UME: Unstructured Mesh Explorations

Description

UME is a collection of data structures and algorithms for parallel unstructured meshes. It is intended to be used to explore the co-design space of hardware and software to support computational fluid dynamics simulation packages.

Installation

UME is a CMake-based package written in C++20. The basic steps for building UME are:

1. Install external packages;
2. Do a cmake configuration step;
3. Build the libraries and executables.

The primary product of UME is a library and header files, as well as some example driver executables.

External Packages

UME relies on several external packages. These can be installed easily using the spack package manager. The spack specs for all packages are:

spack:
  specs:
    - catch2
    - cmake
    - doxygen
    - git
    - graphviz
    - kokkos+openmp std=20
    - openmpi
  packages:
    catch2:
      require: '@3:'

CMake Configuration Options

CMake options can be specified either with the interactive ccmake tool, or on the command line of cmake, preceded by the -D flag. Package-specific CMake options for UME include:

• USE_CATCH2=NO will disable Ctest testing that relies on version 3 of the Catch2 testing framework.
• USE_KOKKOS=YES will enable Kokkos support, and requires Kokkos to be found in the environment (UME does not currently use Kokkos).
• USE_MPI=YES will enable MPI support, and requires MPI compilers to be available on the search path.
• USE_OPENACC=YES will enable OpenACC support, and requires the C++ compiler to support it (UME does not currently use OpenACC).

Generated Executables

After a successful build, two example driver executables will be generated that are of interest to the general community: ume_serial and ume_mpi. Each exercises the gradient operators found in the Ume/gradient.hh headers. The MPI version ume_mpi is the most realistic. The usage is:

% ume_serial <filename>

Where <filename> is the complete file name for an UME binary input file (distributed separately). Or

% mpirun -np <n> ume_mpi <prefix>

Where there is expected to be a set of files named as if generated with sprintf(filename, '%s.%05d.ume', prefix, rank) for 0 <= rank < n.

There will also be two utility executables: txt2bin and scale_mesh. The txt2bin utility takes in an .umetxt file and converts it to a binary representation that is ingestible by Ume.

% txt2bin <infile> <filename>

Where <infile> is the complete file name for an UME text input file and <filename> is the name of the UME binary file to be used by UME.

The scale_mesh utility takes in an UME binary input file and increases the size of the mesh by a user-chosed factor. The scaling factor is currently constrained to be a factor of 2.

% scale_mesh <prefix> 8

Where <filename> is the base file name for a set of UME binary input files.

You can also run scale_mesh with MPI

% mpirun -np <n> scale_mesh <prefix> 8

Run gradient kernel multiple times

By default, all the kernels is run once. There is an option to run these kernels in ume_mpi multiple times to increase the computations for benchmarking purposes using command line argument.

 % mpirun -np <n> ume_mpi <prefix> -i <number of iterations>

Project Name

"Ume" is also the romanization of the Japanese word for "plum" (梅, or うめ). Sort of pronounced in English as if you were asking the question "oo-meh?"

Authors

UME was created by Paul Henning ([email protected]).

UME is currently maintained by: Austin Bullock ([email protected]) Sumathi Lakshmiranganatha ([email protected]) Jered Dominguez-Trujillo ([email protected])

Release

UME is released as open source under a BSD-3 License. For more details, please see the LICENSE.md and NOTICE.md files.

LANL C23016

© 2023. Triad National Security, LLC. All rights reserved.