Precisely controlling the duration of fault injection campaigns: a Statistical view

Fault injection has become the main approach to evaluate the dependability of a circuit with respect to soft errors. It can be applied early during the design process and refined along all design steps. However, the main limitation is the huge number of faults that should be injected when large circuits and/or complex workloads and complex error models are considered. This usually leads to practical campaigns based on a randomly selected subset of all potential faults. However, results obtained in these conditions are only estimations and the error on the estimation is never quantified in the literature. This paper presents an approach to quantify both the error on the presented results and the confidence on the computed error interval. Conversely, the computation of the required number of faults to inject in order to achieve a given confidence and error interval is also discussed. Experimental results are shown, fully supporting the presented approach.