Optimizing Large-Scale Fault Injection Experiments through Martingale Hypothesis: A Systematic Approach for Reliability Assessment of Safety-Critical Systems

Abstract

Functional safety of complex systems in safety-critical domains such as automotive, robotics and healthcare systems is paramount. Fault injection techniques play a pivotal role in the rigorous evaluation of functional safety. This paper introduces a novel approach to optimizing large-scale fault injection experiments by leveraging the Martingale hypothesis. By integrating such probabilistic models, our method enhances the efficiency of fault injection studies and enables deriving high confidence estimates of the diagnostic capabilities of safety critical systems using information gleaned from limited experiments. We demonstrate a pathway to achieving high functional safety by leveraging a combination hardware and software diagnostics. We validate our approach with extensive gate-level fault injection experiments performed over two years on NVIDIA’s Graphics Processing Units (GPUs) spanning over 11 million simulation hours. The presented findings and methodologies have broad implications for functional safety analysis, highlighting the transformative potential of this approach in overcoming previous challenges and advancing the state-of-the-art in fault injection.