A Bayesian model for predicting reliability of software systems at the architectural level
QoSA'07 Proceedings of the Quality of software architectures 3rd international conference on Software architectures, components, and applications
| Hi-index | 0.00 |
Modeling and estimating software reliability during testing is useful in quantifying the quality of the software systems. However, such measurements applied late in the development process leave too little to be done to improve the quality and dependability of the software system in a cost-effective way. Reliability, an important dependability attribute, is defined as the probability that the system performs its intended functionality under specified design limits. We argue that reliability models must be built to predict the system reliability throughout the development process, and specifically when exact context and execution profile of the system is unknown, or when the implementation artifacts are unavailable. In the context of software architectures, various techniques for modeling software systems and specifying their functionality have been developed. These techniques enable extensive analysis of the specification, but typically lack quantification. Additionally, their relation to dependability attributes of the modeled software system is unknown. In this dissertation, we present a software architecture-based approach to predicting reliability. The approach is applicable to early stages of development when the implementation artifacts are not yet available, and exact execution profile is unknown. The approach is two fold: first, the reliability of individual components is predicted via a stochastic reliability model built using software architectural artifacts. The uncertainty associated with the execution profile is modeled using Hidden Markov Models, which enable probabilistic modeling with unknown parameters. The overall system reliability is obtained compositionally as a function of the reliability of its constituent components, and their complex interactions. The interactions form a causal network that models how reliability at a specific time in a system's execution is affected by the reliability at previous time steps. We evaluate our software architecture-based reliability modeling approach to demonstrate that reliability prediction of software systems architectures early during the development life-cycle is both possible and meaningful. The coverage of our architectural analyses, as well as our defect classification is evaluated empirically. The component-level and system-level reliability prediction methodology is evaluated using sensitivity, uncertainty, and complexity, and scalability analyses.