A multi-GPU and CUDA-aware MPI-based spectral element formulation for ultrasonic wave propagation in solid media.

Area of Science:

• Computational Mechanics
• Solid Mechanics
• Wave Propagation

Background:

• Accurate simulation of ultrasonic wave propagation in solids is crucial for non-destructive testing and material characterization.
• Existing methods often face computational limitations in terms of speed and scalability for complex geometries.

Purpose of the Study:

• To introduce a new multi-GPU spectral element (SE) formulation for enhanced simulation of ultrasonic wave propagation.
• To improve communication efficiency between GPUs for faster and more scalable computations.
• To demonstrate the formulation's effectiveness in simulating wave interactions with realistic structural defects.

Main Methods:

• Developed a novel multi-GPU spectral element formulation utilizing CUDA-aware Message Passing Interface (MPI).
• Implemented two new message exchange strategies for direct GPU-to-GPU communication of nodal forces.
• Benchmarked the new formulation against a multi-CPU, classical MPI-based approach.

Main Results:

• Achieved remarkable acceleration across all computational stages: matrix assembly, time integration, and message exchange.
• Demonstrated scalability of computational efficiency and degree-of-freedom limit with increasing numbers of GPUs.
• Successfully simulated the interaction of Lamb waves with thickness loss defects in plates.

Conclusions:

• The new multi-GPU formulation offers significant speedups and scalability for ultrasonic wave propagation simulations.
• This technique shows potential for efficient, accurate, and robust analysis of wave propagation in complex engineering structures.
• The direct GPU-to-GPU communication strategy enhances computational performance compared to CPU-mediated methods.