A Study on GPU Parallelization and Performance Optimization of the Athena++ Simulation Code in High-Performance Computing Environments
Hyun Mi Jung, Hyunjo Lee, Kimoon Jeong, Cheol-Joo Chae · Korean Institute of Smart Media · 2025
This paper identifies performance hotspots in the magnetohydrodynamics (MHD) simulation code Athena++ and proposes parallel optimization techniques for computational acceleration in high-performance computing (HPC) environments. Through code profiling tools such as gprof, valgrind, and vtune, the primary computational hotspot was found to be the Hydro::RiemannSolver module. Performance evaluation was conducted in the Google Colab environment using the A100 GPU. The results demonstrated over a 25-fold improvement in average execution time per computation cycle compared to the CPU implementation, confirming substantial enhancement in computational efficiency through parallel processing of repetitive structures. These findings suggest that parallelization in heterogeneous CPU-GPU environments can significantly improve the performance of high-fidelity simulation codes. Moreover, architecture design and data flow optimization tailored for next-generation computing devices such as CXL and DPUs are expected to play a critical role in future HPC application performance improvement.