Reverse Engineering of Parametric Behavioural Service Performance Models from Black-Box Components

Integrating heterogeneous software systems becomes increasingly important. It requires combining existing components to form new applications. Such new applications are required to satisfy non-functional properties, such as performance. Design-time performance prediction of new applications built from existing components helps to compare design decisions before actually implementing them to the full, avoiding costly prototype and glue code creation. But design-time performance prediction requires understanding and modeling of data flow and control flow accross component boundaries, which is not given for most black-box components. If, for example one component processes and forwards files to other components, this effect should be an explicit model parameter to correctly capture its performance impact. This impact should also be parameterised over data, but no reverse engineering approach exists to recover such dependencies. In this paper, we present an approach that allows reverse engineering of such behavioural models, which is applicable for blackbox components. By runtime monitoring and application of genetic programming, we recover functional dependencies in code, which then are expressed as parameterisation in the output model. We successfully validated our approach in a case study on a file sharing application, showing that all dependencies could correctly be reverse engineered from black-box components.