12JCSDA Benchmarking and Optimizing the Computational Performance of the Joint Effort for Data Assimilation Integration System in the Unified Forecast System

Efforts are currently underway at the National Centers for Environmental Prediction (NCEP) to replace the Gridpoint Statistical Interpolation (GSI) software at the core of the Global Data Assimilation System (GDAS), used to compute operational analyses, with the Joint Effort for Data Assimilation Integration (JEDI) system, part of a broader transition to the Unified Forecast System (UFS). Though JEDI’s unified framework has promising potential to accelerate the development of cutting-edge research and operations in numerical weather prediction at NCEP, the high computational intensity of data assimilation algorithms, such as the hybrid 4D ensemble variational method used by GDAS, demands careful attention be paid to code optimization. In this presentation, we will discuss our progress at optimizing the atmospheric component of JEDI for the Finite-Volume Cubed Sphere Dynamical Core (FV3) used in UFS. Special attention is paid to scalability across various resolutions and quantities of observations, as well as the relative performance of various error minimization algorithms that underlie the variational analyses. Profiles of scalability and algorithm performance in terms of computational speed and observation error statistics will be presented, recent code changes and improvements will be discussed, and plans and suggestions for future research avenues will be outlined.