Dear PETSc team: Greetings. I keep working on mesh I/O utilities using DMPlex. Specifically for the output stage, I need a solid grasp on the global numbers and ideally how to set them into the DMPlex during an input operation and carrying
ZjQcmQRYFpfptBannerStart
This Message Is From an External Sender
This message came from outside your organization.
ZjQcmQRYFpfptBannerEnd
Dear PETSc team:
Greetings.
I keep working on mesh I/O utilities using DMPlex.
Specifically for the output stage, I need a solid grasp on the global numbers and ideally how to set them into the DMPlex during an input operation and carrying the global numbers through API calls to DMPlexDistribute() or DMPlexMigrate() and hopefully also
through some of the mesh adaption APIs. I was wondering if the GlobalToNatural PetscSF manages these global numbers. The next most useful object is the PointSF, but to me, it seems to only help establish DAG point ownership, not DAG point global indices.
This is a good question, and gets at a design point of Plex. I don't believe global numbers are the "right" way to talk about mesh points, or
even a very useful way to do it, for several reasons. Plex is designed to run just fine without any global numbers. It can, of course, produce
them on command, as many people remain committed to their existence.
Thus, the first idea is that global numbers should not be stored, since they can always be created on command very cheaply. It is much more
costly to write global numbers to disk, or pull them through memory, than compute them.
The second idea is that we use a combination of local numbers, namely (rank, point num) pairs, and PetscSF objects to establish sharing relations for parallel meshes. Global numbering is a particular traversal of a mesh, running over the locally owned parts of each mesh in local order. Thus an SF + a local order = a global order, and the local order is provided by the point numbering.
The third idea is that a "natural" order is just the global order in which a mesh is first fed to Plex. When I redistribute and reorder for good performance, I keep track of a PetscSF that can map the mesh back to the original order in which it was provided. I see this as an unneeded expense, but many many people want output written in the original order (mostly because processing tools are so poor). This management is what we mean by GlobalToNatural.
Otherwise, I have been working with the IS obtained from DMPlexGetPointNumbering() and manually determining global stratum sizes, offsets, and numbers by looking at the signs of the involuted index list that comes with that IS. It's working for now (I can monolithically
write meshes to CGNS in parallel), but it is resulting in repetitive code that I will need for another mesh format that I want to support.
What is repetitive? It should be able to be automated.
Thanks,
Matt
Sincerely:
J.A. Ferrand
Embry-Riddle Aeronautical University - Daytona Beach - FL
Ph.D. Candidate, Aerospace Engineering
M.Sc. Aerospace Engineering
B.Sc. Aerospace Engineering
B.Sc. Computational Mathematics
Phone: (386)-843-1829
Email(s): ferranj2@my.erau.edu
jesus.ferrand@gmail.com