A cost model determines system costs for fault-tolerant software systems. The model finds the optimal number of program versions to achieve minimum system cost of the fault-tolerant software techniques: N-version programming, recovery block, and consensus recovery block. In this case, all versions, the voter, and the acceptance test have the same reliability. When the parameters for the versions, acceptance test, and voter in the cost function are all equal, the cost of a 3-version system is always optimal, Cost(CRB)/spl Lt/Cost(RB)/spl Lt/Cost(NVP) for each target reliability differing by as much as two orders of magnitude in some cases. The cost functions were increasing functions of N. When the parameters are not equal, optimality occurred for other values of N. This was especially the case when the cost exponent for version-1 was larger than the exponents for the other versions and the acceptance test or voter. As the values of the cost exponents for the version reliabilities become larger, a smaller difference was required between the version-1 exponent and the others to produce alternate optima.

In situations in which computers are used to manage life-critical situations, software errors that could arise due to inadequate or incomplete testing cannot be tolerated. This paper examines three methods of creating fault-tolerant software systems, Recovery Block, N-Version Programming, and Consensus Recovery Block, and it presents reliability models for each. The models are used to show that one method, the Consensus Recovery Block, is more reliable than the other two.