BUG: makeFaMesh fails with the random decomposition method

Summary

makeFaMesh fails when the random decomposition method is used.

Steps to reproduce

• cd $FOAM_TUTORIALS/finiteArea/liquidFilmFoam/cylinder
• change method simple; to method random; in system/decomposeParDict
• ./Allrun-parallel

What is the current bug behaviour?

Floating point exception.

Environment information

    base1    = develop
    api      = 2106
    patch    = 0
    HEAD     = 70c697fdcb
    version  = com
    compiler = Clang (system)
             = clang version 9.0.1 
    mpi      = SYSTEMOPENMPI
             = mpirun (Open MPI) 1.10.7.0.5e373bf1fd
    OS       = Description:     openSUSE Leap 15.1
    opts     = linux64ClangDPInt32Opt

added bug finiteArea parallel labels

assigned to @kuti

bug or known current restriction? See that sketch on #2084 (closed)

with random decomposition I highly suspect that you have a finite-area that has the fan-like connection.

Sounds like we may need to revisit this topic

Adding to the discussion about makeFaMesh failing in parallel in the fan-like connections mentioned by @mark:

Though this is a known limitation, it seems to appear quite often. Using the attached Ahmed body case, Ahmed-FaMesh.tar.gz, an error of the following type

[4] --> FOAM FATAL ERROR: (openfoam-2106)
[4] Cannot handle edge on processor-processor connection at (0 93.7534 -428.554) (0 95.0822 -425.704)
[4] 
[4] 
[4]     From Foam::List<Foam::LabelledItem<Foam::edge> > Foam::faMesh::getBoundaryEdgePatchPairs() const
[4]     in file faMesh/faMeshPatches.C at line 624.
[4] 
FOAM parallel run aborting

appears using the following decompositions

hierarchical: coeffs (4 2 2) (5 4 2), scotch and kahip with 40 domains

Most probably, it will appear with other decompositions including a moderate number of processors and above.

When the number of processors is small (e.g. hierarchical decomposition with (2 1 1)) and connections like these are less likely to appear, everything seems fine.

Notes:

1. Just to save the time of snappyHexMesh in each decomposition setup, the Allrun script uses reconstructParMesh at the end to get the reconstructed mesh and re-decomposes with other decompositions setups.

2. I couldn't compress the stl to a size below the 10MB threshold allowed by the attachment, so the Allrun scripts downloads it.

Hi @vaggelisp - found some time to pick this up again. It is indeed a processor-processor connection via an edge, which is exactly what normally doesn't happen ... except that it does.

Will be rejigging in the next bit.

unassigned @kuti

assigned to @mark

mentioned in commit f3a1137d

mentioned in merge request !490 (merged)

mentioned in commit 6730c727

mentioned in issue #2237 (closed)

mentioned in commit 145cad56

mentioned in commit 8e451089

confirmed that the first error has been vanished - thanks.

closed