Fault tolerant systems based on the use of software design diversity may be able to achieve high levels of reliability more cost-effectively than other approaches, such as heroic debugging. Earlier experiments have shown that the reliabilities of multi-version software systems are more reliable than the individual versions. However, it is also clear that the reliability benefits are much worse than would be suggested by naive assumptions of failure independence between the versions. It follows that it is necessary to assess the reliability actually achieved in a fault tolerant system. The difficulty here mainly lies in acquiring knowledge of the degree of dependence between the failures processes of the versions. The paper addresses the problem using Bayesian inference. In particular, it considers the problem of choosing a prior distribution to represent the beliefs of an expert assessor. It is shown that this is not easy, and some pitfalls for the unwary are identified. (Presented at...
Bev Littlewood, Peter T. Popov, Lorenzo Strigini