Understanding a program's resiliency through error propagation

Aggressive technology scaling trends have worsened the transient fault problem in high-performance computing (HPC) systems. Some faults are benign, but others can lead to silent data corruption (SDC), which represents a serious problem; a fault introducing an error that is not readily detected nto an HPC simulation. Due to the insidious nature of SDCs, researchers have worked to understand their impact on applications. Previous studies have relied on expensive fault injection campaigns with uniform sampling to provide overall SDC rates, but this solution does not provide any feedback on the code regions without samples.

In this research, we develop a method to systematically analyze all fault injection sites in an application with a low number of fault injection experiments. We use fault propagation data from a fault injection experiment to predict the resiliency of other untested fault sites and obtain an approximate fault tolerance threshold value for each site, which represents the largest error that can be introduced at the site without incurring incorrect simulation results. We define the collection of threshold values over all fault sites in the program as a fault tolerance boundary and propose a simple but efficient method to approximate the boundary. In our experiments, we show our method reduces the number of fault injection samples required to understand a program's resiliency by several orders of magnitude when compared with a traditional fault injection study.