Program-Structure–Guided Approximation of Large Fault Spaces

Abstract

Due to shrinking structure sizes and operating voltages, hardware becomes more susceptible to transient faults. Fault injection campaigns are a common approach to systematically assess the resilience of a system and the effectiveness of software-based counter measures. However, experimentally injecting all possible faults to achieve full fault-space coverage is infeasible in practice. While precise pruning techniques, such as def/use pruning, already provide a significant reduction of the campaign size, the number of injections remains still challenging for even medium-sized systems. We propose fault-space regions (FSRs) as a method to approximately cover the complete fault space with a significantly lower number of required injections. Instead of probabilistic subsampling of the fault space, our approximation exploits the actual program structure and execution trace (e.g., flow of basic blocks) to identify injection points that are representatives for a larger set of faults. We identify such data-flow regions and inject only data values that flow across region boundaries. Thereby, we can further reduce the number of injections by up to 76 percent, while the results divert only by less than 2.7 percent from those of a complete and precise fault-injection campaign. Furthermore, we keep the locality of the results regarding silent data corruptions to a deviation of less than 6.9 percent.