SC14 New Orleans, LA

The International Conference for High Performance Computing, Networking, Storage and Analysis

Performance Optimization and Evaluation of a Global Climate Application Using a 440m Horizontal Mesh on the K Computer.

Authors: Masaaki Terai (RIKEN), Hisashi Yashiro (RIKEN), Kiyotaka Sakamoto (Fujitsu), Shin-ichi Iga (RIKEN), Hirofumi Tomita (RIKEN), Masaki Satoh (University of Tokyo), Kazuo Minami (RIKEN)

Abstract: Using almost all the compute nodes on the K computer, we evaluated the performance of NICAM, a climate application for a high-resolution global atmospheric simulation study. First, based on profiling results, we improved single node performance using kernel programs extracted from the application, resulting in increases in the peak performance ratio of the kernel programs. Thereafter, we applied the optimization results to the application, and evaluated the performance by weak scaling. In the evaluation, to retain the same problem size per node, we reduced the horizontal mesh size as the number of nodes increased, decreasing it to 0.44km with 81,920 nodes. Although we achieved good scalability of the dynamical step, the elapsed time of the physical step increased with the number of nodes owing to inter-node load imbalances in the cloud microphysics. Finally, the peak performance ratio of the main loop of NICAM reached 8.3% with 81,920 nodes.

Poster: pdf
Two-page extended abstract: pdf

Poster Index