RRTMGP: A High-Performance Broadband Radiation Code for the Next Decade
Lead PI: Dr. Eli Mlawer, Atmospheric and Environmental Research, Inc.
Start Year: 2014 | Duration: 5 Years
Partners: Atmospheric and Environmental Research, Inc., University of Colorado, NCAR, NRL Monterey
This proposal is to develop a high-performance broadband radiation code for the current generation of computational architectures. This code, called RRTMGP, will be a completely restructured and modern version of the accurate RRTMG radiation code that has been implemented in many General Circulation Models (GCMs) including the Navy Global Environmental Model (NAVGEM), the NCAR Community Earth System Model (CESM), and NOAA’s Global Forecast System (GFS). Our proposed development will significantly lessen a key bottleneck in these highly complex and coupled models, namely the large fraction of computational time currently required for the calculation of radiative fluxes and heating rates. This will allow these models to increase their resolution and/or the complexity of their other physical parameterizations, thereby enabling a large potential increase in model performance. We will preserve the strengths of the existing RRTMG parameterization, especially the high accuracy of the k-distribution treatment of absorption by gases, while rewriting the entire code to provide highly efficient computation across a range of architectures. We will pay special attention to highly-parallel platforms including Graphical Processing Units (GPUs) and Many- Integrated-Core processors (MICs), which preliminary experience shows can accelerate broadband radiation calculations by as much as a factor of fifty. Our redesign will include refactoring the code into discrete kernels corresponding to fundamental computational elements (e.g. gas optics), optimizing the code for operating on multiple columns in parallel, simplifying the subroutine interface, and revisiting the existing gas optics interpolation scheme to reduce branching. Our work will make extensive use of the lessons our team has learned on two recent related efforts, the development of a specialized GPU-accelerated version of RRTMG and a vectorized reimplementation of the code that includes a novel spectral sub-selection algorithm. Representatives of two global models that will benefit greatly from the proposed development, NAVGEM and CESM, are part of our collaboration. Each modeling group will participate in the redesign phase of the project so that the specifics of each computational environment are taken into account in the development of RRTMGP.