Application Optimization and Scalability of NICAM-LETKF on Fugaku
DescriptionTechnological trends in supercomputers change year by year, and the bottleneck factors for weather forecasting and climate prediction simulations are also changing accordingly. Weather/climate models are collections of interdisciplinary algorithms with no obvious computational hotspots, and the entire application code must be tuned. Developing and maintaining those models are high-cost and require decision-making among many trade-offs. In the development project of the supercomputer Fugaku, we selected the weather data assimilation experiment using NICAM-LETKF as one of the target problems. This practical benchmark test required performance optimization in all data transfer paths, from processing units to file I/O. Through our system-application co-design activities, we conducted the following optimizations: 1) Distributed file I/O using local SSDs, 2) Efficient inter-node data transposition between parallel multi-node simulations and ensemble data assimilation, 3 ) Elimination of time loss that is difficult to capture by performance profilers, 4) Aggressive use of mixed-precision floating-point arithmetic. These optimizations contributed to the realization of a 3.5 km, 1024-member ensemble data assimilation experiment with the help of Fugaku's key features, such as large memory bandwidth, high thread scalability, eco-mode ALU, and the aid of a special-purpose eigenvalue solver.
TimeMonday, June 2615:00 - 15:30 CEST
LocationSanada I
Event Type
Climate, Weather and Earth Sciences