XH5For

The Fortran way to easy write parallel partitioned XDMF/HDF5 meshes

License

What is XH5For?

Is a library to read and write parallel partitioned FEM meshes taking advantage of the Collective/Independent IO provided by the HDF5 library. XH5For is not a general-purpose XDMF library, it only reads XDMF files written by itself.

Some concepts

• XDMF Model and Format is a standarized way to describe common scientific data. It separates the Light data (XML description of the data model) and the Heavy data (raw values) usually saved in HDF5 or binary Files.
• HDF5 is a model and data format designed for efficient Collective/Independent I/O and work with big amounts of data.

How to get XH5For

git clone --recursive https://github.com/victorsndvg/XH5For.git

Compilation

XH5For compile with GNU Fortran compiler 5.1 (and newer versions) and Intel Fortran compiler 15.0.1 (and newer versions).

XH5For uses CMake as a portable compilation system.

The easiest way to compile XH5For under Linux is:

$ cd XH5For
$ mkdir build
$ cd build
$ cmake [cmake-settings] ../
$ make

To compile XH5For under Windows use de equivalent commands

XH5For CMake basic settings

CMake compilation system settings allows us to manage compilation system behaviour. Look at the CMake command line-interface to know how to pass CMake settings. There are several ways to configure XH5For compilation system with CMake

The most important XH5For settings are:

• XH5For_ENABLE_HDF5: ON if HDF5 library is automagically found in your system (default OFF)
• XH5For_ENABLE_MPI: ON if MPI library is automagically found in your system (default OFF)
• XH5For_ENABLE_TESTS: (default OFF)
• XH5For_ENABLE_EXAMPLES: (default OFF)

Getting started with XH5For

The code below writes an hexahedron per MPI task:

program xh5for_hexa_per_task

    use xh5for
    use mpi

implicit none

    type(xh5for_t)        :: xh5
    real, dimension(24)   :: geometry = (/0.0, 0.0, 0.0, &
                                          0.0, 0.0, 1.0, &
                                          0.0, 1.0, 1.0, &
                                          0.0, 1.0, 0.0, &
                                          1.0, 0.0, 0.0, &
                                          1.0, 0.0, 1.0, &
                                          1.0, 1.0, 1.0, &
                                          1.0, 1.0, 0.0/)
    integer, dimension(8) :: topology = (/0, 1, 2, 3, 4, 5, 6, 7/)
    integer, dimension(8) :: temperature = (/0, 1, 2, 3, 4, 5, 6, 7/)
    integer               :: mpierr, rank

    call MPI_INIT(mpierror)
    call MPI_Comm_rank(MPI_COMM_WORLD, rank, mpierr)
    geometry = geometry+rank

    call xh5%Open(FilePrefix='xh5for_unstructured_hexahedron', Strategy=XDMF_STRATEGY_CONTIGUOUS_HYPERSLAB, Action=XDMF_ACTION_WRITE)
    call xh5%SetGrid(NumberOfNodes=8, NumberOfElements=1,TopologyType=XDMF_TOPOLOGY_TYPE_HEXAHEDRON, GeometryType=XDMF_GEOMETRY_TYPE_XYZ)
    call xh5%WriteTopology(Connectivities=topology)
    call xh5%WriteGeometry(XYZ=geometry)
    call xh5%WriteAttribute(Name='Temperature', Type=XDMF_ATTRIBUTE_TYPE_SCALAR ,Center=XDMF_ATTRIBUTE_CENTER_NODE , Values=temperature)
    call xh5%Close()
    call xh5%Free()

    call MPI_FINALIZE(mpierror)

end program xh5for_hexa_per_task

See more info about XH5For_t

Some considerations

• XH5For use XDMF Model and Format to describe the Light data and HDF5 for the Heavy data.
• HDF5 API must be linked to read/write the Heavy data.
• Paralell HDF5 API must to be linked to take advange of the Collective IO.
• Paralell HDF5 API only works if an MPI library is also linked.
• XML syntax is be based on XDMF v2.x.
• The handling of XML files is developed on top of FoX (Fortran XML library) for reading/writing XDMF files. Particularly, FoX_wxml library is used for writing XDMF files and FoX_dom will be used for parsing.

Zen of XH5For:

• Use HDF5 high level library based on MPI_IO
• Open in the right mode
• Number of XDMF/HDF5 files must be independent of the number of processes
• Reduce simultaneous open files to the minimun
• Prefer collective IO
• Do not open/close files too frequently
• Avoid small accesses
• Try MPI_IO hints
• Write/Read as infrequently as possible
• Reduce precision if possible

Best practices for parallel IO and MPI-IO hints

Development roadmap:

• XDMF:
  • Series:
    • Static grid
    • Non-Static grid
  • Grid:
    • Structured
    • Unstructured
  • Element:
    • Linear
      • Triangle
      • Quadrangle
      • Tretrahedron
      • Hexahedron
      • Mixed
      • ...
  • Field:
    • Center on:
      • Node
      • Cell
      • Grid
      • Edge
      • Face
    • Type:
      • Scalar
      • Vector
      • Tensor
      • Matrix
      • Tensor6
      • GlobalID
      • NoType
• HDF5:
  • Groups
  • DataSets
  • HyperSlabs
  • Strategies:
    • Contiguous hyperslab
    • One dataset per process
    • Chunking

UML class diagram:

Automatically generated with ForUML.

How to collaborate

If you want to actively collaborate in the project, please feel free to fork and submit pull requests. Any help or suggestions would be greatly appreciated :)