Fault tolerant QR-decomposition algorithm and its parallel implementation

Oleg Maslennikov and Juri Kaniewski and Roman Wyrzykowski

Abstract
A fault tolerant algorithm based on Givens rotations and a modifiedweighted checksum method is proposed for the QR-decomposition of matrices. The aim is to detect and correct calculation errors occured due to transient hardware faults during computations. The proposedalgorithm enables to correct a single error in each row or column ofan input M x N matrix A, occured at any among N steps of the algorithmimplementation. This effect is obtained by incresing the computational complexity of the original algorithm slightly. A parallel version of this algorithm is designed, dedicated for realisation on a fixed-sizedlinear processor array with fully local communications and lowI/O requirements.
Contact
Oleg Maslennikov
Institute of Electronics,Technical University of Koszalin,Partyzantow 17,75-411 Koszalin,Poland,
ppam@matinf.pcz.czest.pl