TABLE OF CONTENTS Section 7.1 explained that software failures can frequently be overcome by just restarting the program because the failure occurred only when a particular data value or timing relationship was encountered. When the program is not changed, the failure can be expected to recur at a frequency that is dependent on encountering the conditions that trigger the failure. When a program is in widespread use (so that local variations are smoothed out) the failure rate due to specific causes will remain constant [30].
Where faults are being removed as they are found, such as is typical in testing, the failure rate can be expected to decrease, leading to reliability growth. In most cases reference to software reliability models means the software reliability growth model. A number of these models have been defined in the previously mentioned ANSI standard [4]. Most assume that the fault exposure ratio is constant. Typical of these is the Musa-Okumoto model
where ? o is the initial failure rate, ? is an empirical decay factor, and t is execution time. In contrast, the Littlewood-Verall (L-V) model allows for a changing fault exposure ratio. As discussed previously, there is frequently a transition from failures that occur under routine conditions to those that occur only under exception conditions and ultimately to those that occur under multiple exception conditions. These transitions make it likely that there are changes in the fault exposure ratio and thus make the L-V model preferable. However, the L-V...
