of-Service (QoS) requirements. Self-adaptive systems achieve this compliance under changing environmental conditions, evolving requirements and system failures by using closed-loop control to modify their behaviour and structure in response to these events. Runtime quantitative verification (RQV) is a mathematically-based approach that implements the closed-loop control of self-adaptive systems. Using runtime observations of a system and its environment, RQV updates stochastic models whose formal analysis underpins the adaptation decisions made within the control loop. The approach can identify and, under certain conditions, predict violation of QoS requirements, and can drive self-adaptation in ways guaranteed to restore or maintain compliance with these requirements. Despite its merits, RQV has significant computation and memory overheads, which restrict its applicability to small systems and to adaptations affecting only the configuration parameters of the system. In this thesis, we introduce RQV variants that improve the efficiency and scalability of the approach and extend its applicability to larger and more complex self-adaptive software systems, and to adaptations that modify the structure of a system. First, we integrate RQV with established efficiency improvement techniques from other software engineering areas. We use caching of recent analysis results, limited lookahead to precompute suitable adaptations for potential future changes, and nearly-optimal reconfiguration to eliminate the need for an exhaustive analysis of the entire reconfiguration space. Second, we introduce an RQV variant that incorporates evolutionary algorithms into the RQV process facilitating the efficient search through large reconfiguration spaces and enabling adaptations that include structural changes. Third, we propose an RQV-driven approach that decentralises the control loops in distributed self-adaptive systems. Finally, we devise an RQV-based methodology for the engineering of trustworthy self-adaptive systems. We evaluate the proposed RQV variants using prototype self-adaptive systems from several application domains, including an embedded system for unmanned underwater vehicles and a foreign exchange service-based system. Our results, subject to the adaptation scenarios used in the evaluation, demonstrate the effectiveness and generality of the new RQV variants.