Bug: Triangulation::load()/save() with large output files

[tjhei]

We have a bug reported for ASPECT related to Triangulation::load/save that got introduced between 9.2 and 9.3 (see 1). After a quick look into tria_base.cc, I am sure that we will run into issues when the number of cells and/or the total file size runs above 32bits. I have not tested this, but this is suspicious (we need to use types::global_cell_index or MPI_Offset):

dealii/source/distributed/tria_base.cc

Lines 1606 to 1607 in c090a8c

	const unsigned int offset =
	sizes_fixed_cumulative.size() * sizeof(unsigned int);

dealii/source/distributed/tria_base.cc

Line 1656 in c090a8c

const unsigned int offset = global_num_cells * sizeof(unsigned int);

dealii/source/distributed/tria_base.cc

Line 1383 in c090a8c

const unsigned int global_num_cells,

dealii/source/distributed/tria_base.cc

Line 1250 in c090a8c

unsigned int offset = numbers::invalid_unsigned_int;

I see the following issues:

• fixed data transfer for >4GB files
• data transfer for >2GB per process, see fix checkpointing with >2GB per process #12973
• more than 2^32 cells
• more than 2^32 DoFs (check that this works)
• variable data transfer (similar to above) variable checkpointing fixes for >4GB files #12832
• p::fd::T, LargeCount MPI_write/read_at for p::fd:Tria #13867
• MPI IO file write with >2GB individual writes #12873

[marcfehling]

The code snippets you mentioned were already there in version 9.2.0, but at a different spot. Using the v9.2.0 tag for example

dealii/source/distributed/tria.cc

Lines 1989 to 1990 in b42193f

	const unsigned int offset =
	sizes_fixed_cumulative.size() * sizeof(unsigned int);

The entire DataTransfer class has been moved from p:d:Tria to p:DistributedTriangulationBase to be used in combination with p:fd:Tria. DataTransfer has not been changed between releases AFAIK. I would bet that the error might be somewhere else, but I don't have an idea right now :(

---

Anyways, I agree that we should use a different type here.

[bangerth]

At least the first of these code snippets appears correct to me since it pertains to per-cell data. I agree that everything that invokes global_num_cells looks suspicious and should probably be types::global_cell_index.

[tjhei]

At least the first of these code snippets appears correct to me since it pertains to per-cell data. I

I disagree. This is an offset into the data file and is used in the MPI IO call where the type is MPI_Offset (64 bit). If your file is bigger than 4 GB, this will fail.

[tjhei]

@marcfehling I have the fix for the variable transfer ready to go as well. Do I see this right, that we do not test this anywhere?