Fast Coherent Particle Advection through Time-Varying Unstructured Flow Datasets

Area of Science:

• Computational Fluid Dynamics
• Scientific Visualization
• High-Performance Computing

Background:

• Particle path tracing is crucial for flow visualization and analysis.
• Advecting particles on unstructured meshes causes memory access issues due to particle path divergence.
• Existing parallelization methods struggle with incoherent memory accesses on graphics processing units (GPUs).

Purpose of the Study:

• To develop a novel GPU-based approach for efficient particle advection on unstructured meshes.
• To improve memory coherence and enhance the performance of shared-memory GPUs in particle advection tasks.
• To address the challenges of parallelizing particle advection in complex flow fields.

Main Methods:

• Introduced a block advection GPU approach that bundles particles into spatially coherent groups.
• Reorganized particle trajectories to optimize memory access patterns on GPUs.
• Applied the method to unstructured meshes meeting specific criteria (CFL condition, uniform element size, sufficient GPU memory).

Main Results:

• The block advection method significantly improves memory coherence during particle advection.
• Achieved enhanced shared-memory GPU performance compared to traditional parallelization techniques.
• Demonstrated effectiveness for flows satisfying the CFL criterion on suitable unstructured meshes.

Conclusions:

• The block advection GPU approach offers a viable solution for accelerating particle path tracing in flow analysis.
• This method enhances computational efficiency for scientific visualization on modern GPU architectures.
• Future work could explore applicability to more complex mesh types and flow conditions.