Three modified dependency-tests for software failures

Abstract

Multiple version software strategy was proposed for the fault-tolerant system. An N-version software (NVS) system raises reliability by employing independently developed software versions. Various NVS systems were experimented upon. While all reported reliability improvement, it was also evidenced that independently developed software component versions still risked sharing common faults and resulted in coincident failures. Previously, three dependency tests were proposed, namely, the approximate z-test, dependency factor, and /spl chi//sup 2/-test. Unfortunately, rather than testing for the dependency of failures, those tests examine the dependency of outputs that include not only failures but also successful ones. For the remedy, those tests are modified to fit for the test of failure dependency. Demonstration examples using previously reported experimental data are presented.