Application of GPGPU for the Reduction of Computational Time of Smoothed Particle Hydrodynamics Based Numerical Models to Simulate Morphological Changes of Single Plant Cells during Drying.

Abstract

Drying, which happens in a complex manner inside the food m9.terial, is a popular method for preserving food materials by reducing water from t1e cells and tissues. Drying process can be simulated with the aid of numerical modelling in order to understand the fundamentals involved in the process: allowing better-controlled drying processes yielding better food products. In recent time, Meshfree methods such as Smoothed Particle Hydrodynamics (SPH) has become a popular numerical modelling technique due to the capability of modelling complex mechanisms, large deformations, and discrete cellular network. Also, such SPH based numerical models are conventionally designed to run on parallel CPU cores in order to reduce the computational time. However, when the targeted problem domain becomes larger, the computational time can become prohibitively higher since the CPU contains a limited number of cores. As a computationally efficient alternative, this study presents the combined use of CPU and General Purpose Graphical Processing Unit (GPGPU) in plant cell simulations. Accordingly, the source code was developed by combining CPU and GPGPU parallel processing through a Compute Unified Device Architecture (CUDA) GPU programming language. In order to simulate drying of a plant cell, this source code was run on a High-Performance Computer (HPC) with an Intel® Core™ i7-6700, 3.4 GHz x 8 core CPU, 32 GB RAM and NVIDIA QUADRO K1200 GPGPU with 4 GB RAM and 512 CUDA cores, in Ubuntu 14.04 interface. It was observed that compared to the CPU only source code, the CPU-GPGPU combined source code leads to 24% of computational time saving highlighting the significance of the GPGPU meshfree particle-based modelling. The performance of the proposed method can be further improved by enhancing the memory transferring and workload distribution between the CPU and GPGPU. Such developments will assure convenient simulations of a large problem domain in order to reveal the underline fundamental mechanisms involved.