Upgrade to Gridap 1.6 issues

[amartinhuertas]

Hi @fverdugo, @santiagobadia,

FYI ... in PR #32, I ported GridapDistributed.jl to Gridap 0.16.3. Along the process, I annotated a set of concerns that you can find in the list below. Apart from solving this, I think it would be great to speak about which are the priorities/future plans with the development of GridapDistributed.jl.

Several concerns I have detected along the process.

• SparseMatrixAssembler in Gridap v0.16 does no longer conceptually hold test and trial FE spaces inside, while DistributedSparseMatrixAssembler does.
• Introduction of ArtificeSparseMatrixToLeverageDefaults. Something more clever to solve this issue?
• Review the need for the allocate_local_vector function and the workflow associated to it. Is there a better way to design this taking into account the way assembly operations have been redesigned in Gridap v0.16?
• I am skipping symbolic_loop_vector! all the way through.
• [Solved in cc56f4d] Bilinear and linear functions for multifield problems cannot be iterated over (i.e., tuples) as in Gridap.
• I had to recover Reindexed and some of the associated machinery from Gridap 0.14.3. I need it for GridapDistributed.jl. Should we bypass this need? If yes, how?
• Review DistributedFEFunctions and their current hierarchy.
• Gridap.Algebra.add_entry!: We are inserting entries one by one. To see how we can introduce all cell entries at once with the new assembly interface.

[amartinhuertas]

@fverdugo ... FYI ... I had a meeting yesterday with @santiagobadia related to this issue. We agreed to proceed with the following tasks (in order of priority):

1. [UPDATE PR Gridap distributed high level api #34] Explore a high-level API for GridapDistributed.jl that follows as much as possible the one of Gridap.jl.
2. Develop new specialization of GridapDistributed.jl types and functions (DistributedData, DistributedIndexSet, etc.) using PartitionedArrays.jl as back-end.
3. Implement a first prototype non-overlapping balancing domain decomposition preconditioner.
4. To study what is really needed from GridapDistributedPETScWrappers.jl and to include it into GridapPETSc.jl (less priority).

[fverdugo]

Sounds good. For the BDDC part, one needs to add some multi-level data structures in PartitionedArrays.

[amartinhuertas]

Sounds good. For the BDDC part, one needs to add some multi-level data structures in PartitionedArrays.

Ok. Perhaps when I am about to start this work (I am about to start point 2. above now) it might help to have a meeting so that we may exchange ideas in this regard.