Comparison Criticality in Sorting Algorithms

Fault tolerance techniques often presume that the end-user computation must complete flawlessly. Though such strict correctness is natural and easy to explain, it's increasingly unaffordable for extreme-scale computations, and blind to possible preferences among errors, should they prove inevitable. In a case study on traditional sorting algorithms, we present explorations of a criticality measure defined over expected fault damage rather than probability of correctness. We discover novel 'error structure' in even the most familiar algorithms, and observe that different plausible error measures can qualitatively alter criticality relationships, suggesting the importance of explicit error measures and criticality in the wise deployment of the limited spare resources likely to be available in future extreme-scale computers.