Mdrun , running simulation on over 10 GPUS machines

GROMACS version: 18
GROMACS modification: No

I have got access to computer having 16 GPUs where I cannot install jobs scheduler such as slurm or PBS.

I wrote a small script to submit a mdrun instance on each GPU
gmx mdrun -s md.tpr -c md.gro -gpu_id 0 -nt 4 &
gmx mdrun -s md.tpr -c md.gro -gpu_id 1 -nt 4 &

everything is fine until I reach -gpu_id 10 where gromacs recognise 10 as 1 & 0 , 11 as 1&1 12 as 1&2 ect …

I have tried with quotes ‘10’ , “10” but mdrun continue to see it as a list …
how can I use the gpu with ids over 10 ?

Thank you very much

sorry already figured out

gmx mdrun -s md.tpr -c md.gro -gpu_id 0, -nt 4 &
gmx mdrun -s md.tpr -c md.gro -gpu_id 1, -nt 4 &

gmx mdrun -s md.tpr -c md.gro -gpu_id 10, -nt 4 &
gmx mdrun -s md.tpr -c md.gro -gpu_id 11, -nt 4 &

seems to force the ‘,’ as a list delimiter

sorry for posting too fast, but was not really intuitive…

For your information, you can run many trajectories on one computer using -multidir option:

Note that launching many mdrun instances without pinning them to a set of cores can lead to considerable performance loss. For that reason, you should either pin using manual offsets (see the second to last example here) or use the mdrun multidir functionality as suggested by @zhmurov which will ca handle automatically the placement of each mdrun instance it launches.

@zhmurov Might actually be applied in my case and haven’t thought of it, even if I have used it before, so thanks! Howerer I would need mdrun_mpi, which I am unsure I can get or not. (On my workstation or typical cluster wouldn’t be a problem. )

@pszilard thanks for your note. I will definitely have a closer look at that !