Dear all,
I am a newbie on using MUMPS 4.7.3 and I have been using it to solve the
matrix problem AX=B where both A and B are sparse and in double complex
type. I just modified the zsimpletest.F provided by MUMPS. From the manual
I
have read that using iterative refinement (ICNTL(10)>0) the solution can
be
more accurate. However when I tried different positive values of ICNTL(10)
(I have used 1, 2, 3 and 30) the solution of X provided seems to deviate
much further away from the correct solution (I used Matlab as a
comparison).
With ICNTL(10)=0, the result of X is accurate for around 5-7 decimal
places,
but with a positive ICNTL(10), it seems the result of X becomes wrong;
even
the order of magnitude for many of the elements does not match.
Is it true that I have missed out something that causes the problem?
Moreover, are there any ways to make my solution more precise? The values
of
X in my problem solution are often very small, so a precision of 5 d.p.
may
not be good enough. (In particular, I found that the amount of non-zeros
calculated by MUMPS is much higher than that in Matlab, and by calculating
the residue value of B-AX, MUMPS gives a value of 10^-7 for my problem
while
Matlab gives a value of 10^-12.)
Best Regards,
LaBird (Benny).
[Email: Please remove all underscores for the correct email address.]
|
