Software reliability measurement and modeling for multiple releases of commercial software

This paper summarizes our experience and findings in measuring and modeling software reliability for a large IBM software product. Four consecutive releases of this product were studied, with various reliability models fitted for individual releases. The models are evaluated by how good they can fit the observed data and how predictive they are. Model sensitivity and performance analysis are performed for across release comparisons, resulting in the models being ranked according to their robustness and accuracy in predicting failures across different data sets. Various ways of combining the multiple release information for generalization and extrapolation to successor releases are also explored.