feelpp_mesh_partitioner(1)

1. NAME

feelpp_mesh_partitioner - a mesh partitioner for Feel++

2. SYNOPSIS

feelpp_mesh_partitioner --part num [num2…] --ifile file [--dim d] [--json file.json] [--shape s] [--order num] [--by-markers] [--by-markers-desc markers] [--ofile file] [--odir dir]

3. DESCRIPTION

feelpp_mesh_partitioner is a simple application which can generate a partitioned mesh and save it in a Feel++ specific json+hdf5 file format.

The generated mesh can then be loaded very efficiently in parallel.

3.1. Options

feelpp_mesh_partitioner requires some options.

Table 1. Table of command-line feelpp_mesh_partitioner options
Name	Description	Default value
dim	dimension of the mesh	3
shape	shape of the mesh elements	Simplex
order	mesh geometric order	1
by-markers	partitioning on each marker
by-markers-desc	partitioning by specifying the markers
part	number of desired partitions
ifile	name or path to the mesh
ofile	output filename prefix (without extension)
odir	output directory
json	json file that describes the partitioning

4. EXAMPLES

We are now using the feelpp/feelpp-toolboxes:latest docker images as described in book.feelpp.org/user-manual/#getting_started to demonstrate feelpp_mesh_partitioner usage. We use the meshes in src/feelpp/data/gmsh/primitives in the docker image.

Starting the docker feelpp/feelpp-toolboxes:latest

docker run --rm -ti -v $HOME/feel:/feel feelpp/feelpp-toolboxes

4.1. Generating a single mesh partitioning

We generate a mesh partitioned on 4 cores with the following command

feelpp_mesh_partitioner --part 4  --ifile src/feelpp/data/gmsh/primitives/torus.geo --ofile torus

You should have in the current directory 4 files

ls torus*
torus.geo  torus.h5  torus.json  torus.msh

Now the file torus.json can be loaded in a Feel++ application distributed on 4 cores.

4.2. Generating multiple mesh partitioning

Often we are interested in a set of partitioned meshes in order to do a speed-up study. feelpp_mesh_partitioner is the right tool for that. Let’s consider a set of partitions \(\mathcal{P}=\{2, 4, 8, 16, 32 \}\).

feelpp_mesh_partitioner --part 2 4 8 16 32  --ifile src/feelpp/data/gmsh/primitives/torus.geo --odir torus-mesh

You should have in the directory torus-mesh (thanks to the odir option) 5 partitioned meshes

ls torus*
torus_p16.h5    torus_p2.h5    torus_p32.h5    torus_p4.h5    torus_p8.h5
torus_p16.json  torus_p2.json  torus_p32.json  torus_p4.json  torus_p8.json

The mesh filenames contain the partition information.

If you want a different prefix than torus, like torus-coarse, you can use the --ifile torus-coarse option and you will have a set of files named torus-coarse_p*.* in the torus-mesh directory.

4.3. Generating a mesh partitioning by markers

1) partitioning of each marker present in the mesh :

feelpp_mesh_partitioner --ifile  toto.geo --by-markers --part 4 8 15 --odir hola

2) partitioning of marker1 and partitioning of the union of marker2 and marker3 :

feelpp_mesh_partitioner --ifile toto.geo --by-markers-desc marker1:marker2,marker3 --part 4 8 15  --odir hola

In by-markers-desc option, the , keyword create an union of the markers and : keyword allow to split the partitioning regions.

4.4. Aggregating elements of a mesh

For some numerical methods, it is necessary to aggregate elements that share faces marked with a specific marker. This is done to

avoid these faces are also interprocess faces
avoid having degrees of freedom split over different processes

To do so we use the --json option to specify the partitioning.

The following code will aggregate the elements of the spheres marked with CenterSphere, LeftSphere and RightSphere in the mesh on the same processor. The aggregation is done by adding weights to the elements sharing the same face marked with the same marker which then force the graph partitioner to put them on the same processor.

{
    "partitioner":
    {
        "aggregates":
        {
            "spheres":{
                "type": "faces",
                "markers": ["CenterSphere","LeftSphere","RightSphere"]
            }
        }
    }
}

5. SEE ALSO

Feel++ Book: book.feelpp.org