====== Gaussian on DARWIN ======

<WRAP center round important 60%>
The University of Delaware owns a site license to Gaussian software. Only cluster accounts using a UDelNet Id are covered by that license and have access to the software described on this page.  All users -- including HPC guest and ACCESS accounts -- are free to [[abstract:darwin:install_software:workgroup-sw|install their own copies]] of Gaussian software, either from [[https://gaussian.com/g16/g16_plat.pdf|binary packages]] or by compiling from source (Portland Group compilers are available to all cluster users).
</WRAP>

A Slurm job submission script template for batch Gaussian jobs is available at ''/opt/shared/templates/slurm/applications/gaussian.qs''.  Copy and edit the template based on your job's resource needs; the comments in the template describe each Slurm option and environment variable present.  

<note tip>It is a good idea to periodically check ''/opt/shared/templates/slurm/applications'' for changes in existing templates, or the addition of new templates, designed to provide the best performance for a particular version of Gaussian on that cluster.</note>

At the very least, the Gaussian input file variable must be altered accordingly:

<code>
#
# Gaussian input file:
#
GAUSSIAN_INPUT_FILE="h2o.com"
</code>

The template sources an external script to complete the setup of the job environment, including fixup of the input file to account for CPU counts and memory limits imposed by Slurm, and specification of CPU and GPU bindings necessary for Gaussian '16 to utilize GPU offload.

===== GPU Offload =====

Gaussian '16 introduced the ability to offload some computation to GPU devices.  Each GPU associated with the job //must// have a corresponding CPU dedicated to controlling the GPU and moving data to and from it; thus, OpenMP processor bindings are a necessary part of starting ''g16'' executable.

The Gaussian job environment script, part of the DARWIN Slurm templates package (''/opt/shared/slurm/templates/libexec/gaussian.sh''), includes code to produce the necessary PGI OpenMP CPU bindings and CPU-to-GPU bindings strings.  These values are communicated to Gaussian via the environment variables ''GAUSS_CDEF'' and ''GAUSS_GDEF'', respectively.

<note important>By default, the job environment script removes any ''nproc'', ''nprocshared'', ''cpu'', and ''gpucpu'' route 0 lines from the input file (as part of the input file fixup).  The number of CPUs is then inferred by Gaussian from ''GAUSS_CDEF''.  Setting ''GAUSSIAN_SKIP_INPUT_FIXUP'' will use the input file verbatim, which could yield errors if those route 0 directives are present.</note>

To use GPU offload, first ensure the job script is adding a GPU-enabled version of Gaussian to the job environment:

<code>
#
# Add Gaussian to the environment:
#
vpkg_require gaussian/g16b01:gpu
</code>

Next, request GPU partition, either in a ''#SBATCH'' comment in the script or from the command line:

<code bash>
$ sbatch --partition=gpu-v100 --gpus=2 ...
</code>

See [[abstract:darwin:runjobs:queues#the-gpu-partitions|the GPU partitions]] on DARWIN for details on requesting GPU resources.