[petsc-users] DMView and DMLoad

[Berend van Wachem]

Dear Vaclav, Lawrence,

following your example, we have managed to save the DM with a wrapped 
Vector in h5 format (PETSC_VIEWER_HDF5_PETSC) with:

     DMPlexTopologyView(dm, viewer);
     DMClone(dm, &sdm);
     ...
     DMPlexSectionView(dm, viewer, sdm);
     DMGetLocalVector(sdm, &vec);
     ...
     DMPlexLocalVectorView(dm, viewer, sdm, vec);

The problem comes with the loading of the "DM+Vec.h5" with:

     DMCreate(PETSC_COMM_WORLD, &dm);
     DMSetType(dm, DMPLEX);
     ...
     DMPlexTopologyLoad(dm, viewer, &sfO);
     ...
     PetscSFCompose(sfO, sfDist, &sf);
     ...
     DMClone(dm, &sdm);
     DMPlexSectionLoad(dm, viewer, sdm, sf, &globalDataSF, &localDataSF);
     DMGetLocalVector(sdm, &vec);
     ...
     DMPlexLocalVectorLoad(dm, H5Viewer, sdm, localDataSF, vec);

The loaded DM is different to the one created with DMPlexCreateFromfile 
(for instance, no "coordinates" are recovered with the use of 
DMGetCoordinatesLocal). This conflicts with our code, which relies on 
features of the DM as delivered by the DMPlexCreateFromfile function.

We have also noticed that the "DM+Vec.h5" can not be loaded directly 
with DMPlexCreateFromfile because it contains only the groups "topology" 
and "topologies" while the groups "geometry" and "labels" are missing 
(and probably other conflicts).

Is this something which can be changed? We would need to reload a DM 
similar to the one created with DMPlexCreateFromfile.

Best regards,

Berend.



On 9/22/21 8:59 PM, Hapla Vaclav wrote:
> To avoid confusions here, Berend seems to be specifically demanding XDMF 
> (PETSC_VIEWER_HDF5_XDMF). The stuff we are now working on is parallel 
> checkpointing in our own HDF5 format (PETSC_VIEWER_HDF5_PETSC), I will 
> make a series of MRs on this topic in the following days.
> 
> For XDMF, we are specifically missing the ability to write/load DMLabels 
> properly. XDMF uses specific cell-local numbering for faces for 
> specification of face sets, and face-local numbering for specification 
> of edge sets, which is not great wrt DMPlex design. And ParaView doesn't 
> show any of these properly so it's hard to debug. Matt, we should talk 
> about this soon.
> 
> Berend, for now, could you just load the mesh initially from XDMF and 
> then use our PETSC_VIEWER_HDF5_PETSC format for subsequent saving/loading?
> 
> Thanks,
> 
> Vaclav
> 
>> On 17 Sep 2021, at 15:46, Lawrence Mitchell <wence at gmx.li 
>> <mailto:wence at gmx.li>> wrote:
>>
>> Hi Berend,
>>
>>> On 14 Sep 2021, at 12:23, Matthew Knepley <knepley at gmail.com 
>>> <mailto:knepley at gmail.com>> wrote:
>>>
>>> On Tue, Sep 14, 2021 at 5:15 AM Berend van Wachem 
>>> <berend.vanwachem at ovgu.de <mailto:berend.vanwachem at ovgu.de>> wrote:
>>> Dear PETSc-team,
>>>
>>> We are trying to save and load distributed DMPlex and its associated
>>> physical fields (created with DMCreateGlobalVector)  (Uvelocity,
>>> VVelocity,  ...) in HDF5_XDMF format. To achieve this, we do the 
>>> following:
>>>
>>> 1) save in the same xdmf.h5 file:
>>> DMView( DM         , H5_XDMF_Viewer );
>>> VecView( UVelocity, H5_XDMF_Viewer );
>>>
>>> 2) load the dm:
>>> DMPlexCreateFromfile(PETSC_COMM_WORLD, Filename, PETSC_TRUE, DM);
>>>
>>> 3) load the physical field:
>>> VecLoad( UVelocity, H5_XDMF_Viewer );
>>>
>>> There are no errors in the execution, but the loaded DM is distributed
>>> differently to the original one, which results in the incorrect
>>> placement of the values of the physical fields (UVelocity etc.) in the
>>> domain.
>>>
>>> This approach is used to restart the simulation with the last saved DM.
>>> Is there something we are missing, or there exists alternative routes to
>>> this goal? Can we somehow get the IS of the redistribution, so we can
>>> re-distribute the vector data as well?
>>>
>>> Many thanks, best regards,
>>>
>>> Hi Berend,
>>>
>>> We are in the midst of rewriting this. We want to support saving 
>>> multiple meshes, with fields attached to each,
>>> and preserving the discretization (section) information, and allowing 
>>> us to load up on a different number of
>>> processes. We plan to be done by October. Vaclav and I are doing this 
>>> in collaboration with Koki Sagiyama,
>>> David Ham, and Lawrence Mitchell from the Firedrake team.
>>
>> The core load/save cycle functionality is now in PETSc main. So if 
>> you're using main rather than a release, you can get access to it now. 
>> This section of the manual shows an example of how to do 
>> thingshttps://petsc.org/main/docs/manual/dmplex/#saving-and-loading-data-with-hdf5 
>> <https://petsc.org/main/docs/manual/dmplex/#saving-and-loading-data-with-hdf5>
>>
>> Let us know if things aren't clear!
>>
>> Thanks,
>>
>> Lawrence
>