EFFICIENT COMPUTATIONAL APPROACHES TO GPU-BASED MONTE CARLO RADIATION TRANSPORT
No Thumbnail Available
Date
2025-04-24
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Nazarbayev University School of Engineering and Digital Sciences
Abstract
This thesis focuses on efficient computational approaches for Monte Carlo Radiation Transport (MCRT) simulations using modern Graphics Processing Units (GPUs). Over the last decade, GPUs have become an important part of scientific computing due to their capability to perform large-scale parallel computations, particularly in areas such as radiation transport. Modern MCRT applications are complex simulations that require immense computational power. This work addresses the challenges and opportunities in using GPU architectures for the MCRT simulations and contributes to the understanding of how to optimize these simulations for better performance and energy efficiency.
A detailed performance examination of several parallel pseudorandom number generators (PRNGs) running on various Nvidia GPU cards is presented in the thesis. MRG32k3a, MTGP32, PHILOX4_32_10, MT19937, and XORWOW are five PRNGs from the cuRAND library that are evaluated for their efficiency in producing uniform and non-uniform random numbers using a range of implementation options, including GPU-only, CPU-only, and hybrid CPU/GPU approaches. This assessment advances our knowledge of PRNG performance optimization on GPUs, particularly with regard to the Monte Carlo (MC) simulations.
The thesis also evaluates two popular Python-based GPU programming platforms, CuPy and Numba, benchmarking against CUDA C for the MCRT simulations. This evaluation is based on performance and energy consumption using memory-intensive operations and compute-heavy problems. The analysis was conducted on Nvidia GeForce RTX3080, Tesla V100, and Tesla A100 GPU cards. It offers information about the advantages and disadvantages of these platforms, which is valuable to the scientific community when selecting tools for GPU-based simulations.
Further, the work investigates the performance scaling of MCRT simulations on multiple GPUs, focusing on strong and weak scaling, optimization strategies such as fast math and block-thread configuration, and energy consumption. Using an Nvidia DGX-2 server with up to 10 GPUs, the study demonstrates how different scaling strategies and optimization techniques affect both performance and energy efficiency. This research provides practical recommendations for improving the use of multiple GPUs in large-scale MCRT simulations, contributing to the knowledge of multi-GPU programming and optimization.
Overall, this thesis contributes to the understanding of how to efficiently run and optimize MCRT simulations on GPUs. It includes a detailed analysis of PRNGs performance, evaluates popular Python-based computing tools, and explores how well these platforms can scale their applications across multiple GPUs. This work provides useful insights for researchers, students, and professionals who work with GPU computing, particularly in the field of MCRT simulations.
Description
Keywords
Type of access: Open access
Citation
Askar, Tair. (2025). Efficient computational approaches to GPU-based Monte Carlo radiation transport. Nazarbayev University School of Engineering and Digital Sciences