[Trilinos-Users] aztecoo+sundance problem
Gideon Simpson
grs2103 at columbia.edu
Wed Jan 17 20:35:58 MST 2007
I should have posted this too:
AZ_check_input: Setup information on processor 0
solver: 1
convergence flag: 0
maximum iterations: 5000
reordering: 1
preconditioner: 14
polynomial order: 3
GMRES ill conditioning threshold: 1.0e+11
GMRES restart size: 30
orthogonalization: 0
tolerance: 1.0e-12
drop: 0.0e+00
fill-in: 1
overlap: 0
typeoverlap: 0
Number of internal unknowns: 0
Number of border unknowns: 10691
Total number of unknowns: 10691
Number of external unknowns: 0
Number of internal blocks: 0
Number of border blocks: 10691
Total number of blocks: 10691
Number of external blocks: 0
Number of processors: 1
Node number: 0
Number of neighbors: 0
Number of unknowns sent to neighbors: 0
========================================================================
=======
*******************************************************
***** Problem: Epetra::CrsMatrix
***** Preconditioned GMRES solution
***** ILU(1) domain decomp. without overlap
***** No scaling
***** NOTE: convergence VARIES when the total number of
***** processors is changed.
*******************************************************
iter: 0 residual = 1.000000e+00
*********************************************************************
***** Condition number estimate for subdomain preconditioner on PE 0
= 1.4400e+02
*********************************************************************
I'm getting this through sundance, not by explicit manipulation of
the epetra matrices, so I hope this is the calculation you wanted.
In addition, I did try adjusting the alpha and rho threshold
parameters as described in the AztecOO User's Guide and saw no
improvement.
-gideon
On Jan 17, 2007, at 10:04 PM, Mike Heroux wrote:
> Gideon,
>
> This is often a symptom of a poorly conditioned preconditioner. In
> other words, the LU factors of ILU are poorly conditioned, which
> can be true even if your original matrix is well-conditioned. You
> can check if this is the case by calling the Condest() method on
> your AztecOO object. If the value comes back as 1E+14 or more,
> then this is likely the issue. It is possible to reduce the
> condition number and still have an effective preconditioner by
> using the absolute and relative threshold values to perturb the
> diagonal of the preconditioner. This is described in the AztecOO
> User Guide, starting on page 27:
>
> http://software.sandia.gov/trilinos/packages/aztecoo/
> AztecOOUserGuide.pdf
>
> If this is not the problem, let us know. Also, there may be other
> simpler ways to reduce the condition number of the preconditioner,
> but we need to know more about your problem and the particular ILU
> parameters you are using.
>
> Mike
>
>
> From: trilinos-users-bounces at software.sandia.gov [mailto:trilinos-
> users-bounces at software.sandia.gov] On Behalf Of Gideon Simpson
> Sent: Wednesday, January 17, 2007 7:52 PM
> To: trilinos-users at software.sandia.gov
> Subject: [Trilinos-Users] aztecoo+sundance problem
>
> I've been getting the following error when solving some problems
> that come out of Sundance, using GMRES+DomainDecomposition+ILU
>
> ***************************************************************
>
> Warning: a breakdown in this method
> has occurred and solution has not converged.
>
> Solver: gmres
> number of iterations: 1
>
> Recursive residual = nan
>
> Calculated Norms Requested Norm
> -------------------------------------------- --------------
>
> ||r||_2 / ||r0||_2: 1.000000e+00 1.000000e-12
>
> ***************************************************************
>
>
>
> Solution time: 9.740000 (sec.)
> total iterations: 1
> Epetra ERROR -2, ../../../../packages/aztecoo/src/AztecOO.cpp, line
> 824
> Solve failed! state = Crashed
> message=Aztec solver failed: numerical breakdown
> iters taken = 1
> final residual = nan
> matrix written to badMatrix.dat
> RHS vector written to badVector.dat
>
> The matrix and RHS data files look fine, and when I solve the
> problem using an Amesos direct solver, the answer is consistent.
> What is the nature of this error and how might I correct it?
>
> -Gideon Simpson
> Department of Applied Physics and Applied Mathematics
> Columbia University
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://software.sandia.gov/mailman/private/trilinos-users/attachments/20070117/0947c863/attachment-0001.html
More information about the Trilinos-Users
mailing list