Interpretation of Behaviour Models at Runtime: Performance Benchmark and Case Studies

Modelling system behaviour by means of statechart-based formalisms, such as the state machine formalism defined in the Unified Modeling Language, is an established practice in software engineering. As part of a model-driven workflow, engineers usually employ a code generation approach to create software components that implement an intended runtime behaviour. Although this approach yields software components with a good runtime performance, the resulting system behaviour is static. Changes to the behaviour model necessarily provoke an iteration in the code generation workflow and a re-deployment of the generated artefacts. In the area of autonomic systems engineering, it is assumed that systems are able to adapt their runtime behaviour in response to a changing context. Thus, the constraints imposed by a code generation approach make runtime adaptation difficult, if not impossible. We investigate a solution to this problem by employing interpretation techniques for the runtime execution of behaviour models, enabling the adaptability of a system’s runtime behaviour on the level of single model elements. This is done by devising concepts for behaviour model interpretation, which are then used in proof-of-concept implementations to demonstrate the feasibility of the approach. It is insufficient to show only the general feasibility of behaviour model interpretation, as the usefulness of this approach depends on the context the model is used in, which is determined by a specific application domain. Therefore, an assessment of the approach is carried out, using quantitative as well as qualitative methods. For the quantitative evaluation, a novel benchmark is introduced, that enables a performance comparison between the proof-of-concept implementations and generated code. The qualitative assessment is based on use case studies conducted in the area of network and systems management. Behaviour model interpretation has a performance overhead when compared to code generation. In the context of the network and systems management domain, the performance of the approach is found to be adequate for the vast majority of situations, except when dealing with high-throughput or delay-sensitive data.