A minimal implementation of the random ray method for neutral particle transport. This purpose of this code is to serve as an open source performance benchmark for random ray that primarily targets FPGAs.
Key Features:
- Much simpler than the full application (ARRC) -- allowing it to act as a reference implementation
- Complex enough to converge eigenvalues and power distributions for neutron transport problems (see picture below generated with minray)
- Simplified geometry treatment limits complexity (can only simulate 2D heterogeneous Cartesian geometries)
- Hard coded reactor simulation problem of interest: 2D C5G7
- Kernelized for easy porting to accelerator languages
The CPU version of minray is written in C with no dependencies.
The OpenCL version of minray requires an OpenCL capable compiler.
Use the included makefile each the desired source directory to install. Several options are available as toggles at the top of the makefile.
The command line options are:
./minray <options>:
-r <rays>Number of discrete rays-d <distance per ray>Travel distance per ray (cm)-i <inactive iterations>Set fixed number of inactive power iterations-a <active iterations>Set fixed number of active power iterations-s <seed>Random number generator seed (for reproducibility)-m <1, 2, 4, 8, 16, 32>Multiplier to increase/decrease problem size/resolution-pEnables plotting-v <small, medium, large>Executes a specific validation probem to test for correctness
Run the appliation as ./minray to get the default problem. This is a short performance run that uses a realistic problem size per iteration. However, only 20 iterations are run so as to keep the overall runtime low -- meaning the solution will not be converged. The default mode is intended for performance analysis.
When devloping the code or running it on a new system or compiler, it is advisable to test it out using the optional validation mode. Passing the -v <small, medium, large> argument will run the code with a pre-configured problem size and compare the generated eigenvalue to a reference value. Passing requires the eigenvalues match to within a small delta (very small differences due floating point non-associativity are expected).
The -v small should run within a second or so and is meant as a "smoke test" to rapidly debug code. The -v medium and -v large options run progressively larger problems, with the large option being representative of a full simulation to convergence.
Besides the benchmark and validation modes, there are a number of other options for doing custom simulations. For instance, if one wanted to investigate the effect of the Cartesian mesh resolution on simulation accuracy, you could use the mesh multiplier option -m <value> to increase the mesh fineness. At the coarsest setting (-m 1) only a single mesh region is assigned to each pin cell in the 2D C5G7 problem. There are input files available for the -m <1, 2, 4, 8, 16, 32> settings. If only the -m argument is given, minray will automatically increase the number of rays used so as to ensure the mesh is adequately sampled.
The user is also able to manually alter the number of rays used (-r <# of rays>) and the distance per ray in cm (-d <distance>).
By default, the number of inactive and active iterations used is only 10 so as ti keep the runtime low. However, this is not enough iterations to converge a simulation. For a problem of this scale, it is more likely that 1000 inactive and 1000 active iterations would be required, which can be set as: -i 1000 -a 1000.
By default, unless running a validation problem, the seed used to sample the random rays is based on the time of program launch. A seed can manually be set using the -s <seed> argument, which may be useful for debugging when reproducibility is desired.
To plot material/geometry data and several flux spectrums, the -p argument can be given. Plots are output in binary .vtk format, which can be directly loaded into plotting programs like Paraview. If generating plots, it is highly advised that you converge the simulation by increasing the number of inactive and active iterations, e.g.: ./minray -i 1000 -a 1000 -p, and you may also wish to increase the mesh resolution.
The random ray method of neutral particle transport is a recently developed stochastic method derived from the traditionally deterministic Method of Characteristics (MOC). There are a variety of major differences compared to traditional MOC that allow random ray to make full scale 3D reactor simulation practical. Compared to traditional deterministic MOC, random ray is able to converge solutions to large problems in competitive or faster runtimes using far less memory and not requiring non-linear acceleration techniques like CMFD.
More background information on the derivation of the method and how it works in practice can be found in John Tramm's Thesis.
While full random ray applications like ARRC are capable of simulating arbitrary 3D geometries using a constructive solid geometry treatment capable of representing arbitrary second order surfaces (e.g., cylinders, spheres, cones, etc), Minray is only capable of simulating strictly Cartesian geometries in 2D. This simplification is made to keep the source code as compact as possible so as to allow running on a variety of HPC architectures such as FPGAs. In the future, we may add in a more complex (and accurate) geometry treatment.
The result of this simplification is that only crude approximations to real reactor geometries can be simulated. For the target problem of interest, 2D C5G7, since we are using a Cartesian mesh we are not accurately representing the curves of the fuel pins, so our solutions will have some error (compared to the problem reference) even for a very fine mesh. See the below figure for an example comparing the typical constructive solid geometry mesh used by the full appliation ARRC with the simplified Cartesian mesh used in minray. With that in mind, the particular mesh type and resolution required for most problems is already well known and studied, so we are not concerned with that problem in this application where we are focused on testing the basic algorithm on FPGAs and other HPC architectures.
Papers wishing to cite random ray, ARRC, or minray should in general refer to:
John R. Tramm, Kord S. Smith, Benoit Forget, and Andrew R. Siegel. The Random Ray Method for neutral particle transport. Journal of Computational Physics, 342:229 - 252, 2017. https://doi.org/10.1016/j.jcp.2017.04.038
For convenience, the bibtex entry is given below:
@article{Tramm2017, title = {{The Random Ray Method} for neutral particle transport}, journal = "Journal of Computational Physics", volume = "342", pages = "229 - 252", year = "2017", issn = "0021-9991", doi = "https://doi.org/10.1016/j.jcp.2017.04.038", url = "http://www.sciencedirect.com/science/article/pii/S0021999117303170", author = "John R. Tramm and Kord S. Smith and Benoit Forget and Andrew R. Siegel", }