Intel Many-Integrated Core (MIC) architecture-based parallel computation and optimization of finite-difference time-domain (FDTD) schemes for acoustic simulation

Abstract

IntroductionRecent years have seen a growing interest in many-corebased parallel computing with a graphics processing unit (GPU) or with Intel many-integrated core (Intel MIC) accelerators in a broad range of fields including acoustic simulation. The movement toward applying GPU to scientific computing has been ongoing since 2005 [1]. A number of studies have reported the application of GPU [2-4] to largescale and long-time acoustic simulation for finite-difference time-domain (FDTD) methods [5][6][7]. On the other hand, the first generation of Intel MIC architecture is newer than GPU because it was only released in 2013. Therefore, fewer studies applying Intel MIC to FDTD methods have been reported [8]; hence, the performance of this accelerator is not known. An Intel MIC accelerator has higher availability than a GPU because its programming model shares many similarities with regular CPUs, which means that Intel MIC can be utilized by regular OpenMP parallelization code written in C/C++ or Fortran. In contrast, GPUs are programmed through APIs such as OpenCL or CUDA (for the NVIDIA GPU).Therefore, in this study, we evaluate the performance of three kinds of acoustic FDTD schemes on the Intel MIC architecture. In addition, we perform software optimizations known in the field of high-performance computing (HPC) [9] to reveal the attainable performance of the Intel MIC architecture. The acoustic FDTD schemes we examined are the standard FDTD(2,2) [10], FDTD(2,4) [11], and WE (Wave Equation) -FDTD(2,2) [3,12] schemes.The purpose of this study is to evaluate the performance of acoustic FDTD schemes on the Intel MIC architecture and to reveal the attainable performance by performing software optimizations.