GPU-Oriented Algorithms for Continuous Energy Monte Carlo Neutron Transport
The advent of graphics processing units (GPUs) has brought computing to new heights with deep learning models, now deployed ubiquitously and touching the lives of many. While GPU hardware may be ideal for deep learning, its full potential in various scientific computing applications has yet to be re...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2024
|
| Online Access: | https://hdl.handle.net/1721.1/157192 |
| _version_ | 1824458389714894848 |
|---|---|
| author | Ridley, Gavin |
| author2 | Forget, Benoit |
| author_facet | Forget, Benoit Ridley, Gavin |
| author_sort | Ridley, Gavin |
| collection | MIT |
| description | The advent of graphics processing units (GPUs) has brought computing to new heights with deep learning models, now deployed ubiquitously and touching the lives of many. While GPU hardware may be ideal for deep learning, its full potential in various scientific computing applications has yet to be realized. Often, paradigm shifts in the data formalisms and algorithmic choices used to solve scientific computing problems must take place to fully leverage GPUs. A quintessential example of this shift has been the move towards matrix-free, high-order finite element formulations researched under the Exascale Computing Project. Similar groundbreaking shifts are only starting to take place in continuous energy Monte Carlo (MC) neutron transport simulations. These simulations play a crucial role in designing fission, fusion, and security systems that may play a pivotal role in the transition to a decarbonized world. This work contributes to adapting continuous energy MC neutron transport simulations for the GPU computing era. We first summarize some changes made to other scientific computing applications that led to performance gains on GPUs, which informed our independent development of a CUDA-based version of OpenMC, an open-source continuous energy MC neutron and photon transport code. Fortunately, the historical event-based MC simulation modality developed extensively through the 1980s for vector computers provides an excellent basis for GPU computing. Adapting a full-physics, continuous energy MC neutron transport simulation for GPUs is a feat only completed by a few institutions across the world, so we share some software development tricks that facilitated this task. We then identify a variety of algorithmic optimizations that improved the performance of the baseline CUDA application, and identify areas for further development. 3 Based on experience adapting a full-physics continuous energy MC code for GPU, we identify two pieces of the simulation which can be improved for GPU computing: resonance upscatter handling and unresolved resonance modeling. Our new method for modeling resonance upscatter is based on a novel, fundamental observation regarding the resonance upscatter effect. The relative speed tabulation (RST) method developed by other GPU MC researchers can be underpinned by a universal special function we have named the incomplete Faddeeva function, which is closely related to the incomplete Goodwin-Staton integral. Our research develops numerical algorithms for efficient, accurate computation of the incomplete Faddeeva function and identifies some properties of the function. We then present a specialized root-finding algorithm that takes advantage of the structure of the problem to efficiently sample the resonance upscatter effect on GPUs. This obviates the need to rely on RST tables or a zero kelvin pointwise cross section, freeing precious GPU memory while using a GPU-friendly memory access pattern. Continuing in the same direction, we focus on unresolved resonance region (URR) crosssection modeling, which was shown to induce a 30% computational efficiency degradation on GPUs. We review the requirements to model cross sections in the unresolved resonance regime, and provide what is to our knowledge the first rigorous demonstration that URR modeling can be reduced to a one-dimensional probabilistic model in addition to some expectation values of partial cross sections conditioned on the total. Through three asymptotic arguments covering different resonant behavior regimes, we show that the normal inverse Gaussian distribution is the natural choice for modeling the total neutron cross-section distribution. Rather than inducing a performance degradation, we show the new URR modeling technique in fact outperforms a pointwise infinite-dilute approach when it is used to model the URR region. |
| first_indexed | 2025-02-19T04:25:07Z |
| format | Thesis |
| id | mit-1721.1/157192 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2025-02-19T04:25:07Z |
| publishDate | 2024 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1571922024-10-10T03:05:58Z GPU-Oriented Algorithms for Continuous Energy Monte Carlo Neutron Transport Ridley, Gavin Forget, Benoit Massachusetts Institute of Technology. Department of Nuclear Science and Engineering The advent of graphics processing units (GPUs) has brought computing to new heights with deep learning models, now deployed ubiquitously and touching the lives of many. While GPU hardware may be ideal for deep learning, its full potential in various scientific computing applications has yet to be realized. Often, paradigm shifts in the data formalisms and algorithmic choices used to solve scientific computing problems must take place to fully leverage GPUs. A quintessential example of this shift has been the move towards matrix-free, high-order finite element formulations researched under the Exascale Computing Project. Similar groundbreaking shifts are only starting to take place in continuous energy Monte Carlo (MC) neutron transport simulations. These simulations play a crucial role in designing fission, fusion, and security systems that may play a pivotal role in the transition to a decarbonized world. This work contributes to adapting continuous energy MC neutron transport simulations for the GPU computing era. We first summarize some changes made to other scientific computing applications that led to performance gains on GPUs, which informed our independent development of a CUDA-based version of OpenMC, an open-source continuous energy MC neutron and photon transport code. Fortunately, the historical event-based MC simulation modality developed extensively through the 1980s for vector computers provides an excellent basis for GPU computing. Adapting a full-physics, continuous energy MC neutron transport simulation for GPUs is a feat only completed by a few institutions across the world, so we share some software development tricks that facilitated this task. We then identify a variety of algorithmic optimizations that improved the performance of the baseline CUDA application, and identify areas for further development. 3 Based on experience adapting a full-physics continuous energy MC code for GPU, we identify two pieces of the simulation which can be improved for GPU computing: resonance upscatter handling and unresolved resonance modeling. Our new method for modeling resonance upscatter is based on a novel, fundamental observation regarding the resonance upscatter effect. The relative speed tabulation (RST) method developed by other GPU MC researchers can be underpinned by a universal special function we have named the incomplete Faddeeva function, which is closely related to the incomplete Goodwin-Staton integral. Our research develops numerical algorithms for efficient, accurate computation of the incomplete Faddeeva function and identifies some properties of the function. We then present a specialized root-finding algorithm that takes advantage of the structure of the problem to efficiently sample the resonance upscatter effect on GPUs. This obviates the need to rely on RST tables or a zero kelvin pointwise cross section, freeing precious GPU memory while using a GPU-friendly memory access pattern. Continuing in the same direction, we focus on unresolved resonance region (URR) crosssection modeling, which was shown to induce a 30% computational efficiency degradation on GPUs. We review the requirements to model cross sections in the unresolved resonance regime, and provide what is to our knowledge the first rigorous demonstration that URR modeling can be reduced to a one-dimensional probabilistic model in addition to some expectation values of partial cross sections conditioned on the total. Through three asymptotic arguments covering different resonant behavior regimes, we show that the normal inverse Gaussian distribution is the natural choice for modeling the total neutron cross-section distribution. Rather than inducing a performance degradation, we show the new URR modeling technique in fact outperforms a pointwise infinite-dilute approach when it is used to model the URR region. Ph.D. 2024-10-09T18:27:20Z 2024-10-09T18:27:20Z 2024-09 2024-10-09T14:28:05.299Z Thesis https://hdl.handle.net/1721.1/157192 In Copyright - Educational Use Permitted Copyright retained by author(s) https://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Ridley, Gavin GPU-Oriented Algorithms for Continuous Energy Monte Carlo Neutron Transport |
| title | GPU-Oriented Algorithms for Continuous Energy Monte
Carlo Neutron Transport |
| title_full | GPU-Oriented Algorithms for Continuous Energy Monte
Carlo Neutron Transport |
| title_fullStr | GPU-Oriented Algorithms for Continuous Energy Monte
Carlo Neutron Transport |
| title_full_unstemmed | GPU-Oriented Algorithms for Continuous Energy Monte
Carlo Neutron Transport |
| title_short | GPU-Oriented Algorithms for Continuous Energy Monte
Carlo Neutron Transport |
| title_sort | gpu oriented algorithms for continuous energy monte carlo neutron transport |
| url | https://hdl.handle.net/1721.1/157192 |
| work_keys_str_mv | AT ridleygavin gpuorientedalgorithmsforcontinuousenergymontecarloneutrontransport |