Type Inference for Soft-Error Fault-Tolerance Prediction

Abstract

Software systems are becoming increasingly vulnerable to a new class of soft errors, originating from voltage spikes produced by cosmic radiation. The standard technique for assessing the source-level impact of these soft errors, fault injection - essentially a black-box testing technique - provides limited high-level information. Since soft errors can occur anywhere, even control-structured white-box techniques offer little insight. We propose a type-based approach, founded on data-flow structure, to classify the usage pattern of registers and memory cells. To capture all soft errors, the type system is defined at the assembly level, close to the hardware, and allows inferring types in the untyped assembly representation. In a case study, we apply our type inference scheme to a prototype brake-by-wire controller, developed by Volvo Technology, and identify a high correlation between types and fault-injection results. The case study confirms that the inferred types are good predictors for soft-error impact.