Post by Paddy DoyleHi Loris,
Post by Loris BennettHi,
I'm installing GAMESS on our cluster for the first time and find that it
has a slightly crufty, Slurm-unaware start script.
Has anyone already adapted this for Slurm and is willing to share?
It's a tricky one alright!
See attached a custom rungms.slurm that I hacked together for our system...
I had a little bit of fighting to get it to work with Slurm! :) The rungms
script is quite hard-coded with local values to begin with (and it's csh which I
don't like as a scripting language).
Our rungms.slurm script unfortunately still has some hard-coded paths from our
setup, and I used the python hostlist tool, so there's definitely some changes
to be made for your setup.
I compiled GAMESS with Intel 14 and IMPI 4.
I used a bunch of SLURM_* environment variables in the script.
I tried to leave the control of the runtime setup (nodes/cores) in the hands of
the Slurm submission script.
E.g. if the user wanted 16 cores (in our cluster there are 8 per node), they'd
just use
#SBATCH -n 16
..and the rungms.slurm would figure out that it was 16 cores over 2 nodes, and
start 8 on each.
#SBATCH -n 16
#SBATCH -N 4
..and again the rungms.slurm would use the Slurm environment to figure out how
many nodes and how many processes per node.
I used the slurm-aware hydra launcher, which takes care of those details I
think.
mpiexec.hydra -bootstrap slurm ...
Hopefully it will be of use to you.
Thanks,
Paddy
Thanks, Paddy. I had noticed you mention the script on your GAMESS page
at TCHPC, so I was already hoping you might be able to help me out ...
GAMESS seems to have a very strange notion of what 'version' means).