Soft Error Resilience of Probabilistic Inference Applications

With shrinking device size and increasing complexity, soft errors are becoming an issue in the reliability of digital systems. To make efficient robust systems, it is important to understand how soft errors affect the quality of output for the target applications. Probabilistic inference applications are interesting since they produce non-exact results and yet are useful in many different fields. Our fault injection experiments show that some of these applications can mask or quickly recover from most transient data errors. In addition, their approximate nature enables low cost fault recovery mechanisms for control flow errors. This allows us to use simple software modifications and checkpointing to drastically reduce the number of program crashes. These adjustments can be combined with control flow checking as an effective way to protect approximate probabilistic inference applications.