HP MLIB User's Guide Vol. 2 7th Ed.

1038 HP MLIB User’s Guide
DSEVEX Eigenextraction
• A block size of 1 is usually less efficient than a block
size of 2.
• Block sizes larger than 10 are usually less efficient
than smaller block sizes.
•Ifmxbksz is zero or negative, a default block size
optimized to the computer system will be chosen.
In all cases, the Lanczos algorithm may find it
necessary to reduce the block size to fit the problem
into the storage available.
tolact An accuracy tolerance for the eigenvalues, which must
be in the range from ≤ tolact ≤∈, where ∈ is the
machine precision. If a negative or zero value is
supplied, a default value of ∈
2/3
will be used. Special
care should be taken if the value of tolact is reduced
below its default value. In particular, it is important to
be sure that the computed eigenvectors remain
adequately orthogonal. Subroutine DSEVCK can
facilitate such monitoring.
shfscl An estimate of the magnitude of the smallest nonzero
eigenvalues. The package will compute a heuristic
estimate if the user does not supply an estimate
(signaled by shfscl ≤ 0.0). This argument is provided
only to assist the algorithm in cases where zero lies in
the interval of interest or when it is not appropriate to
use a shift of 0.0 either because A or K is singular (has
rigid body modes) or because this is a buckling
analysis.
nusrvc Number of user starting vectors provided. Any
approximate eigenvectors supplied by the user will be
used to create a starting block for the Lanczos
algorithm. Provision of starting vectors can be effective
when only a very few eigenvalues are to be computed,
but it has limited effect otherwise.
usrsvc Array containing the starting vectors, one per column.
Referenced only if nusrvc is positive. norder is the
problem size as set in DSEVIN.
Updated global Global communications array for this problem. This
array must be passed, untouched by the user, to
successive subroutines in this package.
ε