2026-03-09T06:39:33Z https://ipsj.ixsq.nii.ac.jp/oai

oai:ipsj.ixsq.nii.ac.jp:00237618 2025-01-19T08:49:26Z 1164:1579:11464:11703

Acceleration of Data Assimilation for Rainfall Prediction with Co-design of Hardware and Software Acceleration of Data Assimilation for Rainfall Prediction with Co-design of Hardware and Software Sameer, Deshmukh Arata, Amemiya Takumi, Honda Elmor, Lang Ian Sameer, Deshmukh Arata, Amemiya Takumi, Honda Elmor, Lang Ian ビックデータ処理・信頼性 Rainfall forecasting is an important application of scientific computing that has a far-reaching impact on humanity. Accurate and timely forecasts can make the difference between life and death for people affected by severe weather conditions. High quality forecasts depend mainly upon availability of data and computational resources. The MP-PAWR at Saitama University generates high quality distributions of 3D radar reflectivity at 30 second intervals. The data from this radar has been used to generate rainfall density forecasts using the SCALE-LETKF algorithm. The SCALE-LETKF algorithm is very inefficient on currently available architectures, utilizing only about 5% of the available computational capacity of Fugaku. Improvement in the computational efficiency of SCALE-LETKF can hold tremendous value for the accuracy, speed and energy efficiency of rainfall forecasts. LETKF from SCALE-LETKF occupies a majority of the computational resources. LETKF is designed for processing large amounts of data, and has been written over several years, comprising more than 20,000 lines of FORTRAN code. We propose a mini-application that imitates the most important performance and numerical characteristics of LETKF. We then propose an alternative hardware accelerator that can potentially achieve higher computational efficiency than currently available hardware. Rainfall forecasting is an important application of scientific computing that has a far-reaching impact on humanity. Accurate and timely forecasts can make the difference between life and death for people affected by severe weather conditions. High quality forecasts depend mainly upon availability of data and computational resources. The MP-PAWR at Saitama University generates high quality distributions of 3D radar reflectivity at 30 second intervals. The data from this radar has been used to generate rainfall density forecasts using the SCALE-LETKF algorithm. The SCALE-LETKF algorithm is very inefficient on currently available architectures, utilizing only about 5% of the available computational capacity of Fugaku. Improvement in the computational efficiency of SCALE-LETKF can hold tremendous value for the accuracy, speed and energy efficiency of rainfall forecasts. LETKF from SCALE-LETKF occupies a majority of the computational resources. LETKF is designed for processing large amounts of data, and has been written over several years, comprising more than 20,000 lines of FORTRAN code. We propose a mini-application that imitates the most important performance and numerical characteristics of LETKF. We then propose an alternative hardware accelerator that can potentially achieve higher computational efficiency than currently available hardware. technical report 情報処理学会 2024-08-01 application/pdf 研究報告システム・アーキテクチャ（ARC） 24 2024-ARC-258 1 6 2188-8574 AN10096105 https://ipsj.ixsq.nii.ac.jp/record/237618/files/IPSJ-ARC24258024.pdf eng