Petascale turbulence simulation using a highly parallel fast multipole method
Dept. of Mechanical Engineering, Boston University, Boston, MA 02215
arXiv:1106.5273v1 [cs.NA] (26 Jun 2011)
@article{Yokota:2011cs,
author={Yokota, Rio and Narumi, Tetsu and Barba, L. A. and Yasuoka, Kenji},
title={Petascale turbulence simulation using a highly parallel fast multipole method},
year={2011},
eprint={1106.5273},
archivePrefix={arXiv},
primaryClass={cs},
SLACcitation={%%CITATION=1106.5273;%%}
}
We present a 0.5 Petaflop/s calculation of homogeneous isotropic turbulence in a cube of 2048^3 particles, using a highly parallel fast multipole method (FMM) using 2048 GPUs on the TSUBAME 2.0 system. We compare this particle-based code with a spectral DNS code under the same calculation condition and the same machine. The results of our particle-based turbulence simulation match quantitatively with that of the spectral method. The calculation time for one time step is approximately 30 seconds for both methods; this result shows that the scalability of the FMM starts to become an advantage over FFT-based methods beyond 2000 GPUs.
June 28, 2011 by hgpu
