next up previous contents index
Next: Eigenvalue Problems Up: Solving Linear Systems of Previous: Solving Linear Systems of

Solving Linear Least Squares Problems

 

Table 5.11  summarizes performance results obtained for the ScaLAPACK routine PSGELS /PDGELS  that solves full-rank linear least squares problems. Solving such problems of the form tex2html_wrap_inline16768, where x and b are vectors and A is a rectangular matrix having full rank is traditionally achieved via the computation of the QR factorization of the matrix A. In ScaLAPACK, the QR factorization   is based on the use of elementary Householder   matrices of the general form
displaymath16782
where v is a column vector and tex2html_wrap_inline14435 is a scalar. This leads to an algorithm with excellent vector performance, especially if coded to use Level 2 PBLAS.

The key to developing a distributed block form of this algorithm is to represent a product of K elementary Householder matrices of order N as a block form of a Householder matrix.   This can be done in various ways. ScaLAPACK uses the form [108]
displaymath16792
where V is an N-by-K matrix whose columns are the individual vectors tex2html_wrap_inline16800 associated with the Householder matrices tex2html_wrap_inline16802 and T is an upper triangular matrix of order K. Extra work is required to compute the elements of T, but this is compensated for by the greater speed of applying the block form.

  table4001
Table 5.11: Speed in Mflop/s of PSGELS/PDGELS for square matrices of order N



Susan Blackford
Tue May 13 09:21:01 EDT 1997