A Python-based flow solver for numerical simulations using an immersed boundary method on single GPUs

Area of Science:

• Computational fluid dynamics
• High-performance computing
• Numerical simulation

Background:

• Simulating three-dimensional incompressible flow around complex geometries is computationally intensive.
• Existing methods often face challenges with communication overheads in parallel processing.

Purpose of the Study:

• To develop an efficient single-GPU implementation for simulating incompressible flow around moving bodies with complex geometries.
• To optimize performance by leveraging GPU architectures and the Immersed Boundary Method.

Main Methods:

• Implementation of the Immersed Boundary Method tailored for GPU using Nvidia CUDA, Numba, and Python.
• Exploitation of different GPU grid architectures for performance optimization.
• Focus on single-GPU execution to eliminate communication overheads and keep data in global memory.

Main Results:

• The single-GPU code achieves a speedup of 34 to 54 times compared to a CPU-based parallel solver.
• Significant speedups are observed in solving linear systems, with more modest gains in non-linear term computations (×1.6-3).
• The code demonstrates strong scaling on different GPU hardware for both external (bioinspired aerodynamics) and internal (cardiovascular) flow applications.

Conclusions:

• The developed single-GPU implementation provides a highly efficient solution for complex fluid flow simulations.
• This approach offers a substantial performance improvement over traditional CPU-based methods.
• The method is validated and applicable to diverse fluid dynamics problems, including bioinspired and cardiovascular flows.