International Journal of Aerospace and Lightweight Structures

doi: 10.3850/S2010428612000232

High Performance Computation by Multi-Node GPU Cluster-Tsubame2.0 on the Air Flow in an Urban City Using Lattice Boltzmann Method

Xian Wang¹ and Takayuki Aoki^{2 1}State Key Laboratory for Strength and Vibration of Mechanical Structures, School of Aerospace, Xi'an Jiaotong University, Xi'an, Shannxi, 710049, China. wangxian@mail.xjtu.edu.cn ²Global Scientific Informational and Computing Center, Tokyo Institute of Technology, 2-12-1, Meguro-ku, Tokyo, 152-8550, Japan. taoki@gsic.titech.ac.jp

ABSTRACT

General Purpose Graphic Processing Unit (GPGPU) has drawn much attention on accelerating non-graphic applications. The simulation by D3Q19 model of Lattice Boltzmann method was executed successfully on multi-node GPU cluster by using CUDA programming and MPI library. The numerical code ran on the multi-node GPU cluster TSUBAME2.0 of Tokyo Institute of technology, which includes 1408 computing nodes and all of them are equipped with three NVIDIA Tesla M2050 GPU accelerators, total 4224 GPUs are equipped. In the present work, a large-scaled computation on the air flow in an urban city was carried on using 120 GPUs of TSUBAME2.0. The number of computational grids was 3072£ 2000 £ 256 and the computational domain was decomposed in a three dimensional way for parallel computation. The large eddy simulation (LES) was adopted for the turbulence. As a result, for the conditions of 1km x 1km area and 273 s flow time, the time for simulation is 585 seconds, in which the computational time is 461 seconds and data communicational time is 124 seconds, respectively. The achieved performance is about 6 TFLOPS.

Keywords: GPGPU, Parallel computation, Lattice Boltzmann method.

---

Back to TOC

FULL TEXT(PDF)