DocumentCode :
659248
Title :
Real-world computer simulation by using a GPU-rich supercomputer
Author :
Aoki, Toyohiro
Author_Institution :
Global Sci. Inf. & Comput. Centre, Tokyo, Japan
fYear :
2013
fDate :
13-14 Sept. 2013
Abstract :
Real-world simulations based on fluid dynamics require huge amount of computer resources. Only by large-scale computation on world top-class supercomputers the computational demands can be satisfied. We demonstrate three applications carried out on the whole TSUBAME 2.0 system, which has 2.4 PFLOPS of the peak performance with more than 4,000 GPUs (Graphics Processing Unit) at the Tokyo Institute of Technology. The upgrade of TSUBAME is on-going and the performance will be improved to 5.7 PFLOPS (double precision) and 17 PFLOPS (single precision). Since there are more than 2,000 small processing cores in a single GPU chip, it is necessary to make programs based on data-parallel algorithm to extract high performance from the GPU. First, we show high resolution meso-scale atmosphere model ASUCA that is being developed by the Japan Meteorological Agency (JMA) for the purpose of the next-generation weather forecasting service. We have succeeded in a weather prediction with 500-m resolution (cf. Current JMA weather forecast uses 5km mesh). We also show an air flow simulation for 10 km × 10 km area of a central part of metropolitan Tokyo with 1-m resolution by Lattice Boltzmann method, in which a new Large-Eddy Simulation model has been employed. Finally we talk about a phase-field simulation to develop new materials by studying the dendritic solidification of Al-Si alloy. We achieved 2.0 PFLOPS in single precision, which is 44.5% of the peak performance. The GPU-based TSUBAME has high electrical power efficiency and we have obtained simulation results with less energy.
Keywords :
geophysics computing; graphics processing units; lattice Boltzmann methods; parallel algorithms; parallel machines; solidification; weather forecasting; 17 PFLOPS; 2.0 PFLOPS; 2.4 PFLOPS; 5.7 PFLOPS; ASUCA; GPU-rich Supercomputer; JMA; Japan Meteorological Agency; TSUBAME 2.0 system; Tokyo Institute of Technology; air flow simulation; aluminum silicon alloy; computational demands; computer resources; data-parallel algorithm; dendritic solidification; electrical power efficiency; fluid dynamics; graphics processing unit; high performance extract; high resolution meso-scale atmosphere model; large-eddy simulation model; lattice Boltzmann method; metropolitan Tokyo; next-generation weather forecasting service; peak performance; phase-field simulation; real-world computer simulation; single GPU chip; small processing cores; weather prediction; Abstracts; Atmospheric modeling; Computational modeling; Computer simulation; Graphics processing units; Supercomputers; Weather forecasting;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Emerging Trends and Applications in Computer Science (ICETACS), 2013 1st International Conference on
Conference_Location :
Shillong
Print_ISBN :
978-1-4673-5249-9
Type :
conf
DOI :
10.1109/ICETACS.2013.6691381
Filename :
6691381
Link To Document :
بازگشت