Incremental caffeination of a terrestrial hydrological modeling framework using Fortran 2018 teams

Abstract

We present Fortran 2018 teams (grouped processes) running a parallel ensemble of simulations built from a pre-existing Message Passing Interface (MPI) application. A challenge arises around the Fortran standard's eschewing any direct reference to lower-level communication substrates, such as MPI, leaving any interoperability between Fortran's parallel programming model, Coarray Fortran (CAF), and the supporting substrate to the quality of the compiler implmentation. Our approach introduces CAF incrementally, a process we term "caffeination." By letting CAF initiate execution and exposing the underlying MPI communicator to the original application code, we create a one-to-one correspondence between MPI group colors and Fortran teams. We apply our approach to the National Center for Atmospheric Research (NCAR)'s Weather Research and Forcecasting Hydrological Model (WRF-Hydro). The newly caffeinated main program replaces batch job submission scripts and forms teams that each execute one ensemble member. To support this work, we developed the first compiler front-end and parallel runtime library support for teams. This paper describes the required modifications to a public GNU Compiler Collection (GCC) fork, an OpenCoarrays [1] application binary interface (ABI) branch, and a WRF-Hydro branch.