Metrics of software evolution as effort predictors - a case study

Abstract

Despite its importance, cost estimation in the context of continuing software evolution has been relatively unexplored. This paper addresses this omission by describing some models that predict effort as a function of a suite of metrics of software evolution. It presents a case study relating to the evolution of the kernel of a mainframe operating system. Six models based on eight different indicators of evolution activity are proposed, and their predictive power is examined and compared to that of two baseline models. Predictions with errors of the order of 20% of the actual values have been obtained from the models, when fitted to and tested against historical data over a segment of 10 years of the kernel's continuing evolution. The appropriateness of the proposed models as predictors appears to be restricted to homogeneous evolution segments, i.e. periods with relatively small variations in the level of effort applied. It was found that models based on coarse granularity measures, such as "subsystem counts", provided a mean magnitude of relative error which was similar to those based on finer alternatives, such as "module counts".