The 1000-Member Ensemble Kalman Filtering with the JMA Nonhydrostatic Mesoscale Model on the K Computer

The ensemble Kalman filter (EnKF) approximates background error covariance by using a finite number of ensemble members. Although increasing the ensemble size consistently improves the EnKF analysis, typical applications of the EnKF to realistic atmospheric simulations are conducted with a small ensemble size due to limited computational resources. The finite ensemble size introduces a sampling error into the background error covariance, leading to a degradation of the accuracy of the analysis fields. As a representative of EnKF applications, a local ensemble transform Kalman filter (LETKF) was implemented on the K computer, the flagship supercomputer in Japan, which enables demanding computations with larger ensembles. This study investigated the performance of the LETKF on the K computer and evaluated the influence of sampling noise on the background error covariance estimated from 1000-member ensemble forecasting with the Japan Meteorological Agency nonhydrostatic model covering Japan with a 15-km horizontal resolution. The LETKF on the K computer achieved a high peak performance ratio of 14.7 % without special optimization, showing the suitability of the LETKF for high-performance parallel computing. The background error covariance estimated from 1000 ensemble members contained negligible sampling noise even at distant locations without covariance localization. The results indicated that for the case in the current study, an ensemble size of 500 would be large enough to approximate the error covariance under the configuration with a horizontal resolution of 15 km. The results also suggest that improving input/output performance will become a primary goal in the design of next-generation supercomputers.