Included are scripts for generating Docker images for fv3gfs using either the GNU or Intel compiler collection.
Docker should be already installed on your system to generate the images.
You can compile with GNU or INTEL compiler by executing one of the build scripts. In addition, to give your image a specific repository name, pass the respository name to the scripts as
REPO=noaagsl ./build_gnu_images.sh
Generating the GNU images is straightforward
REPO=noaagsl ./build_gnu_images.sh
This will generate five separate images for mpi+netcdf, ESMF, fv3, NCEPLIBS and GFSv16b. The OS within the container is ubuntu:18.04
You can install a version of mpi you like by modifying build_mpi.sh
Using Intel compiler is a bit more complicated because we chose to use the host compiler in the container as well. Licensing issues coupled with complicated installation procedure within container make using Intel compiler inside a container difficult. This also means the host and container OS should be the same, we use centos:8 for our purpose.
First, set the location of the intel compiler collection in build_intel_images.sh
and scripts/intel_comp.sh
INTEL_COMP_DIR=/data/intel
This directory will be mounted to the container whenever we want to compile something with icc/ifort.
Also modify other parts of scripts/intel_comp.sh
as needed. This script will be sourced before compilation begins.
After these steps, compilation of the images is straightforward
./build_intel_images.sh
You can also build using Intel MPI with GNU compiler with
COMP=gnu ./build_intel_images.sh
Though, currently this breaks down when building HDF5 (will be installed in local directory instead of /usr/local).
Also the Intel build does not work properly with OpenMP -- I suspect it is because of Intel 19.0 version that is probably untested. You can turn off OpenMP in conf/configure.fv3.linux.intel or don't use threads in your forecasts.