Parallel computing for high-resolution/large-scale flood simulation using the K supercomputer

Abstract

This paper reports the implementation of high-performance computing using the K supercomputer in Kobe, Japan, for large-scale/high-resolution flood simulation. Supercomputer K was developed in 2012 by RIKEN and Fujitsu and ranked first in the list of Top 500 supercomputer sites in 2011 during its development stage. A two-dimensional inundation simulation model developed based on a shallow water equation using an existing numerical scheme was parallelized with the K supercomputer. Osaka and other cities along the Yodo River were chosen as application sites and the area discretized by 12824442 (= 3453 × 3714) nodes with a resolution of 10 m. The computational time for the five-hour flood simulation was measured by changing the number of 8-core CPUs of the K supercomputer. As a result, computational time was decreased to 9.3 min by using 128 × 64 = 8192 8-core CPUs. The computational time was 1423.7 min for one 8-core CPU. Thus, the simulation speed increased by a factor of 153.2 with the use of the K supercomputer.