Design and evaluation of fault tolerance techniques for highly parallel architectures

Summary form only given. The author discusses fault tolerance techniques for computer systems, including a new technique, which he calls algorithm-based fault tolerance, for error detection and correction when computations are performed using multiple processor systems. The technique uses knowledge about the algorithm to reduce the amount of overhead necessary for fault tolerance. This is done by appropriately encoding the data and tailoring the algorithms to operate on the encoded data and produce encoded output data. Examples are given of applications including matrix operations, fast Fourier transforms, and computation of eigenvalues