Important notice
Linda is not supported on our system (i.e. only shared-memory parallelisation is supported). All Gaussian jobs must be run on a single node. Invoking the resource -R "span[hosts=1]" ensures this.
Gaussian code performs notoriously bad in highly parallel calculations. There is essentially no speed-up beyond ca. 16 cores.
Wrappers take care of setting the Gaussian environment. No "module load" action is required on behalf of the user.
Since it is a common mistake to set different number of cores in the Gaussian input file (via the Link0 instruction %NProcShared) and in the LSF batch script (via -n), the wrapper ensures both values match. The LSF option overrides the input file value.
A very common file format issue is that Gaussian requires an empty line at the end of the input file.
Benchmark
SLURM Submit script example
For more information use the Job Script Generator.
Tutorial
You can follow this tutorial about geometric optimisation with Gaussian to get hands-on with the program.