Hi Birger and Tri Dat,
are you not getting any convergence at all or do you get convergence at
a smaller time step size?
Do you get the error in the first Newton step of a time step or in
between somewhere?
I have similar problems for parallel computations with bad aspect ratios
(300:10). Additionally I have horizontal layers with large differences
in permeability. I am using AMG (also with SuperLU as coarse solver) and
the fully implicit box model (1p2c).
Running the same grid sequentially leads to much larger time step sizes
until my time step is constrained by the linear solver.
I also ran the my grid for a homogeneous case with simplified physics
which the linear solver was able to handle.
I assume that the problem in my case is a combination of heterogeneities
and the bad aspect ratio.
Unfortunately I was not able to resolve the problem. Some actions that
slightly improved the behavior were:
(1) In amgproperties.hh setting the Preconditioner from Dune::SeqSSOR to
Dune::SeqILUn (This also improved convergence behavior of the linear
solver for the sequential case)
(2) If you use Dune ALUGrud. Try setting METIS_PartGraphRecursive in
your alugrid.cfg file. Another point I observed is that the partitioning
method played a huge role. Metis worked the best for me.
(3) In amgbackend.hh increasing the coarsen target to a value larger
than 2000 (for example 10000, 50000 or 100000). In the following line:
Dune::Amg::Parameters params(15,2000,1.2,1.6,Dune::Amg::atOnceAccu);
You can see the effect of the latter when you set the linear solver
verbosity to 2 in your input file:
[LinearSolver]
Verbosity = 2
AMG will print the different agglomeration levels for each Newton step.
If you increase the coarsen target there should be fewer levels. On the
highest level the coarse solver (SuperLU) is used. The convergence in my
case improved with increasing number of unknowns in the highest level.
The downside to this is that it takes much more time to solve the coarse
linear system with SuperLU which for me meant that I was much faster in
the sequential case than in the parallel one in the end. But maybe you
have more luck.
Best regards
Alex
On 11/26/2015 04:05 PM, Birger Hagemann wrote:
Hi Tri Dat,
I had recently the same problem with parallel computation on a
realistic reservoir grid where also the horizontal dimensions are much
larger than the vertical. The error message was the same. Bernd gave
me the hint to install SuperLU. However, after installing SuperLU the
error was only changed to:
“Newton: Caught exception: "NumericalProblem
>>[newtonSolveLinear:…/dumux/dumux/nonlinear/newtoncontroller.hh:380]:
>>Linear solver did not converge"
Maybe installing SuperLU will solve your problem. My problem is also
still unsolved, thus, any hints are welcome. I am also using a fully
implicit cell-centered model. The same simulation is working when
started on only one processor and the same model is working in
parallel on more simple grids.
Kind regards
Birger
*Von:*Dumux [mailto:[email protected]] *Im
Auftrag von *Tri Dat NGO
*Gesendet:* Dienstag, 24. November 2015 20:07
*An:* DuMuX User Forum
*Betreff:* Re: [DuMuX] Convergence problem for 3D simulations (2p
cell-centered model, ALUGrid)
Sorry, I forgot to add my files.
Kind regards,
Tri Dat
2015-11-24 20:03 GMT+01:00 Tri Dat NGO <[email protected]
<mailto:[email protected]>>:
Hi Dumuxers,
I would like to run parallel simulations of CO2 injection into a 3D
reservoir of which the horizontal extents are much higher than the
vertical one: Lx/Lz = Ly/Lz >> 1 (e.g. a reservoir of 1000m x 1000m x
50 m). A 2p cell-centered model is used.
The domain is described on a cartesian grid of 100x100x100 (1E+6)
cells. The CO2 is injected into the domain at X=500m and Y = 500m by
using the Peaceman's well model (bhp). The pressure at the boundaries
perpendicular to the Y-axis is set to the initial hydrostatic
pressure. All others boundaries are no-flux Neumann boundaries. At the
first step, we consider, for simplicity, an homogeneous reservoir. You
can find in attachment the *.hh files of my test.
The simulation converges only if the reservoir is anisotropic (Kxx/Kzz
= Kyy/Kzz >>1), meaning that the medium is lowly permeable in
Z-direction. Nevertheless, when the domain is isotropic (Kxx = Kzz), a
problem occurs: Newton solver did not converge. The error message is
"MathError
[mgc:/home/share/soft/dumux-2.6/include/dune/istl/paamg/amg.hh:825]:
Coarse solver did not converge".
Moreover, the same test works correctly on isotropic but thicker
domains (e.g. 1000m x 1000m x 250 m, the grid is always of 100x100x100).
I don't understand why a same test works on a grid but not on another
grid. Is this related to the numerical scheme (fully-implicit,
cell-centered)?
Any help will be greatly appreciated!
Kind regards,
Tri Dat
_______________________________________________
Dumux mailing list
[email protected]
https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
--
Alexander Kissinger
Institut für Wasser- und Umweltsystemmodellierung
Lehrstuhl für Hydromechanik und Hydrosystemmodellierung
Pfaffenwaldring 61
D-70569 Stuttgart
Telefon: +49 (0) 711 685-64729
E-Mail: [email protected]
_______________________________________________
Dumux mailing list
[email protected]
https://listserv.uni-stuttgart.de/mailman/listinfo/dumux
