2014/01/13 by James John Elliott, Elliott, James, Mark Frederick Hoemmen +3 · 1 citation
Engineering · Decision Sciences · Computer Science · #Radiation Effects in Electronics #Risk and Portfolio Optimization #Blockchain Technology Applications and Security
paper · pdf · doi:10.48550/arxiv.1401.3013
Future extreme-scale computer systems may expose silent data corruption (SDC)\nto applications, in order to save energy or increase performance. However,\nresilience research struggles to come up with useful abstract programming\nmodels for reasoning about SDC. Existing work randomly flips bits in running\napplications, but this only shows average-case behavior for a low-level,\nartificial hardware model. Algorithm developers need to understand worst-case\nbehavior with the higher-level data types they actually use, in order to make\ntheir algorithms more resilient. Also, we know so little about how SDC may\nmanifest in future hardware, that it seems premature to draw conclusions about\nthe average case. We argue instead that numerical algorithms can benefit from a\nnumerical unreliability fault model, where faults manifest as unbounded\nperturbations to floating-point data. Algorithms can use inexpensive "sanity"\nchecks that bound or exclude error in the results of computations. Given a\nselective reliability programming model that requires reliability only when and\nwhere needed, such checks can make algorithms reliable despite unbounded\nfaults. Sanity checks, and in general a healthy skepticism about the\ncorrectness of subroutines, are wise even if hardware is perfectly reliable.\n