Lay Summary

Simulating 3D fluid dynamics around intricate real-world geometries—such as urban buildings or complex machinery—demands immense computing power distributed across thousands of processor cores. When fluid motion varies unpredictably across the simulation domain, some processors become overloaded while others sit idle. This paper presents an open-source 3D fluid dynamics framework featuring automatic dynamic load balancing. By continuously reallocating computational tasks among processors as the flow evolves, the algorithm maintains even workload distribution, dramatically improving simulation efficiency and execution speed on supercomputing clusters. (84 words)

Full Abstract

We present an open-source computational fluid dynamics (CFD) code designed for scalable 3D flow simulations in complex geometries subject to severe and dynamic load imbalances. Leveraging asynchronous message-driven execution and dynamic load balancing via Charm++, the framework continuously redistributes unstructured mesh elements across available processing elements. This capability prevents computational bottlenecks caused by transient local features, adaptive mesh refinement, or multi-physics couplings. Performance evaluation across thousands of supercomputer cores shows near-ideal speedup, confirming the framework’s suitability for large-scale engineering and environmental flow applications.

Metadata

• Publication Date: 2024
• Author: József Bakosi, Mátyás Constans, Zoltán Horváth, Ákos Kovács, László Környei, Marc Charest, Aditya Pandare, Paula Rutherford, Jacob Waltz
• Journal: Computer Physics Communications / Journal of Computational Physics