Increasing the multiscale/multiphysics capability of CAM-SE using implicit time integration and GPU accelerators

Tuesday, 16 December 2014
Rick Archibald1, Katherine J Evans1, Patrick Worley1, Matthew R Norman2, Aaron Lott3, Andrew Salinger4 and Carol S Woodward3, (1)Oak Ridge National Laboratory, Oak Ridge, TN, United States, (2)Oak Ridge National Lab, Oak Ridge, TN, United States, (3)Lawrence Livermore National Laboratory, Livermore, CA, United States, (4)Sandia National Laboratory, Albuquerque, NM, United States
The recent focus on regional refinement in the Community Atmosphere Model (CAM5) has created a strong need to develop time-stepping methods capable of accelerating throughput on high performance computing for climate dynamics across multiple spatial and temporal scales. This research is focused on developing implicit methods that can be executed at scale on GPU based machines. Efforts to port the scalable spectral element dynamical core to incorporate these developments is presented, including both 2D and 3D benchmark test case results. The current implicit solver and preconditioner implementations utilize a Fortran interface package within the Trilinos project, third party software that allows fully tested, optimized, and robust code with a suite of parameter options to be included a priori. Merging this coding strategy with GPU libraries will be discussed along with beneficial optimization gains and data structure requirements to evaluate Trilinos binded residual calculations on GPU processors.