[Trilinos-Users] [EXTERNAL] preconditioned GMRES: no convergence in explicit residual
C.S. Natarajan
csnataraj at gmail.com
Fri Jan 31 17:02:29 MST 2014
Don't mean to take this discussion off point. The solution looks
interesting for a couple of reasons (assuming it is a surface graph of the
solution)
1) Looks like the solution is violating the discrete maximum principle.
Isn't this an isotropic elliptic problem? (Probably just me misinterpreting
something here)
2) With preconditioners, both solutions seem off in capturing the physics
as well. AMG is just bonkers but ilu looks a little diffusive at the
interior and terrible at the boundary. Not sure what this means, just an
observation.
Cheers,
C.S.N
> Are you using left or right preconditioning, what happens when you
> switch to the other side?
Good question! I was using left preconditioning, but it doesn't matter to
which side it is applied: the result will be the same.
I checked the results that come out when not paying attention to the large
explicit residual: The solutions are indeed off if a preconditioner is used
(see attachments). I'm not sure what to make of this.
I tried CG and MINRES in comparison, and both work well in all cases.
Any further ideas?
--Nico
On Fri, Jan 31, 2014 at 8:59 PM, Thornquist, Heidi K <hkthorn at sandia.gov>
wrote:
> Hi Nico,
>
> This is interesting. Are you using left or right preconditioning,
> what happens when you switch to the other side?
> I know that left preconditioning, because you are starting with a
> preconditioned RHS, can give a residual that is off from the residual
> of the original system. But this is pretty extreme.
>
> I've tried running the preconditioned examples of pseudo block GMRES
> varying tolerances and I'm not seeing any abnormalities.
>
> Hmmmm....
>
> Heidi
>
> --
>
> Heidi K. Thornquist
> Electrical Models & Simulation
> Sandia National Laboratories
> Albuquerque, NM 87185-1323
>
>
>
>
>
> On 1/31/14 10:58 AM, "Nico Schlömer" <nico.schloemer at gmail.com> wrote:
>
>>Hi all,
>>
>>I've got this matrix corresponding a finite-element discretization of
>>the Poisson problem on the unit square (with Dirichlet BC, no issues
>>here). Running the problem through Belos with CG and ML, Ifpack gives
>>no problems at all. The PseudoBlockGMRES solver works fine when
>>unpreconditioned, but yields something weird when using a
>>preconditioner:
>>
>>============== *snip* ==============
>>*******************************************************
>>***** Belos Iterative Solver: Pseudo Block Gmres
>>***** Maximum Iterations: 100
>>***** Block Size: 1
>>***** Residual Tests (SEQ):
>>***** Test 1 : Belos::StatusTestGenResNorm<>: (2-Norm Imp Res Vec) /
>>(2-Norm Prec Res0), tol = 1e-13
>>***** Test 2 : Belos::StatusTestGenResNorm<>: (2-Norm Exp Res Vec) /
>>(2-Norm Res0), tol = 1e-13
>>*******************************************************
>>Iter 0, [ 1] : 1.000000e+00 ---
>>Iter 1, [ 1] : 6.120593e-01 ---
>>Iter 2, [ 1] : 2.695238e-01 ---
>>Iter 3, [ 1] : 5.704995e-02 ---
>>Iter 4, [ 1] : 1.272673e-02 ---
>>Iter 5, [ 1] : 2.783810e-03 ---
>>Iter 6, [ 1] : 5.753817e-04 ---
>>Iter 7, [ 1] : 2.045848e-04 ---
>>Iter 8, [ 1] : 5.724907e-05 ---
>>Iter 9, [ 1] : 1.633313e-05 ---
>>Iter 10, [ 1] : 3.055064e-06 ---
>>Iter 11, [ 1] : 4.947173e-07 ---
>>Iter 12, [ 1] : 6.945227e-08 ---
>>Iter 13, [ 1] : 1.859584e-08 ---
>>Iter 14, [ 1] : 5.185228e-09 ---
>>Iter 15, [ 1] : 5.922331e-10 ---
>>Iter 16, [ 1] : 8.149474e-11 ---
>>Iter 17, [ 1] : 1.010586e-11 ---
>>Iter 18, [ 1] : 1.600644e-12 ---
>>Iter 19, [ 1] : 3.034731e-13 ---
>>Iter 20, [ 1] : 3.452280e-14 7.213315e-01
>>[...]
>>============== *snap* ==============
>>
>>The ExpResVec never converges at all.
>>
>>What may cause this? Any known fix/workaround?
>>
>>Cheers,
>>Nico
>>_______________________________________________
>>Trilinos-Users mailing list
>>Trilinos-Users at software.sandia.gov
>>http://software.sandia.gov/mailman/listinfo/trilinos-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20140131/b146d4ce/attachment.html
More information about the Trilinos-Users
mailing list