An experimental study investigating models for N-version programming

N-version programming uses different programs, all written to the same specification, running in parallel; a majority vote on the respective outputs is taken as the correct answer. In an earlier experiment, failure counting and voting procedures for an iterative applications problem were developed. An experimental 3-version failure probability for a pool of program versions was obtained that was lower than a predicted value based on a simple model that assumes versions fail independently. Work is described that extends the previous experiment. The new 3-version experimental results again achieve lower failure probability than predicted by the simple model. The model is modified to take into account an intelligent voting algorithm that handles the no-majority case. A further modification incorporates into the model the number of iterations computed before failure status is determined. Experimental results are compared with the predictions of the current models