Visualizing large 3D geodesic grid data with massively distributed GPUs

Jinrong Xie, Hongfeng Yu, Kwan-Liu Ma

Abstract: Geodesic grids become increasingly prevalent in large weather and climate applications. The deluge amount of simulation data demands efficient and scalable visualization capabilities for scientific exploration and understanding. Given the unique characteristics of geodesic grids, no current techniques can scalably visualize scalar fields defined on a geodesic grid. In this paper, we present a new parallel ray-casting algorithm for large geodesic grids using massively distributed GPUs. We construct a spherical quadtree to adaptively partition and distribute the data according to the grid resolution of simulation, and ensure a balanced workload assignment over a large number of processors from different view angles. We have designed and implemented the entire rendering pipeline based on the MPI and CUDA architecture, and demonstrated the effectiveness and scalability of our approach using an example of large application on a supercomputer with thousands of GPUs.