A Monte Carlo Neutron Transport Mini-App
This application is a simplified Monte Carlo neutral particle transport mini-app that support a number of physical processes as particle histories are tracked.
Before building the dependent neutral application, it is necessary to clone the application into the arch project. The instructions can be found on the arch project README.
git clone [email protected]:uob-hpc/arch
cd arch
git clone [email protected]:uob-hpc/neutral
cd neutral
The neutral build process is intended to be simple, and has been tested on a number of platforms.
make KERNELS=omp3 COMPILER=INTEL ARCH_COMPILER_CC=mpicc
The KERNELS option determines the particular kernel set that will be used when building the project. At this time the difference between the kernel sets is that they are written with different programming models. When you clone the repository there are multiple directories that contain duplicates of the core computational kernels, ported to different models and the name of the directory is the value to be used with the KERNELS option.
A number of other switches and options are provided:
DEBUG=<yes/no>- 'yes' switches off optimisation and adds debug flagsMPI=<yes/no>- 'yes' turns off any use of MPI within the application.- The
OPTIONSmakefile variable is used to allow visit dumps, with-DVISIT_DUMP, and profiling, with-DENABLE_PROFILING.
Please note: We do not support granular profiling with the over particles parallelisation scheme because it has a negative impact on the performance of the application and gives spurious results.
Upon building, an binary file will be output with the extension of the value of KERNELS. e.g. neutral.omp3. You can run the application with, for example:
./neutral.omp3 problems/scatter
The problems/scatter problem is described in the following section.
The configuration files expose a number of key parameters for the application.
iterations- the number of iterations the application will proceed throughdt- the timestep for the applicationnx- the number of cells in the x-dimensionny- the number of cells in the y-dimensioninitial_energy- the initial energy that all particles will be set to
The performance of the Monte Carlo application is highly problem dependent, and so we provide multiple configuration files that present different computation problems:
problems/scatter- the particles will mostly collide within the slabproblems/stream- the particles stream across the mesh multiple times per timestep, colliding infrequentlyproblems/csp- the particles stream until they encounter a region of high density in the center of the slabproblems/split- the particles are spread evenly in regions of high and low density to match the number of events of each type
TODO: Describe the problem and source descriptions in the parameter file.
The implementation is currently in an active development phase. There are multiple branches that are exploring algorithmic changes and other optimisations in order to test the performance of the application on modern architectures.
master- the main branch of the code, where parallelisation is over particles.event-based- adjusts the parallelisation strategy so that events are handled for all particles in a synchronous manner.hybrid- supports event-based parallelisation with progress through successive events.master-soa- changes the master branch to use an SoA data structure.tiled- attempts to tile the over particles parallelisation strategy to improve cache locality.
The mini-app currently supports elastic scattering, with realistic cross sections. We intend to extend the application to include particle production.
Random123 is a highly usable counter-based random number generator that we use for random number generation, https://www.deshawresearch.com/resources_random123.html.