CriticalFault: Amplifying Soft Error Effect Using Vulnerability-Driven Injection

Abstract

As future microprocessors will be prone to various types of errors, researchers have looked into cross-layer hardware-software reliability solutions to reduce overheads. These mechanisms are shown to be effective when evaluated with statistical fault injection (SFI). However, under SFI, a large number of injected faults can be derated, making the evaluation less rigorous. To handle this problem, we propose a biased fault injection framework called Ciritical Fault that leverages vulnerability analysis to identify faults that are more likely to stress test the underlying reliability solution. Our experimental results show that the injection space is reduced by 30% and a large portion of injected faults cause software aborts and silent data corruptions. Overall, Critical Fault allows us to amplify soft error effects on reliability mechanism-under-test, which can help improve current techniques or inspire other new fault-tolerant mechanisms.