Increasing the efficiency and efficacy of selective-hardening for parallel applications

Selective hardening is a promising solution to efficiently improve the reliability of high performance and safety-critical real-time applications. One of the most significant challenges of selective hardening is to choose how many resources or code portions to protect to avoid unnecessary system performances degradation (at least 2x overhead for the naive duplication). In this paper, we propose a selective hardening strategy for parallel algorithms. We first identify through extensive fault-injection campaigns the code portions whose protection significantly increases the algorithm reliability. Then, we select the code portions that, once protected, maximize the reliability/overhead ratio. We can achieve fault coverage as high as 60% with a 3% overhead. We show that the hardening efficiency can be higher than 90% when compared to naive full duplication.