Stochastic modeling of individual resource consumption during the programming phase of software development

Reviewer: Bill Curtis

Considerable effort has been invested in developing resource consumption models of large software projects for use in software cost models. However, there has been correspondingly little effort to model resource consumption at the program- mer level. McNicholl and Magel make an initial contribution in this area by presenting distributions of time to complete the implementation of a single module. The behavior they are describing is carefully constrained to coding and module testing, and thus does not encompass the requirements, design, or software integration activity involved in actual projects. However, by constraining the process studied, they are able to statistically characterize its behavior. This paper is important reading for anyone interested in software measurement, but it is not for the statistical faint of heart. McNicholl and Magel's study is important for three reasons. First, their statistical rigor is refreshing. They demonstrated that the log-normal distribution provided the best fit to their data compared to those of five other distributions. They cleverly created a “completion intensity function” to handle the failure of some programmers to complete a program that ran correctly. Second, they discussed conceptual reasons why the characteristics of the log-normal distribution were more appropriate than those of other distributions for describing coding behavior. Third, with further research this approach can be scaled up to provide more accurate characterizations of the entire programming process. Work in software metrics suggests that measurable characteristics of the program design could be used in estimating the task-related parameters of time-to-completion distributions. The authors present their approach as an alternative to the deterministic model of software development which tries to predict completion time from characteristics of the individual, the environment, and the programming task. However, it is from some of these factors that estimates of important parameters of time-to-completion distributions must be estimated if such distributions are to be used effectively in estimating resource consumption on actual projects. Although many individual difference models have not produced strong results in the past, the problem has been in the simple-minded variables many investigators chose to measure. Better research is emerging on programmer knowledge structures, and better individual difference measures should follow. For instance, the problem solving process of experienced professional programmers has been shown to be qualitatively different from that of novices (the authors' data described third semester students). Thus, if such deterministic measures can be used to estimate parameters of time-to-completion distributions, McNicholl and Magel's approach can be made an even more powerful tool for estimating project resources and completion schedules.