14 February 2018

647. Linda/G16 and ecce + slurm

I had the opportunity to test G16 with Linda on my cluster. Suprisingly, I observed a decent reduction in computational time for geometry optimisation, and a fair reduction for vibrational analysis.

Computational time -- geometry optimisation
1x6 cores: 100%
2x6 cores: 56% (2 nodes)
3x6 cores: 44% (3 nodes)

Computational time -- normal mode computation
1x6 cores: 100%
2x6 cores: 63% (2 nodes)
3x6 cores: 57% (3 nodes)

In no case was the speed up good enough that it actually makes long-term sense to use Linda (need less than 50% for 2 nodes and 33% or less for 3 nodes), but if you're itching for quick results, it might be worth it.

Either way, as part of this I wrote a CONFIG.Linda for ECCE and SLURM:
Gaussian-16: /opt/gaussian/g16a/g16/g16 perlPath: /usr/bin/ qmgrPath: /usr/bin/ xappsPath: /usr/bin/ Slurm { #!/bin/csh #SBATCH -p linda #SBATCH --time=$walltime #SBATCH --output=slurm.out #SBATCH --job-name=$submitFile #SBATCH --nodes=$nodes #SBATCH --cpus-per-task=$ppn } Gaussian-16FilesToRemove{ core } Gaussian-16Command{ set path = ( /opt/nbo6/bin $path ) set nnodes = `scontrol show job $SLURM_JOBID|grep NodeList|sed 's/=/\t/'|gawk '{print $2}'|tail -n 1` sed -i "2s/^/%LindaWorkers=$nnodes\n/" g16.g16in sed -i '3s/^/%UseSSH\n/' g16.g16in sed -i "s/nprocshared.*/nprocshared=$ppn./" g16.g16in setenv GAUSS_SCRDIR /home/me/scratch setenv GAUSS_EXEDIR /opt/gaussian/g16a/g16 setenv GAUSS_ARCHDIR /opt/gaussian/g16a/g16/arch setenv GAUSS_BSDDIR /opt/gaussian/g16a/g16/bsd setenv GAUSS_LEXEDIR /opt/gaussian/g16a/g16/linda-exe /opt/gaussian/g16a/g16/g16< $infile > $outfile }
The slurm partition linda has all the nodes that has g16.

No comments:

Post a Comment