Experimental Comparison of Software Reliability Models Based on Code Coverage and on Time Domain

The study of software reliability has its origin in 1967 when the first experiments were performed. Since then, diverse models that use as parameter the time of execution of software or the time between failures have been considered. Another approach indicates that the use of code coverage, instead of execution time, yields better estimates of the reliability of the software. This paper presents the results observed in an experiment carried out to compare the software reliability models Binomial Model based on Coverage (BMBC) and Infinite Failure Software Reliability Model Based on Code Coverage (IFMBC) with the following models based on time domain: Geometric (GEO), Littlewood-Linear Verral (LAV-L), Littlewood-Quadratic Verral (LAVQ), Musa Basic (MS-B), Musa Logarithmic (MS-L) and Non-Homogeneous Poisson Model (NHP). These models, in addition to being much quoted in the literature, are implemented by SMERFS^3, used to calculate its parameters and to estimate reliability. As a result, it has been observed that traditional models do not provide robustness when the profile was changed. On the other hand, code coverage-based models have produced good results for all operational profiles. In addition to the good fit, coverage-based models have generated better estimates of software reliability, as shown by the Kolmogorov-Smirnov test.