SHARC: An efficient metric for selective protection of software against soft errors

Abstract

This paper presents a metric for the efficient application of selective hardening using software-based techniques against soft errors. It offers a method for selecting the resources to be protected obtaining maximum fault coverage with the minimum overhead. Common approaches are based on exhaustive exploration of the solution space or time-consuming fault injection campaigns. Contrarily, our Software based HARdening Criticality metric (SHARC) relies on early estimations of the impact that protection techniques will have on the global reliability of the application. SHARC estimations are based on features extracted from the dynamic analysis of source code and produce a prioritization of the resources involved accordingly. For assessing our approach two case studies were carried out using low-cost embedded microprocessors. Results were compared to traditional approaches like brute-force exploration and the Architectural Vulnerability Factor (AVF) metric. Experiments show that SHARC improves the results between 5% and 21% at a fraction of the effort.