The joint research team of Professor Takayuki Aoki of Tokyo Institute of Technology International Center for Science and Technology, Naoya Maruyama, Team Leader, Program Composition Model Research Team, RIKEN Center for Computational Science, and Research Fellow Mohamed Wahib will be held on November 11th. Announced that it won the Best Paper Award at SC18, the world's best international conference on performance computing.
Computer simulation is widely used in society.In particular, by using a large-scale computer such as a supercomputer, it becomes possible to perform higher-precision and higher-definition simulations at high speed.On the other hand, the speed of improvement of computer performance is gradually slowing down, and it is said that improvement of the simulation method itself will be more important than ever in order to further improve the accuracy and speed of simulation.
The adaptive mesh refinement method is a type of grid method that is one of the methods frequently used in simulation, and it is said that the required calculation and memory usage can be significantly reduced compared to the normal grid method.However, there are many problems in actually using the adaptive mesh refinement method.Therefore, although the adaptive mesh refinement method is a promising method in principle, its actual use is limited as compared with the ordinary grid method.
This time, the joint research team has developed a new software technology that solves the problems of the adaptive mesh refinement method, and realized an environment where the adaptive mesh refinement method can be easily used on a large-scale supercomputer.When the developed software technology was used on a supercomputer, it succeeded in automatically performing high-speed and large-scale simulations using a large number of GPUs of 1,000 units at the same time.This point was highly evaluated in this SC16, and it won the Best Paper Award.
The joint research team will continue to improve and improve the developed method, and at the same time, research and develop software technology that will reduce the development cost of simulation software in future supercomputers by various methods, not limited to the adaptive mesh refinement method. I plan to go.