Experimental analysis of models for correlation in multiversion software

Abstract

Two different classes of models for correlation in multiversion systems have been proposed in the literature. One considers that the version failures are statistically correlated, and the other defines additional failure types to account for related failures. We investigate these two model types, and apply them to several different sets of experimental data available in the literature. Both models provide a good fit for the available data, and thus both seem appropriate for models of multiversion systems. The models which explicitly consider correlation appear better suited to the analysis of software in isolation from the rest of the system, while those models which propose additional failure types are better suited for system-level models that consider hardware as well as software.<>