Performance analysis and optimization of the external mode part of an ocean model#
Masao Kurogi (JAMSTEC)
Abstract#
Compared to the other three-dimensional calculation processes of ocean general circulation models, the strong scaling speedup of the external mode part (two-dimensional calculation process) tends to be low because the data communication cost is relatively high. The external mode part of the ocean model COCO is isolated, and the performance analysis is performed on the CPU nodes (AMD EPYC 7742), VE-equipped nodes (AMD EPYC 7742 + NEC SX-Aurora TSUBASA Type 20B), and GPU-equipped nodes (AMD EPYC 7742 + NVIDIA A100) of the Earth Simulator at JAMSTEC. The optimization method of reducing the frequency of communication by increasing the overlap grids is also evaluated.