Thanks Bernd, BiCGStabSSOR solver converged on my 50 cube problem, and it
appears to be running on 180 cube problem (the largest my workstation memory
can handel).
I really appreciate the quick attention and advice. I will never be able to
figure this out.
I hope I am all set for my current project, and will be happy to share with you
the final results.
Thank you all again for your kind help.
Best,
-Shawn
From: Bernd Flemisch <[email protected]>
To: DuMuX User Forum <[email protected]>
Sent: Monday, April 2, 2012 12:34 PM
Subject: Re: [DuMuX] Permeability heterogeneity and grid resolution
Hey Shawn,
On Mon, 2 Apr 2012 06:35:49 -0700 (PDT)
Shawn Zhang <[email protected]> wrote:
> Hi Bernd,
>
> Thanks for the suggestions.
>
> I am using the release. All 3D runs work if the resolution is low (say,
> 10x10x10), but diverge for higher resolution. If the problem is related to
> lens z-coordinate being undefined, the code should break regardless the
> resolution, do you agree?
No. Pretty much anything can happen. Most often the uninitialized z-values will
be zero, thus there will be no lens. But also other numbers can be in there,
making it unpredictable. So you really have to incorporate the corresponding
stuff from my 1ptestspatialparameters.hh file.
> Moreover, in the example I attached in my last email, I assigned a tensor
> permeability for each cell. Can you please take a quick look, and suggest
> where I did wrong?
This seems fine.
> Again, the code runs with lower resolution, but not higher.
> Waiting nervously for your advice.
No need to become nervous. I attach my current problem file and input file with
which it works in 3d for me up to 50x50x50, I did not test larger.
This time, the trick for me was to use BiCGStabSSOR instead of CGSSOR as a
solver. And I even can come up with a handwaving explanation. In principle, the
matrix should be symmetric. But this is not explicitly enforced in the code, so
a_ij and a_ji are calculated separately. Due to numerical errors, it will be
slightly unsymmetric and the CG solver gets into trouble. BiCGStab does not
care.
Hope this helps. Kind regards
Bernd
>
> Best,
> -Shawn
>
> From: Bernd Flemisch <[email protected]>
> To: DuMuX User Forum <[email protected]> Sent: Monday, April 2,
> 2012 3:17 AM
> Subject: Re: [DuMuX] Permeability heterogeneity and grid resolution
>
> Hey Shawn,
>
> thank you for your further investigations. You are right that the options for
> the linear solver are not well documented yet. You can choose one out of
> dumux/linear/boxlinearsolver.hh. The parameters are documented in
> dumux/linear/linearsolverproperties.hh. About the actual meaning of, e.g., a
> relaxation parameter or a preconditioner, you should consider your favorite
> linear algebra textbook.
>
> Concerning your 3d calculations: are you using the release or the trunk? If
> you use the release, 3d can break, since the z-coordinate of the lens is not
> set. You can use the attached 1ptestspatialparameters.hh. This is already
> fixed in the trunk and will be fixed in a bugfix release 2.1.1.
>
> Kind regards
> Bernd
>
> On Sat, 31 Mar 2012 19:02:42 -0700 (PDT)
> Shawn Zhang <[email protected]> wrote:
>> Hi Bernd and All,
>>
>> I confirm that Bernd's input file resolved the convergence problem I had.
>> Thanks.
>>
>> However (I hate it when people do this to me, but sorry I can't find other
>> way out), I am still puzzled by this convergence issue. It seems the
>> following factors are influencing the convergence,
>>
>> 1. Resolution. The higher the resolution, the more difficult the convergence.
>> 2. Heterogeneity. The higher the heterogeneity, the more difficult the
>> convergence.
>> 3. Linear solver type. Not sure what are the options and how they are
>> related to convergence.
>> 4. Solve controls: relaxation, preconditioners, and number of iterations. I
>> only understand the last one.
>>
>> I most puzzled by #1. Why higher grid resolution causes non-convergence?
>> Attached a more specific problem I am working on. The set up converge at
>> 10x10x10, but diverge at 50x50x50 or higher.
>> Would you mind taking a look and suggest how to make it converge? I need to
>> run it at 200x200x200 resolution.
>>
>> In addition to the problem, I also modified 1pmodel.hh file to handle tensor
>> perm. I also have a question about why using Scalar tensor in all the
>> boxmodels tests, but this is not critical (unless it is related to the
>> convergence issue).
>>
>> Sorry about all the questions, but I am about to pull my hair now.
>>
>> -Shawn
>>
>> From: Bernd Flemisch <[email protected]>
>> To: DuMuX User Forum <[email protected]> Sent: Thursday, March
>> 29, 2012 5:23 AM
>> Subject: Re: [DuMuX] Permeability heterogeneity and grid resolution
>>
>> Sorry, wrong file. Attached is the correct one.
>>
>> Kind regards
>>
>>
>> On Thu, 29 Mar 2012 10:36:57 +0200
>> Bernd Flemisch <[email protected]> wrote:
>>> Hey Shawn,
>>>
>>> I was not too specific about what to change: you should set
>>> PreconditionerIterations to 3, not PreconditionerRelaxation. You can also
>>> set this in the input file, it overrides the settings of the problem file.
>>>
>>> I attach my current working input file for your convenience.
>>>
>>> Kind regards
>>> Bernd
>>> On Wed, 28 Mar 2012 12:42:13 -0700 (PDT)
>>> Shawn Zhang <[email protected]> wrote:
>>>> Hi Bernd and All,
>>>>
>>>> Sorry if I appear to be sticky about the convergence issue, but it kept
>>>> coming back and bite me.
>>>>
>>>> Coming back to this test problem, with the attached 1p project, I have the
>>>> following issue.
>>>>
>>>> Standard test: worked.
>>>> Increase grid size to 100x100, not converging. It worked after I, increase
>>>> # iteration to 1000, reduce convergence criterion to 10e-6.
>>>> Increase grid size to 1000x1000, not converging. I changed the linear
>>>> solver from Dumux::BoxCGILU0Solver to Dumux::BoxCGSSORSolver, and
>>>> PreconditionerRelaxation from 1.0 to 3.0 as suggested by Bernd, it is
>>>> still not converging...
>>>>
>>>> I attached the project folder.
>>>> Do you think this maybe related to the Oracle VBox based Debian
>>>> installation?
>>>>
>>>> Thanks.
>>>> -Shawn
>>>>
>>>> From: Bernd Flemisch <[email protected]>
>>>> To: DuMuX User Forum <[email protected]> Sent: Tuesday,
>>>> March 27, 2012 4:01 AM
>>>> Subject: Re: [DuMuX] Permeability heterogeneity and grid resolution
>>>>
>>>> Hey Shawn,
>>>>
>>>> with the attached input file, it works for 1000x1000 and four orders
>>>> difference. I increased the maximum number of iterations to 1000 and the
>>>> residual reduction to 1e-6.
>>>>
>>>> Kind regards
>>>> Bernd
>>>>
>>>> On Mon, 26 Mar 2012 11:44:21 -0700 (PDT)
>>>> Shawn Zhang <[email protected]> wrote:
>>>>> Well, I celebrated too fast.
>>>>>
>>>>> My goal is to determine, in a system of two material with different
>>>>> perm, when one material can be treated as having negligible impact on
>>>>> pressure drop.
>>>>>
>>>>> The two parameters influencing this are,
>>>>>
>>>>> 1. Permeability difference between two materials.
>>>>> 2. Grid resolution.
>>>>>
>>>>> I use the 1p with lens test problem, and increasing the perm of the lens
>>>>> material.
>>>>>
>>>>> At 100x100 resolution, I derive the attached plot (pressure plot along
>>>>> y=0.5 line), which is saying that when the permeability difference is
>>>>> greater than 3 order of magnitude, the material with high permeability
>>>>> can be seen as open channel flow comparing to the low permeable phase,
>>>>> and the pressure drop is neglegible. Moreover, further increase of the
>>>>> perm difference will have a neglegible effect.
>>>>>
>>>>> At 10x10 resolution, we start seeing this when the perm difference is 2
>>>>> order of magnitude.
>>>>>
>>>>> When I increase the resolution to 1000x1000, with the permeability
>>>>> difference 4 order of magnitude, the solver does not converge again.
>>>>>
>>>>> Is this the expected behavior, as I am trying to solve large
>>>>> discontinuity caused by the big perm difference?
>>>>>
>>>>> -Shawn
>>>>>
>>>>>
>>>>>
>>>>> From: Bernd Flemisch <[email protected]>
>>>>> To: DuMuX User Forum <[email protected]> Sent: Monday,
>>>>> March 26, 2012 12:45 PM
>>>>> Subject: Re: [DuMuX] Permeability heterogeneity and grid resolution
>>>>>
>>>>> Hey Shawn,
>>>>>
>>>>> here, it helps to change the linear solver (or rather the preconditioner
>>>>> from ILU to SSOR). This unfortunately is a compile-time choice done in
>>>>> the problem file. I attach my corresponding problem file, changes are in
>>>>> lines 83 and 86.
>>>>>
>>>>> Kind regards
>>>>> Bernd
>>>>>
>>>>> On Mon, 26 Mar 2012 09:00:18 -0700 (PDT)
>>>>> Shawn Zhang <[email protected]> wrote:
>>>>>> Hi Bernd,
>>>>>>
>>>>>> I switched the permeability between lens and outside, which gives the
>>>>>> non-convergence exception.
>>>>>>
>>>>>> I.e.,
>>>>>>
>>>>>> permeability = 1e-12
>>>>>> permeabilityLens = 1e-10
>>>>>>
>>>>>> Instead of
>>>>>>
>>>>>>
>>>>>> From: Bernd Flemisch <[email protected]>
>>>>>> To: DuMuX User Forum <[email protected]> Sent: Monday,
>>>>>> March 26, 2012 11:08 AM
>>>>>> Subject: Re: [DuMuX] Permeability heterogeneity and grid resolution
>>>>>>
>>>>>> Hey Shawn,
>>>>>>
>>>>>> ok, this is difficult since I cannot reproduce it. Can you please send
>>>>>> the output of your run with 100x100 and the attached input file? I
>>>>>> increased the verbosity of the linear solver, so maybe we can see more
>>>>>> from there.
>>>>>>
>>>>>> Kind regards
>>>>>> Bernd
>>>>>>
>>>>>>
>>>>>> On Mon, 26 Mar 2012 07:53:03 -0700 (PDT)
>>>>>> Shawn Zhang <[email protected]> wrote:
>>>>>>> Hi Bernd,
>>>>>>>
>>>>>>> Thanks a lot for your quick attention. However, with your input file,
>>>>>>> and grid resolution 100x100, I got the same exceptions. I also tried to
>>>>>>> increase ResidualReduction parameter even further without any luck.
>>>>>>>
>>>>>>> It puzzles me that the code runs at lower resolution but not at
>>>>>>> higher...
>>>>>>>
>>>>>>> -Shawn
>>>>>>>
>>>>>>> From: Bernd Flemisch <[email protected]>
>>>>>>> To: DuMuX User Forum <[email protected]> Sent: Saturday,
>>>>>>> March 24, 2012 6:07 AM
>>>>>>> Subject: Re: [DuMuX] Permeability heterogeneity and grid resolution
>>>>>>>
>>>>>>> Hey Shawn,
>>>>>>>
>>>>>>> thank you for reporting this. It helps to relax the threshold for the
>>>>>>> reduction of the residual that the linear solver is required to
>>>>>>> achieve. I attach a corresponding input file, where the parameter is
>>>>>>> adjusted in line 31.
>>>>>>>
>>>>>>> To me, the default threshold of 1e-12 seems to be very small. We will
>>>>>>> discuss whether the default value should be adjusted.
>>>>>>>
>>>>>>> Kind regards
>>>>>>> Bernd
>>>>>>>
>>>>>>> On Fri, 23 Mar 2012 12:36:05 -0700 (PDT)
>>>>>>> Shawn Zhang <[email protected]> wrote:
>>>>>>>> Dear All,
>>>>>>>>
>>>>>>>> I ran into an issue of 1p model not converging. After a series
>>>>>>>> of tests, I trace it to this canonical situation,
>>>>>>>> For standard test/boxmodels/1p problem, if I increase the 2D grid
>>>>>>>> resolution from 10x10 to 100x100, the solver will not converge, with
>>>>>>>> the following error reported.
>>>>>>>>
>>>>>>>> This error, however, will go away if I reduce the perm difference
>>>>>>>> between lens and exterior from 2 order of magnitude into 1 order of
>>>>>>>> magnitude. This makes me speculate that the error is related to sharp
>>>>>>>> gradient at lens/exterior interface caused by the heterogeneity at
>>>>>>>> high resolution.
>>>>>>>>
>>>>>>>> I think the reproducer on standard test makes it likely a serious
>>>>>>>> problem, though I will be very happily corrected.
>>>>>>>>
>>>>>>>> Truly appreciate your quick attention.
>>>>>>>>
>>>>>>>> Best,
>>>>>>>> -Shawn
>>>>>>>>
>>>>>>>> root@szdebian:/home/szhang/dumux/dumux-2.1.0/test/boxmodels/1p#
>>>>>>>> ./test_1p -parameterFile test_1p.input
>>>>>>>> Welcome aboard DuMuX airlines. Please fasten your seatbelts! Emergency
>>>>>>>> exits are near the time integration.
>>>>>>>> Initializing problem '1ptest'
>>>>>>>> Writing result file for "1ptest"
>>>>>>>> Newton iteration 1 done, relative error = 0.666666
>>>>>>>> Newton iteration 2 done, relative error = 5.45083e-06
>>>>>>>> Solve: M deltax^k = rNewton: Caught exception: "NumericalProblem
>>>>>>>> [newtonSolveLinear:../../../dumux/nonlinear/newtoncontroller.hh:398]:
>>>>>>>> Linear solver did not converge"
>>>>>>>> Newton solver did not converge with dt=1 seconds. Retrying with time
>>>>>>>> step of 0.5 seconds
>>>>>>>> Newton iteration 1 done, relative error = 0.666666
>>>>>>>> Newton iteration 2 done, relative error = 5.45083e-06
>>>>>>>> Solve: M deltax^k = rNewton: Caught exception: "NumericalProblem
>>>>>>>> [newtonSolveLinear:../../../dumux/nonlinear/newtoncontroller.hh:398]:
>>>>>>>> Linear solver did not converge"
>>>>>>>>
>>>>>>>> ....
>>>>>>>> Newton solver did not converge with dt=0.0078125 seconds. Retrying
>>>>>>>> with time step of 0.00390625 seconds
>>>>>>>> Newton iteration 1 done, relative error = 0.666666
>>>>>>>> Newton iteration 2 done, relative error = 5.45083e-06
>>>>>>>> Solve: M deltax^k = rNewton: Caught exception: "NumericalProblem
>>>>>>>> [newtonSolveLinear:../../../dumux/nonlinear/newtoncontroller.hh:398]:
>>>>>>>> Linear solver did not converge"
>>>>>>>> Newton solver did not converge with dt=0.00390625 seconds. Retrying
>>>>>>>> with time step of 0.00195312 seconds
>>>>>>>> Newton iteration 1 done, relative error = 0.666666
>>>>>>>> Newton iteration 2 done, relative error = 5.45083e-06
>>>>>>>> Solve: M deltax^k = rNewton: Caught exception: "NumericalProblem
>>>>>>>> [newtonSolveLinear:../../../dumux/nonlinear/newtoncontroller.hh:398]:
>>>>>>>> Linear solver did not converge"
>>>>>>>> Newton solver did not converge with dt=0.00195312 seconds. Retrying
>>>>>>>> with time step of 0.000976562 seconds
>>>>>>>> Dune reported error: Dune::MathError
>>>>>>>> [timeIntegration:../../../dumux/boxmodels/common/boxproblem.hh:485]:
>>>>>>>> Newton solver didn't converge after 10 time-step divisions.
>>>>>>>> dt=0.000976562
>>>>>>>
>>>>>>> ___________________________________________________________
>>>>>>>
>>>>>>> Bernd Flemisch phone: +49 711 685 69162
>>>>>>> IWS, Universitaet Stuttgart fax: +49 711 685 67020
>>>>>>> Pfaffenwaldring 61 email: [email protected]
>>>>>>> D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
>>>>>>> ___________________________________________________________
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> Dumux mailing list
>>>>>>> [email protected]
>>>>>>> https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
>>>>>>
>>>>>> ___________________________________________________________
>>>>>>
>>>>>> Bernd Flemisch phone: +49 711 685 69162
>>>>>> IWS, Universitaet Stuttgart fax: +49 711 685 67020
>>>>>> Pfaffenwaldring 61 email: [email protected]
>>>>>> D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
>>>>>> ___________________________________________________________
>>>>>>
>>>>>> _______________________________________________
>>>>>> Dumux mailing list
>>>>>> [email protected]
>>>>>> https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
>>>>>>
>>>>>>
>>>>>>
>>>>>> permeability = 1e-10
>>>>>> permeabilityLens = 1e-12
>>>>>>
>>>>>> Sorry about the ignorance. I should notice this the first time. Can you
>>>>>> give a try and see whether you can reproduce?
>>>>>>
>>>>>> -Shawn
>>>>>
>>>>> ___________________________________________________________
>>>>>
>>>>> Bernd Flemisch phone: +49 711 685 69162
>>>>> IWS, Universitaet Stuttgart fax: +49 711 685 67020
>>>>> Pfaffenwaldring 61 email: [email protected]
>>>>> D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
>>>>> ___________________________________________________________
>>>>>
>>>>> _______________________________________________
>>>>> Dumux mailing list
>>>>> [email protected]
>>>>> https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
>>>>
>>>> ___________________________________________________________
>>>>
>>>> Bernd Flemisch phone: +49 711 685 69162
>>>> IWS, Universitaet Stuttgart fax: +49 711 685 67020
>>>> Pfaffenwaldring 61 email: [email protected]
>>>> D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
>>>> ___________________________________________________________
>>>>
>>>> _______________________________________________
>>>> Dumux mailing list
>>>> [email protected]
>>>> https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
>>>
>>> ___________________________________________________________
>>>
>>> Bernd Flemisch phone: +49 711 685 69162
>>> IWS, Universitaet Stuttgart fax: +49 711 685 67020
>>> Pfaffenwaldring 61 email: [email protected]
>>> D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
>>> ___________________________________________________________
>>
>> ___________________________________________________________
>>
>> Bernd Flemisch phone: +49 711 685 69162
>> IWS, Universitaet Stuttgart fax: +49 711 685 67020
>> Pfaffenwaldring 61 email: [email protected]
>> D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
>> ___________________________________________________________
>>
>> _______________________________________________
>> Dumux mailing list
>> [email protected]
>> https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
>
> ___________________________________________________________
>
> Bernd Flemisch phone: +49 711 685 69162
> IWS, Universitaet Stuttgart fax: +49 711 685 67020
> Pfaffenwaldring 61 email: [email protected]
> D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
> ___________________________________________________________
>
> _______________________________________________
> Dumux mailing list
> [email protected]
> https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
___________________________________________________________
Bernd Flemisch phone: +49 711 685 69162
IWS, Universitaet Stuttgart fax: +49 711 685 67020
Pfaffenwaldring 61 email: [email protected]
D-70569 Stuttgart url: www.hydrosys.uni-stuttgart.de
___________________________________________________________
_______________________________________________
Dumux mailing list
[email protected]
https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
_______________________________________________
Dumux mailing list
[email protected]
https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
