Parallel Mesh By Choice

[friedmud]

We need to configure libMesh with ParallelMesh always... and then optionally use "parallel = true" in the Mesh block to turn it on. We will default to SerialMesh.

Just don't use Mesh() to create the Mesh... use either SerialMesh or ParallelMesh depending on that option.

[jwpeterson]

In 4e282ef:

Make MooseMesh::_mesh protected and force clients to use the accessor.

Refs #2105.

[jwpeterson]

In 2e96c32:

Adding documentation in MooseMesh.

Refs #2105.

[jwpeterson]

In 92f695d:

Move function definitions to C file.

Refs #2105.

[jwpeterson]

In fe82ced:

MooseMesh::_mesh is now a pointer, in preparation for letting it
dynamically choose between being a SerialMesh or ParallelMesh.

Refs #2105.

[jwpeterson]

In 8b77942:

Consistently use getMesh().

That way we can insulate code from possible future changes to the type
of MooseMesh::_mesh.

Refs #2105.

[jwpeterson]

In :

Be explicit about the use of SerialMesh where necessary.

Refs #2105.

[jwpeterson]

In b676994:

Don't use LIBMESH_ENABLE_PARMESH to determine whether MooseMesh
holds a ParallelMesh.

MooseMesh will (eventually) be able to contain either a SerialMesh
_or_ a ParallelMesh regardless of the value of LIBMESH_ENABLE_PARMESH,
so we can't rely on that anymore.

Refs #2105.

[jwpeterson]

Since the plan is to use a MooseEnum {PARALLEL, SERIAL, DEFAULT} to read the user's input, we probably don't want to call it 'parallel' any more:

parallel = parallel

or even better

parallel = serial

is just crazy-pants. What do we want to call it? Right now I'm using

mesh_decomposition_type = parallel

But I'm willing to admit that name's a little cumbersome...

[YaqiWang]

or 'distribute_mesh = true/false'

[aeslaughter]

How about something simple: mesh_type

mesh_type = parallel

mesh_type = serial

mesh_type = default

[jwpeterson]

Andrew, I first thought about 'mesh_type' but since we already have the 'type' parameter, I thought it might be too similar? If anyone else thinks mesh_type is fine I'd be happy to use that instead.

Yaqi, your suggestion is good too, but we do need something more than just a boolean type. The 'default' mesh_type means that the input file will work with either SerialMesh or ParallelMesh, and the user can override it by specifying --parallel-mesh on the command line.

[andrsd]

Why not to use just: decomposition = {default | parallel | serial} ?

[friedmud]

I like something close to what Yaqi said:

parallel_distribution = { default | serial | parallel }

or maybe just

distribution = { default | serial | parallel }

"decomposition" is getting too close to "partitioning" I could see people getting confused by that.

[jwpeterson]

OK going with this

distribution = { default | serial | parallel }

[jwpeterson]

In 354392c:

Picking a better name for _is_parallel (it really means _is_nemesis).

The accessor is removed, but it could always be readded.

'nemesis' is a parameter on MooseMesh instead of FileMesh.

Reformatting long doc strings onto separate lines.

Adding new command line option to MOOSE: --parallel-mesh.

MooseMesh now uses the MooseApp object to determine if
--parallel-mesh has been specified on the command line.

MooseMesh reads an enumeration for the type of mesh "distribution" to use.

Note: if the user does not specify anything for 'distribution' in the
Mesh block, it is possible to set it to parallel via the
--parallel-mesh command line argument, otherwise it defaults to
serial.  This command line arg is not allowed to override the user if
he does specify 'distribution = serial' in the Mesh block.

Refs #2105.

[jwpeterson]

In d9fb1d4:

Nemesis output can be performed even if the underlying
libMesh mesh is a SerialMesh.

Removing an expect_err test and modifying another so
that it runs for serial meshes as well.

Refs #2105.

[jwpeterson]

In 37fbf05:

Transitioning from mesh_mode = 'SERIAL' (which is no longer sufficient)
to adding 'distribution = serial' in Periodic BCs tests.

Periodic BCs do not currently work with ParallelMesh on more than one
processor, but most of these input files will actually work with
ParallelMesh enabled if you run on 1 processor...

Refs #2105.
Refs #2119.

[jwpeterson]

In 4b3f7b1:

Transitioning from mesh_mode = 'SERIAL' (which is no longer sufficient)
to adding 'distribution = serial' in all "geomsearch" tests.

The geometric search stuff does not currently work with ParallelMesh on
more than one processor, but most of these input files will actually
work with ParallelMesh enabled if you run on 1 processor...

Refs #2105.
Refs #2121.

[jwpeterson]

In 3c91442:

ExodusII_IO::write_element_data() is causing a large number of tests
to fail in parallel with ParallelMesh enabled.  If we can fix that in
libmesh, we can hopefully remove the 'distribution = serial' line from
all the input files in this commit.

Refs #2105.
Refs #2122.

[jwpeterson]

In 4da24ee:

This test no longer needs to be marked with mesh_mode = 'PARALLEL'
since nemesis=true in the Mesh block implies the use of ParallelMesh.

Refs #2105.

[jwpeterson]

In 6e81a40:

This test has to use SerialMesh because stitch_meshes() is only implemented for SerialMesh.

Note that running:

../../../moose_test-opt -i tiled_mesh_test.i --mesh-only --parallel-mesh

Produces the error:

 Exiting without writing output
 ERROR: Unrecognized file extension: --parallel-mesh

So I think there's something wrong with our command line argument processing...

Refs #2105.
Refs #2124.

[jwpeterson]

In c38f4ef:

The test_parmesh_check is no longer valid, so I removed it from
the chain of prerequisites.

Refs #2105.

[jwpeterson]

In a1f8022:

DiracKernels don't seem to work currently in parallel with ParallelMesh.

The tests in this commit currently segfault when run in parallel with
ParallelMesh, and are therefore restricted to run with SerialMesh.
For details, see #2125.

Refs #2105.
Refs #2125.

[jwpeterson]

In c977ec6:

Marking multiapp transfer stuff as serial-only for now.

Refs #2105.
Refs #2126.

[jwpeterson]

In c1a7def:

Tests using MeshModifiers are restricted to using SerialMesh.
For more information, see #2129.

Refs #2105.
Refs #2129.

[jwpeterson]

In 01cc5ad:

Modifying the Gmsh test slightly.

It now only tests reading a mesh in Gmsh format and writing it back
out.  Should also work with ParallelMesh enabled.

Refs #2105.

[jwpeterson]

In a4d80bd:

The mixed-dimension tests are restricted to run with SerialMesh.
For more information, see #2130.

Refs #2105.
Refs #2130.

[jwpeterson]

In 938e89f:

Make the all_print_pps test run with SerialMesh only.

Apparently just changing the mesh type from SerialMesh to ParallelMesh
is enough to make it do a different number of residual evaluations...

Refs #2105.