Predicting SDC Vulnerability of Instructions Based on Random Forests Algorithm

Silent Data Corruptions (SDCs) is a serious reliability issue in many domains of computer system. Selectively protecting of the program instructions that have a higher SDC vulnerability is one of the research hot spots in computer reliability field at present. A number of algorithms have already been presented to tackle this problem. However, many of them require tens of thousands of fault injection experiments, which are highly time and resource intensive. This paper proposes SDCPredictor, a novel solution that identify the SDC-vulnerable instructions based on random forests algorithm. SDCPredictor are based on static and dynamic features of the program alone, and do not require fault injections to be performed. SDCPredictor selectively protects the most SDC-vulnerable instructions in the program subject to a given performance overhead bound. Our experimental results show that SDCPredictor can obtain higher SDC detection efficiency than previous similar techniques.