branson

branson copied to clipboard

Branson

Introduction

So you've decided to use Branson...

Here are some things to know:

• Branson is not an acronym.
• The point of Branson is to study different algorithms for parallel Monte Carlo transport. Currently it contains a particle passing method for domain decomposition (two mesh passing methods for domain decomposition are available in older versions)
• Many of the parameters that impact parallel performance can be set in the input file.
• Input files are in XML, which makes them easy to generate and change in python.
• Input files are complicated when you want to have multiple spatial regions but are pretty simple otherwise.

How to install

Accessing the sources

• Download a tarball from Github, or
• Fork and clone the git repository.

# after forking the repository
git clone [email protected]:[github-username]/branson
cd branson
git remote add upstream [email protected]:lanl/branson
git checkout -b develop upstream/develop

Installing Branson:

• Build requirements:
  • C/C++ compiler(s) with support for C11 and C++14.
  • CMake 3.9+
  • MPI 3.0+ (OpenMPI 1.10+, mpich, etc.)
• Optional dependencies
  • OpenMP Used for parallelism on an MPI rank, currently just for the transport phase. Note that the number of OpenMP threads is taken from the input file and not the user's environment
  • Metis Used to decompose the mesh, if Metis is not found, Branson uses a simple decomposition where cells are divided evenly between ranks with x being the fastest index and z being the slowest index
  • HDF5
  • Silo
  • If these tools aren't readily avilable on your target machine, consider using a package manager like spack to help you install these tools.
• There are multiple CMake options:
  • CMAKE_BUILD_TYPE, set on the command line with -DCMAKE_BUILD_TYPE=<Debug|Release> where the default is Release.
  • USE_OPENMP, set on the command line with -DUSE_OPENMP=<ON|OFF>, if this variable is not set it will default to ON
  • ENABLE_VERBOSE_GPU_TRANSPORT, set on the command line with DENABLE_VERBOSE_GPU_TRANSPORT=<ON|OFF, defaults to OFF, useful for GPU debugging
• If CMake has trouble finding your installed TPLs, you can try
  • appending their locations to CMAKE_PREFIX_PATH,
  • Setting helper variables like HDF5_ROOT (refer to the cmake documentation or the FindXXX.cmake scripts in Branson's src/config directory for a list of variables), or
  • try running ccmake . from the build directory and changing the values of build system variables related to TPL locations.

EXPORT CXX=`which g++`
cd $build_dir
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=<install-location> ${branson_source_dir}/src
make -j

Testing the build:

cd $build_dir
ctest -j 32

Parameters

• In the common block of the XML input file you can set several parameters related to parallel performance:
  • batch_size: how many particles are processed by an MPI rank before it stops to process particle messages (small means more processing of MPI messages and possibly more interleaving of off rank and on rank work)
  • particle_message_size: the size of the communication buffer for particle data (small means more message but possibly more interleaving of on rank and off rank work)
  • mesh_decomposition: Can be METIS or CUBE, generally use Metis unless you're trying to run a very large problem (Metis is serial and ParMetis can't be used due to licensing restrictions). For a cube decomposition, the number of ranks must be perfect cubes (x^(1/3) is an interger)

Special builds

Fake Multigrop Branson:

• The clean_mg branch has a new capability that mimics the data flow in real multigroup transport.
• This branch sets the number of groups at the configure step in CMake. I know this is messy from a usability standpoint but it makes the data layout very easy to control when the number of groups is known at compile time. Use the N_GROUPS CMake variable to set the number of groups (e.g. cmake -DN_GROUPS=10 ../path/to/CMakeLists.txt).
• The code will still produce gray results! The physical value in each group is the same and it's still set via the input deck.
• This branch samples a group with a uniform PDF (it does not weight the opacity with a Planckian spectrum).

Running Branson on performance problems

• There are two performance problems of interest in the inputs folder, they are both simplified 3D hohlraums and should be run with a 30 group build of Branson (see Special builds section above).
• The 3D_hohlaum_single_node.xml problem is meant to be run on a full node. It is in replicated mode which means there is very little MPI communication (end of cycle reductions). It is run with:

mpirun -n <procs_on_node> <path/to/branson> 3D_hohlaum_single_node.xml

• The 3D_hohlaum_multi_node.xml problem is meant to be run on many nodes. It is in domain decomposed mode which means particles must be communicated between spatial domains. It is run with:

mpirun -n <n_ranks> <path/to/branson> 3D_hohlaum_multi_node.xml

Note that Branson does not currently have any threading capability so n_ranks should usually be n_nodes*n_ranks_per_node. This problem is meant to consume about 30% of a 128 GB node.

Authors

• Branson was written by Alex R. Long
• Random123 is by D. E. Shaw Research, Copyright 2010-2011
• PugiXML is by zeux (Arseny Kapoulkine), MIT License
• RNG.h uses code from Draco, BSD-3 Clause License

Release

• Branson is released under the BSD 3-Clause License. For more details see the LICENSE.md file.

• Branson development follows the development model outlined in the CS-memo-2020-11-18.pdf file, available in the top-level of this repo

• LANL code designation: `C17048