High-level vulnerability over space and time to insidious soft errors

The integrity of computational results is being increasingly threatened by soft errors, especially for computations that are large-scale or performed under harsh conditions. Existing methods for soft error estimation do not clearly characterize the vulnerability associated with a particular result. 1) We propose a metric which captures the intrinsic vulnerability over space and time (VST) to soft errors that corrupt computational results. The method of VST estimation bridges the gap between the inherently low-level faults and the high-level computational failures that they eventually cause. 2) We define a model of an insidious soft error and try to clear up confusion around the concept of silent data corruption. 3) We present experimental results from three vulnerability studies involving floating-point addition, CORDIC, and FFT computations. The results show that traditional vulnerability metrics can be confounded by seemingly reliable but inefficient implementations which actually incur high vulnerability per computation. The VST method characterizes vulnerability accurately, provides a figure-of-merit for comparing alternative implementations of an algorithm, and in some cases uncovers pronounced and unexpected fluctuations in vulnerability.