SmartInjector: Exploiting intelligent fault injection for SDC rate analysis

Recently, researchers have shown that exploiting symptom-based solutions provides a promising way to achieve low-cost fault tolerance. However, these solutions cannot provide enough reliability due to lack of ability to handle silent data corruptions (SDCs). To address SDCs efficiently, it is important to identify them effectively. This paper presents an intelligent fault injection framework, SmartInjector, which can analyze the SDC rates of the instructions in an application accurately with affordable fault simulation efforts. Based on program analysis, SmartInjector firstly prunes the faults whose outcome can be known without fault simulation or the faults which are considered to be equivalent to other faults. SmartInjector also employs the fault outcome prediction technique to reduce the time for a single fault simulation. The decreased requirement of computational resources allows SmartInjector to perform detailed fault injection experiments to analyze SDC rate for each instruction in an application. Validation results show that the fault simulation time needed by SmartInjector is only about 0.15% of the time required for simulating all possible faults, while the analysis accuracy of SDC rate reaches to 94% on average. Unlike existing fault injection frameworks, SmartInjector is the first to exploit fault outcome prediction to reduce the single simulation time.