A Program-Aware Fault-Injection Method for Dependability Evaluation Against Soft-Error Using Genetic Algorithm

Abstract

Decreasing the scale of transistors and exponential increase in the transistor counts has made the soft-errors as one of the major causes of software failures. Fault injection is a powerful method for dependability assessment of a computer system against soft-errors. A considerable number of randomly injected faults in the current methods and tools are effect-less or equivalent. To overcome this problem and reduce the cost of fault injection, this study presents a software based fault-injection method that accurately evaluates the dependability of a computer system with a limited number fault-injection. Using a genetic algorithm (GA) the most vulnerable executable paths of an input program is identified; then only the basic blocs (BBs) into the identified vulnerable paths are considered as the target of fault injection. The results of fault injections on the set of 8 traditional benchmark-programs show that the proposed method reduces about 20% of effect-less faults by avoiding the injection of faults in the error-derating blocks of a program. Furthermore, the number of injected faults is reduced to 60% of its original size in the random injection. Also, the proposed method provides more stable and accurate results than the random injection.