GPU Optimization for High-Quality Kinetic Fluid Simulation.

Area of Science:

• Computational fluid dynamics
• Applied mathematics
• High-performance computing

Background:

• Incompressible Navier-Stokes Equations (INSE) are standard for fluid simulations but create difficult sparse linear systems for parallel computation.
• Kinetic methods, specifically the adaptive-central-moment multiple-relaxation-time (ACM-MRT) model, offer an alternative with high-quality simulation capabilities for laminar and turbulent flows.
• ACM-MRT's local formulation is promising for scalable parallel implementations on Graphics Processing Units (GPUs).

Purpose of the Study:

• To present novel GPU optimization techniques for efficient ACM-MRT kinetic fluid simulations.
• To address computational challenges in simulating fluids with complex solid objects.
• To enable high-quality, scalable fluid simulations on parallel systems.

Main Methods:

• Developed novel GPU optimization techniques including a communication-efficient data layout and a load-balanced immersed-boundary method.
• Implemented a multi-kernel launch strategy with a simplified ACM-MRT formulation for increased parallelism.
• Integrated techniques into a parametric cost model for automated performance optimization and extended to multi-GPU systems.

Main Results:

• Demonstrated state-of-the-art performance and high visual quality in ACM-MRT-based fluid simulations.
• Achieved efficient simulation of complex fluid domains containing solid objects.
• Showcased scalability on parallel systems, including multi-GPU setups.

Conclusions:

• The novel GPU optimization techniques significantly enhance the efficiency and scalability of ACM-MRT kinetic fluid simulations.
• The developed solver provides a powerful tool for high-quality fluid dynamics research, particularly in scenarios involving complex geometries.
• This work paves the way for large-scale, efficient fluid simulations on modern parallel architectures.