Accurate FIT Rate Estimation Through High-Level Software Fault Injection

Reliability is today one of the major issues for computing devices from the embedded domain up to large high-performance systems. To safely deploy a computing device in a system, accurate measurements of the failure in time (FIT) rate are required to ensure that the device complies with the project reliability requirements. In this work, we present a method to provide a more accurate device FIT rate estimation from software fault injection. We use the technology sensitivity factor, obtained from beam experiments, as well as architectural and code features to significantly increase the FIT rate estimation accuracy. We compare the estimated FIT rates with the ones measured from radiation experiments of eight codes executed on ARM Cortex-A5 and Cortex-A9. We show that, on average, we can provide a FIT rate estimation accuracy of less than 20% (overestimation) and 35% (underestimation) the expected FIT rate for A5 and A9, respectively.