[petsc-users] DMPlex Parallel Output and DMPlexCreateSection Crashes when DMPlexCreateGmshFromFile used
Matthew Knepley
knepley at gmail.com
Tue Apr 26 18:33:15 CDT 2022
On Tue, Apr 26, 2022 at 7:27 PM Mike Michell <mi.mike1021 at gmail.com> wrote:
> Thanks for the answers. One last question related to designing my code.
> What I want to do is to build a finite-volume code discretized on "median
> dual" unstructured mesh. So basically it will use a vertex-centered
> discretization scheme by looping over "median dual faces", which is not
> physically existing in my mesh file.
>
> I can acheive this by allocating some vectors to compute required
> geometrical metrics (e.g., face normal vectors, face area, etc. that
> required to integrate over median dual control volume) as I did before even
> without PETSc. However, I think this definitely cannot be an optimal way,
> and I believe there should be much larger benefit that can be obtained, if
> I well design my DMPlex's data structure (e.g., PetscSection).
> Is there any example with this median dual control volume case? or Any
> guideline will be helpful.
> I was thinking to borrow some functions designed for finite element.
>
I have never done a dual scheme before, so let me ask some questions to
better understand the underlying ideas.
1) Is there a reason not to store the dual directly? That seems like the
easiest approach.
2) You can use Plex to layout auxiliary geometric data. I do it the
following way:
DMClone(dm, dmGeom);
<Create the Section s you want>
DMSetLocalSection(dmGeom, s);
PetscSectionDestroy(&s);
then
DMCreateLocalVector(dmGeom, &lv);
will give you a Vec with the local data in it that can be addressed by mesh
point (global vectors too). Also you would be able
to communicate this data if the mesh is redistributed, or replicate this
data if you overlap cells in parallel.
Thanks,
Matt
> Thanks,
> Mike
>
> 2022년 4월 26일 (화) 오후 5:40, Matthew Knepley <knepley at gmail.com>님이 작성:
>
>> On Tue, Apr 26, 2022 at 4:48 PM Mike Michell <mi.mike1021 at gmail.com>
>> wrote:
>>
>>> Below two interesting things are found:
>>> - If DMPlexCreateGmshFromFile() is used, DMPlexDistribute() is not done
>>> by default. I should call DMPlexDistribute() to distribute DMPlex over the
>>> procs.
>>>
>>
>> Here is the explanation. The default distribution comes from
>> DMSetFromOptions(), which I am guessing was not called after
>> DMPlexCreateGmshFromFile().
>>
>>
>>> - If {DMCreate(), DMSetType(), DMSetFromOptions()} in serially used with
>>> "-dm_plex_filename ./new.msh" option in command line, DMPlexDistribute() is
>>> done by default even though DMPlex object is created by reading the same
>>> .msh file.
>>>
>>> Thanks for the modification to the example ex1f90.F90, now it works with
>>> 2 procs.
>>> But still, if I print my output to "sol.vtk", the file has only a part
>>> of whole mesh. However, the file is okay if I print to "sol.vtu".
>>> From "sol.vtu" I can see the entire field with rank. Is using .vtu
>>> format preferred by petsc?
>>>
>>
>> VTK is generally for debugging, but it should work. I will take a look.
>>
>> VTU and HDF5 are the preferred formats.
>>
>> Thanks,
>>
>> Matt
>>
>>
>>> 2022년 4월 26일 (화) 오후 2:26, Matthew Knepley <knepley at gmail.com>님이 작성:
>>>
>>>> On Tue, Apr 26, 2022 at 9:33 AM Mike Michell <mi.mike1021 at gmail.com>
>>>> wrote:
>>>>
>>>>> Thank you for the answers.
>>>>> For the first question, basically, I cannot
>>>>> run "/dm/impls/plex/ex1f90.F90" example with more than 1 proc. I removed
>>>>> DMPlexDistribute() following your comment and what I tried is:
>>>>>
>>>>> - no modification to ex1f90.F90 (as it is)
>>>>> - make "ex1f90"
>>>>> - mpirun -np 2 ./ex1f90
>>>>>
>>>>> It gives me "Bad termination of one of ..." for Rank 1. The code runs
>>>>> okay with "mpirun -np 1 ./ex1f90".
>>>>>
>>>>
>>>> You are correct. It evidently never worked in parallel, since those
>>>> checks will only work in serial.
>>>> I have fixed the code, and added a parallel test. I have attached the
>>>> new file, but it is also in this MR:
>>>>
>>>> https://gitlab.com/petsc/petsc/-/merge_requests/5173
>>>>
>>>> Thanks,
>>>>
>>>> Matt
>>>>
>>>>
>>>>> Thanks,
>>>>> Mike
>>>>>
>>>>> 2022년 4월 26일 (화) 오전 3:56, Matthew Knepley <knepley at gmail.com>님이 작성:
>>>>>
>>>>>> On Mon, Apr 25, 2022 at 9:41 PM Mike Michell <mi.mike1021 at gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>>> Dear PETSc developer team,
>>>>>>>
>>>>>>> I'm trying to learn DMPlex to build a parallel finite volume code in
>>>>>>> 2D & 3D. More specifically, I want to read a grid from .msh file by Gmsh.
>>>>>>> For practice, I modified /dm/impls/plex/ex1f90.F90 case to read &
>>>>>>> distribute my sample 2D grid, which is attached. I have two questions as
>>>>>>> below:
>>>>>>>
>>>>>>> (1) First, if I do not use my grid, but use the default box grid
>>>>>>> built by ex1f90.F90 and if I try "DMPlexDistribute" over mpi processors,
>>>>>>> the output file (sol.vtk) has only some portion of the entire mesh. How can
>>>>>>> I print out the entire thing into a single file? Is there any example for
>>>>>>> parallel output? (Related files attached to "/Question_1/") Paraview gives
>>>>>>> me some error messages about data size mismatching.
>>>>>>>
>>>>>>
>>>>>> For the last release, we made parallel distribution the default.
>>>>>> Thus, you do not need to call DMPlexDIstribute() explicitly here. Taking it
>>>>>> out, I can run your example.
>>>>>>
>>>>>>
>>>>>>> (2) If I create DMPlex object through "DMPlexCreateGmshFromFile",
>>>>>>> "DMPlexCreateSection" part is crashed. I do not understand why my example
>>>>>>> code does not work, because the only change was switching from "DMCreate"
>>>>>>> to "DMPlexCreateGmshFromFile" and providing "new.msh" file. Without the
>>>>>>> PetscSection object, the code works fine. Any comments about this? (Related
>>>>>>> files attached to "/Question_2/")
>>>>>>>
>>>>>>
>>>>>> If I remove DMPlexDistribute() from this code, it is clear that the
>>>>>> problem is with the "marker" label. We do not create this by default from
>>>>>> GMsh since we assume people have defined their own labels. You can pass
>>>>>>
>>>>>> -dm_plex_gmsh_use_marker
>>>>>>
>>>>>> to your code. WHen I do this, you example runs for me.
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Matt
>>>>>>
>>>>>>
>>>>>>> Thanks,
>>>>>>> Mike
>>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> What most experimenters take for granted before they begin their
>>>>>> experiments is infinitely more interesting than any results to which their
>>>>>> experiments lead.
>>>>>> -- Norbert Wiener
>>>>>>
>>>>>> https://www.cse.buffalo.edu/~knepley/
>>>>>> <http://www.cse.buffalo.edu/~knepley/>
>>>>>>
>>>>>
>>>>
>>>> --
>>>> What most experimenters take for granted before they begin their
>>>> experiments is infinitely more interesting than any results to which their
>>>> experiments lead.
>>>> -- Norbert Wiener
>>>>
>>>> https://www.cse.buffalo.edu/~knepley/
>>>> <http://www.cse.buffalo.edu/~knepley/>
>>>>
>>>
>>
>> --
>> What most experimenters take for granted before they begin their
>> experiments is infinitely more interesting than any results to which their
>> experiments lead.
>> -- Norbert Wiener
>>
>> https://www.cse.buffalo.edu/~knepley/
>> <http://www.cse.buffalo.edu/~knepley/>
>>
>
--
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20220426/282ca98f/attachment.html>
More information about the petsc-users
mailing list