Householder bidiagonalization on parallel computers with dynamic ring architecture

A parallel algorithm for Householder bidiagonalization on parallel computers with dynamic ring architecture is presented. The Householder bidiagonalization is the core for singular value decomposition (SVD) which has been found to be very useful as an analytical tool in the presence of roundoff error and inexact data. Two sided Householder reduction/expansion technique is applied for bidiagonalization. Innovative systolic like communication techniques are proposed which eliminate the need for computing explicitly the transpose of the matrix. The experimental study on the CM-5 shows that the parallel algorithm developed in the article achieves high speedup for large matrices.