Experiments in fault tolerant software reliability

Twenty functionally equivalent programs were built and tested in a multiversion software experiment. Following unit testing, all programs were subjected to an extensive system test. In the process sixty-one distinct faults were identified among the versions. Less than 12 percent of the faults exhibited varying degrees of positive correlation. The common-cause (or similar) faults spanned as many as 14 components. However, a majority of these faults were trivial, and easily detected by proper unit and/or system testing. Only two of the seven similar faults were difficult faults, and both were caused by specification ambiguities. One of these faults exhibited variable identical-and-wrong response span, i.e. response span which varied with the testing conditions and input data. Techniques that could have been used to avoid the faults are discussed. For example, it was determined that back-to-back testing of 2-tuples could have been used to eliminate about 90 percent of the faults. In addition, four of the seven similar faults could have been detected by using back-to-back testing of 5-tuples. It is believed that most, if not all, similar faults could have been avoided had the specifications been written using more formal notation, the unit testing phase was subject to more stringent standards and controls, and better tools for measuring the quality and adequacy of the test data (e.g. coverage) were used.