Software complexity: Integrating the task complexity model and cognition.

摘要

Knowledge work is typically thought of as complex, and few types of knowledge work are as complex as software engineering (SE). A large number of measures have been developed to analyze software and its concomitant processes with the goals of evaluating, predicting and controlling its complexity. The first research question for this dissertation is how to determine a parsimonious subset of software measures to use in software complexity analysis? To answer this question, task complexity is studied. The result is that coupling and cohesion stand out as comprising one such subset.;The second research question revolves around how to resolve potential conflicts between lowering coupling and increasing cohesion? Consideration of the information processing view of cognition allows this dissertation to address the second question. In doing so, cohesion is proposed as a moderator on the main relationship between coupling and complexity.;The proposed model is tested with a controlled experimental design using commercial software engineers wherein coupling and cohesion are manipulated across two software maintenance tasks.;This research makes several contributions. The theory-driven approach considers both the task complexity model and cognition and lends significant support to the developed model for complexity. Empirically, the model undergoes validation using a controlled experimental design that simultaneously manipulates coupling and cohesion. Furthermore, examination of the task complexity model steers this dissertation towards considering complexity in the holistic sense of an entire program, rather than a single program unit, as is conventionally done. Finally, it is intended that by focusing software measurement on coupling and cohesion, research can more fruitfully aid both the practice and pedagogy of complexity management in SE.;In terms of results, the interaction of coupling and cohesion was determined to have a significant effect at an alpha level of 0.10 (and demonstrating medium to large effect sizes) on the length of time it took subjects to perform the software maintenance, as predicted by the model.