Container for (spack-manager) CUDA GPU Build of Exawind for NERSC Science Platform

[ajpowelsnl]

The proposed code specifies a spack-manager-based GPU-capable container on NERSC science platforms.

• Use NERSC Container Technology

Build

podman-hpc build --no-cache -t <TAG_NAME> -f Dockerfile-containergpucuda .

Run

podman-hpc run --rm --gpu -it <TAG_NAME>

Expected Output

root@8953e3033a99:/exawind-entry/spack-manager# which exawind
/exawind-entry/spack-manager/snapshots/exawind/containergpucuda/2023-11-03/opt/linux-ubuntu22.04-zen3/gcc-11.4.0/exawind-git.d3c1aa4656fc3c6eccaec8c684671c82a3895172=multiphase-fyborjcfgnruxcks2vwzrc4guyw5toew/bin/exawind
root@8953e3033a99:/exawind-entry/spack-manager# exawind --help
usage: exawind [--awind NPROCS] [--nwind NPROCS] input_file
    -h,--help       Show this help message
    --awind NPROCS      Number of ranks for AMR-Wind (default = all ranks)
    --nwind NPROCS      Number of ranks for Nalu-Wind (default = all ranks)

Helpful Hints

• After exiting a container, you must build the container again using a new tag name.
• If using non-NERSC machines, Docker can be used instead of podman-hpc
• Be aware that spack-manager depends on a pinned version of Spack, a point of configure / build / runtime fragility