Efficient Parallel Particle Advection via Targeting Devices
Loading...
Date
2020-09-24
Authors
Belcher, Kristi
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oregon
Abstract
Particle advection is a fundamental operation for a wide range of flow visualization algorithms. Particle advection execution times can vary based on many factors, including the number of particles, duration of advection, and the underlying architecture. In this study, we introduce a new algorithm for parallel particle advection which improves execution time by targeting devices, i.e., adapting to use the CPU or GPU based on the current work. This algorithm is motivated by the observation that CPUs are sometimes able to better perform part of the overall computation since CPUs operate at a faster rate when the threads of a GPU can not be fully utilized. To evaluate our algorithm, we ran 162 experiments and compared our algorithm to traditional GPU-only and CPU-only approaches. Our results show that our algorithm adapts to match the performance of the faster of CPU-only and GPU-only approaches.
Description
Keywords
Device Targeting, GPUs, Heterogeneous Computing, Particle Advection, Scientific Visualization, VTK-m