Topic/Type: 2.1 MHD, EHD & other fluid methods, Poster
K. Fukazawa1, T. Umeda2, T. Ogino2
1 Department of Earth and Planetary Sciences, Faculty of Sciences, Kyushu University
2 Solar-Terrestrial Environment Laboratory, Nagoya University
In the recent day, more than 99% of the ?top 500? supercomputer systems in the world adopt scalar processor architecture. However it is often mentioned that the performance of electromagnetic fluid codes is not good. Thus the aim of the present study is to make performance tuning of electromagnetic fluid codes for space plasma simulations on scalar massively-parallel supercomputer systems with various type of scalar processors. In this study we use the T2K open supercomputer at University of Tokyo, SR16000 at Kyushu University, and FX1 and HX600 at Nagoya University.
In this presentation we measure performance of two electromagnetic fluid codes, MHD and Vlasov codes. The three-dimensional MHD model is performed by four different decomposition methods for parallelization, one-, two- and three-dimensional decomposition methods and a cache-tuned three-dimensional decomposition method to find out which method is best for MHD model. As for parallelization, the five-dimensional Vlasov model is parallelized with the two-dimensional configuration space decomposition without the velocity space decomposition. As the results of these runs, we have achieved over 10% of peak performance efficiency with both MHD and Vlasov model on T2K, and much better performances on SR16000, FX1 and HX600. In particular we found that the two-dimensional decomposition of the three-dimensional MHD model is suitable for T2K, and for the other scalar computers three-dimensional decomposition gave the best performance. In this study we will show the results of performance measurements and performance tuning techniques for the MHD and Vlasov models on scalar massively-parallel supercomputers in detail.