Portable refined triangulation format

[agrayver]

As time goes, I really think one should have a portable way to save/load refined triangulations. This was touched in #10447 and it was made clear the binary format is not compatible across versions/machines. It is also not compatible between different versions of boost (serialization) library.

Is there a portable (i.e. independent of deal.II/boost versions) way to save a refined triangulation with all the associated information (boundary id, manifold, material properties) and restore it later?

[tjhei]

This is a reasonable request, but not trivial to implement. I don't think there is a way to do this right now.

A few things to consider:

• Does this need to work in parallel? (I assume yes)
• Do you want to store parallel partitioning as well?
• Does this need to work when loading with a different number of MPI ranks? (Again, useful, but makes it more complicated)

If you want to support p::d::Triangulation, the code will need to start with the coarse mesh and execute refinement steps (otherwise we won't be able to end up with the correct p4est). If you want to support parallel computations, I would strongly recommend a single large file with MPI IO (or something like HDF5 as the container).

This would be quite useful to have. For example: one could partition a mesh offline before running a simulation with a fully distributed mesh.

[bangerth]

For the non-parallel kind, there is PersistantTriangulation. For the parallel case, we generally just use the serialization approach. Why don't you like that?

[agrayver]

For the parallel case, we generally just use the serialization approach. Why don't you like that?

Sorry for the late response. I do like it and this is what we use. My problem is that these serialized objects are only usable with the deal.II version they were created with.

For instance, I have run (parallel) simulations with deal.II 9.2. Meanwhile, the code was updated to deal.II 9.3 to use some new functionality, and thus it cannot read older serialized triangulations anymore.

I do not even know how I could potentially convert v9.2 serialized objects to v9.3 other than rerunning all simulations from scratch (which is not feasible).