Characteristics-based integration methods for efficient atmospheric simulation on accelerated parallel architectures
Matthew Norman, North Carolina State University
Immense distributed-memory parallelism requires prioritizing the lowering of communication burden in numerical methods, and yet the explicit time steps must be maximized. Focusing on atmospheric simulation, I propose herein various characteristics-based (CB) integration methods. With fully discrete explicit time stepping, CB methods reduce communication and synchronization while allowing large Courant-Friedrichs-Lewy (CFL) time steps. There will be discussion of theory, GPU implementation and efficiency, results, and future work.
Abstract Author(s): Matthew R. Norman, Ramachandran D. Nair, and Fredrick H. M. Semazzi