Toward automatically quantifying the impact of a change in systems

Abstract

Software maintenance is becoming more challenging with the increased complexity of the software and the frequently applied changes. Performing impact analysis before the actual implementation of a change is a crucial task during system maintenance. While many tools and techniques are available to measure the impact of a change at the code level, only a few research work is done to measure the impact of a change at an earlier stage in the development process. Measuring the impact of a change at the model level speeds up the maintenance process allowing early discovery of critical components of the system before applying the actual change at the code level. In this paper, we present model-based impact analysis approach for state-based systems such as telecommunication or embedded systems. The proposed approach uses model dependencies to automatically measure the expected impact for a requested change instead of relying on the expertise of system maintainers, and it generates two impact sets representing the lower bound and the upper bound of the impact. Although it can be extended to other behavioral models, the presented approach mainly addresses extended finite-state machine (EFSM) models. An empirical study is conducted on six EFSM models to investigate the usefulness of the proposed approach. The results show that on average the size of the impact after a single modification (a change in a one EFSM transition) ranges between 14 and 38 % of the total size of the model. For a modification involving multiple transitions, the average size of the impact ranges between 30 and 64 % of the total size of the model. Additionally, we investigated the relationships (correlation) between the structure of the EFSM model, and the size of the impact sets. Upon preliminary analysis of the correlation, the concepts of model density and data density were defined, and it was found that they could be the major factors influencing the sizes of impact sets for models. As a result, these factors can be used to determine the types of models for which the proposed approach is the most appropriate.