Fault Tolerance of Sorting Algorithms Under Varying Input Characteristics

Bit flips in main memory or in the CPU can lead to silent data corruptions (SDCs) and therefore pose a threat to all modern computer systems. The selection of particularly fault-tolerant algorithms is known to have a strong impact on a system's resilience, but involves time-consuming fault-injection (FI) campaigns. This problem is exacerbated by the fact that the optimal selection also depends on input characteristics like data length or structure. In this paper, we exemplarily analyze different sorting algorithms with respect to their fault tolerance. Confirming and complementing other studies, we show that input-data length and ordering are relevant for algorithm selection. We find that existing hybrid algorithms, like Timsort or Introsort, also make favorable choices for fault tolerance and dominate the field. In a further consideration, we correlate the SDC probabilities with simpler metrics, such as the number of retired instructions. We find that these metrics can also be used to select the most resilient sorting algorithm, as the resulting algorithm ranking is strongly correlated to one obtained from an expensive FI campaign.