Understanding the Effects of Data Corruption on Application Behavior Based on Data Characteristics

In this paper, the results of an experimental study on the error sensitivities of application data are presented. We develop a portable software-implemented fault-injection (SWIFI) tool that, on top of performing single-bit flip fault injections and capturing their effects on application behavior, is also data-level aware and tracks the corrupted application data to report their high-level characteristics (usage type, size, user, memory space location). After extensive testing of NPB-serial (7.8M fault injections), we are able to characterize the sensitivities of data based on their high-level characteristics. Moreover, we conclude that application data are error sensitive in parts; depending on their type, they have distinct and wide less-sensitive bit ranges either at the MSBs or LSBs. Among other uses, such gained insight could drive the development of sensitivity-aware protection mechanisms of application data.