Network analysis of large scale object oriented software systems

Abstract

The evolution of software engineering knowledge, technology, tools, and practices has seen progressive adoption of new design paradigms. Currently, the predominant design paradigm is object oriented design. Despite the advocated and demonstrated benefits of object oriented design, there are known limitations of static software analysis techniques for object oriented systems, and there are many current and legacy object oriented software systems that are difficult to maintain using the existing reverse engineering techniques and tools. Consequently, there is renewed interest in dynamic analysis of object oriented systems, and the emergence of large and highly interconnected systems has fuelled research into the development of new scalable techniques and tools to aid program comprehension and software testing. In dynamic analysis, a key research problem is efficient interpretation and analysis of large volumes of precise program execution data to facilitate efficient handling of software engineering tasks. Some of the techniques, employed to improve the efficiency of analysis, are inspired by empirical approaches developed in other fields of science and engineering that face comparable data analysis challenges. This research is focused on application of empirical network analysis measures to dynamic analysis data of object oriented software. The premise of this research is that the methods that contribute significantly to the object collaboration network's structural integrity are also important for delivery of the software system’s function. This thesis makes two key contributions. First, a definition is proposed for the concept of the functional importance of methods of object oriented software. Second, the thesis proposes and validates a conceptual link between object collaboration networks and the properties of a network model with power law connectivity distribution. Results from empirical software engineering experiments on JHotdraw and Google Chrome are presented. The results indicate that five considered standard centrality based network measures can be used to predict functionally important methods with a significant level of accuracy. The search for functional importance of software elements is an essential starting point for program comprehension and software testing activities. The proposed definition and application of network analysis has the potential to improve the efficiency of post release phase software engineering activities by facilitating rapid identification of potentially functionally important methods in object oriented software. These results, with some refinement, could be used to perform change impact prediction and a host of other potentially beneficial applications to improve software engineering techniques.