Software reliability: Repetitive run experimentation and modeling

A software experiment conducted with repetitive run sampling is reported. Independently generated input data was used to verify that interfailure times are very nearly exponentially distributed and to obtain good estimates of the failure rates of individual errors and demonstrate how widely they vary. This fact invalidates many of the popular software reliability models now in use. The log failure rate of interfailure time was nearly linear as a function of the number of errors corrected. A new model of software reliability is proposed that incorporates these observations.