Exploring System Architectures For Next-Generation Cfd Simulations In The Postpeta-Scale Era

CFD simulations with uniform grids have been paid attention as a next-generation CFD simulation on a large-scale supercomputing system. The Building-Cube Method (BCM) is one of the next-generation CFD methods. The basic idea is to balance loads of calculations among processing elements on a supercomputing system by dividing the whole calculations into many parallel tasks with the same amount of computation. Thus, it is suitable for highly parallel computation on supercomputing systems. This paper firstly implements BCM on five supercomputing systems as an example of a next-generation CFD simulation in the upcoming postpeta-scale era. Then, by theoretical analyses and performance evaluations, this paper clarifies the requirements of future supercomputing systems for a next-generation CFD simulation. The performance evaluations show that as the number of processing elements increases, the imbalance of data exchanges among nodes becomes more serious than that of calculations even in a next-generation CFD simulation. While the calculation time can ideally be reduced according to the number of processing elements, the data transfer time becomes dominant in the total execution time. Different from the massively-parallel system architecture, the number of nodes in a system should be as small as possible to prevent the data transfer. The performance analyses also show that the memory bandwidth limits the performance of BCM and use of an on-chip memory is effective to improve the performance. A memory subsystem that achieves a higher sustained memory bandwidth is required. Therefore, a supercomputing system that consists of a small number of high-performance nodes is essential to achieve high sustained performance of the next-generation CFD in the upcoming postpeta-scale era by reducing the data transfers, which becomes eventually a bottleneck in large-scale simulation.