Misc pending tasks associated with refactoring in branch release-0.2

[fverdugo]

Merging into master PR #50

• Update README
• remove src/OLD, test/OLD, compile
• Add NEWS.md

High priority

• Release Gridap 0.17.0
• Release PartitonedArrays 0.2.3
• Write data into pvtu format (@amartinhuertas) PR Support parallel file formats WriteVTK.jl#1 PVTU files #46
• Create MPI launchers for the test (now we just test in serial mode) @fverdugo Adding mpi tests #45
• Assembly of the rhs vector alone (@amartinhuertas) PR Rhs assembly alone #40
• Solve broken tests in FESpacesTests.jl namely assemble_XXX! functions do not work @fverdugo Nonlinear test (and misc associated fixes) #44
• Implement assembly strategy when one also iterates over ghost cells.
• Interpolation onto distributed multi-field spaces. @fverdugo PR Multi field interpolation and DG test #42
• Port uniformly refined forest of octrees distributed model this new version (in GridapP4est) @amartinhuertas
• Add test drivers including DG examples. @fverdugo PR Multi field interpolation and DG test #42
• Add test drivers including nonlinear examples. @fverdugo PR Nonlinear test (and misc associated fixes) #44
• Do not use MPI.COMM_WORLD in get_part_ids(b::MPIBackend,... PartitionedArrays/PartitionedArrays.jl#33 Duplicating communicator in MPI backend PartitionedArrays/PartitionedArrays.jl#35
• @asserts locally triggered from the processors may lead to deadlock! We should call mpi_abort whenever an AssertionError exception is triggered. PR Improving prun() for MPIBackend. Adding prun_debug() for debugging purposes. PartitionedArrays/PartitionedArrays.jl#39
• AD in parallel. Talk to @amartinhuertas to grasp the challenges there.

Medium priority

• Nicer user API for PETScSolver (to be done in GridapPETSc). In particular, better handling of the lifetime of PETSc objects.
• Implement distributed models from gmsh (in GridapGmsh) @fverdugo
• implement lu! for PSparseMatrix following the same strategy as with \
• Move mpi tests to their own CI job @amartinhuertas PR Separating Seq and MPI tests in Github actions #47
• Automatically generate sysimage to reduce mpi tests runtime @amartinhuertas PR Adding sysimage generation to mpi tests #48
• Think how the user can define the local and global vector type in the FESpace
• Showing NLSolve.jl solver trace only in master MPI rank leads to a dead lock.
• Peridodic BCs
• Implement ZeroMeanFESpace

Low priority

• Implement a more lazy initialization of the matrix exchanger in PartitionedArrasys since the exchanger is not always needed.
• Overlap compression of the sparse matrix with communication of rhs assembly
• Implement another strategy to represent local matrices in PartitionedArrays with a data layout compatible with petsc
• interpolate_everywhere not available in GridapDistributed.jl, only interpolate Solved in PR Dirichlet values #74

[amartinhuertas]

Let me document here a spare set of issues while I go through GridapDistributed.jl (to not forget, these are temporary, I may add new items on the fly, etc.):

• [Minor Solved in PR https://github.com/Rhs assembly alone #40] DistributedSparseMatrixAssembler does not have strategy member variable, and I think it should, as required by the get_assembly_strategy function (among others).
• [Important] If you use FullyAssembledRows() strategy in DistributedSparseMatrixAssembler but do not pass the ghost cells in the assembly raw data, the code does not complain. Conversely if one uses SubAssembledRows() + raw data with ghost cells. I think this is VERY dangerous and prone to errors. We should try to find a solution.
• Why the assemble objects are called Allocation and NOT Allocator ?
• While the function local_views in GridapDistributed.jl is NOT actually called local_view. Why in plural?
• More issues to come ...

[fverdugo]

Hi @amartinhuertas!

I have opened a new branch https://github.com/gridap/GridapDistributed.jl/tree/release-0.2 from the current state.

We can use this branch as code that "is working" and do intermediate devs somewhere else.

[fverdugo]

@amartinhuertas I have classified the pending tasks as high/medium/low priority. (Move tasks from one tier not another if you find it necessary)

For me, we can merge the branch once the high-priority test are fixed. The missing low/medium priority ones can be moved to separate issues.

[amartinhuertas]

For me, we can merge the branch once the high-priority test are fixed. The missing low/medium priority ones can be moved to separate issues.

Ok. I have added a new high-priority task. Agreed.

[fverdugo]

@amartinhuertas I am adding a nonlinear test and I have found a bug associated with the in-place assembly functions assemble_xxx!. I am fixing it.

[amartinhuertas]

@amartinhuertas I am adding a nonlinear test and I have found a bug associated with the in-place assembly functions assemble_xxx!. I am fixing it.

Yes, indeed I added @test_broken macro calls in FESpacesTests.j in relation to these. You will find this in the rhs_assembly_branch. When you fix it, you can replace these by @test.

[fverdugo]

Hi @amartinhuertas

I have added the nonlinear example. In particular,

• the in-place assembly functions are working.
• one can use the parallel assembly strategy to build a valid triangulation e.g. Triangulation(with_ghost,model) and Triangulation(FullyAssembledRows(),model) are equivalent.

[oriolcg]

Hi @amartinhuertas, @fverdugo, what are the developments needed for periodic BCs?

[fverdugo]

At least, the ghost layer has to be implemented taking into account periodicity. And perhaps other devs.

[oriolcg]

Hi @amartinhuertas and @fverdugo, the last task of the list can be checked out once PR #74 is accepted. Also, I suggest to add another task into the list: support ZeroMeanFESpace. I guess that this is only a matter of selecting only one local space to constrain a DOF and use the global volume instead of the local volume as a member of the struct. What do you think about this?

[amartinhuertas]

the last task of the list can be checked out once PR #74 is accepted.

Great @oriolcg ! Thanks for your work!

support ZeroMeanFESpace

Sure. Task added.

I guess that this is only a matter of selecting only one local space to constrain a DOF and use the global volume instead of the local volume as a member of the struct. What do you think about this?

I indeed implemented this in GridapDistributed.jl v0.1.0. See https://github.com/gridap/GridapDistributed.jl/blob/v0.1.0/src/ZeroMeanDistributedFESpaces.jl and https://github.com/gridap/GridapDistributed.jl/blob/v0.1.0/test/ZeroMeanDistributedFESpacesTests.jl. I guess it should be a matter to re-write this using the new code organization in v0.2.0.