A Memory Efficient Image Composition-based Parallel Particle Based Volume Rendering

抄録

Large-scale simulations have widely been conducted on modern High-Performance Computing (HPC) systems in a variety of scientific and engineering fields, and scientific visualization has been a popular approach for analyzing and extracting meaningful information from the simulation results. In this work, we focused on Particle Based Volume Rendering (PBVR) method because of its proven effectiveness for handling non-trivial unstructured volume data, which is still commonly used on numerical simulations in the engineering fields. PBVR possesses a visibility sorting-free characteristics thanks to its use of small and opaque particles as the rendering primitives. However, there is a memory cost and image quality trade-off because of the necessity of storing the entire sets of generated particle data before starting the rendering process. In this paper, we present a memory cost efficient parallel PBVR approach for enabling high-quality and high-resolution PBVR of large-scale numerical simulation results. For this purpose, we focused on the image data gathering and processing instead of traditional particle data gathering and processing by using the sort-last parallel image composition approach. We evaluated its effectiveness on the K computer by using the Binary-Swap-based 234Compositor library, and verified its potential for reducing the memory cost while generating high-quality and high-resolution image data.