ARCHIVES

Weather forecasting typically relies on numerical prediction models that process observational data, such as temperature and humidity, to estimate future conditions. The Korea Meteorological Administration (KMA) uses the UK’s GloSea6 model for this purpose. While effective, running such models—especially for research—requires significant computing power, often exceeding available supercomputer resources. To overcome this, KMA introduced a scaled-down version called Low GloSea6, designed for use on standard research servers. However, Low GloSea6 still demands considerable computational effort, particularly due to its high input/output (I/O) load, which can slow performance. Manual tuning of I/O parameters is inefficient, prompting the need for a smarter solution. This study proposes a machine learning-based approach to optimize both hardware and internal model parameters for Low GloSea6. The process involves collecting performance data using profiling tools and training a machine learning model to predict optimal settings for various research environments. Results demonstrate that the model effectively identifies optimal configurations, achieving only a 16% error rate in execution time prediction. This technique shows promise not just for weather models, but also for enhancing other high-performance computing applications.