Using full GPU node without MPI

User discussions
1

GROMACS version: 2019
GROMACS modification: Yes

Hello,

I am trying to run the gromacs on a multi-instance GPU (MIG). I tried several commands to run Gromacs but each time it used part of the GPU and not the full node.

1. gmx mdrun -v -deffnm md_0_10

2. gmx mdrun -ntmpi 6 -ntomp 12 -nb gpu -pme gpu -npme -1 -bonded gpu -v -deffnm md_0_10

3. gmx mdrun -gputasks 0001 -nb gpu -pme gpu -npme 1 -ntmpi 4 -v -deffnm md_0_10

4. mpirun -np 1 gmx_mpi mdrun -ntomp 6 -nb gpu -gputasks 00 -v -deffnm md_0_10

5. gmx mdrun -ntmpi 2 -nb gpu -pme gpu -bonded gpu -v -deffnm md_0_10

Has anyone faced the same problem earlier? Any help in this regard will be really helpful.

2

Hi,

Can you please clarify what exactly are you trying to accomplish? NVIDIA MIG is used to partition a GPU into multiple slices so independent work can progress concurrently (and in isolation). Are you trying to run many independent jobs on a single node?

Cheers,
Szilárd

3

Hello @pszilard,

I am trying to run a simulation with Gromacs. I would like to run it with a higher speed so one of the platform team members told me to use a full GPU node to optimize the speed.

I just want an optimized speed for the simulation with this MIG architecture.
I hope this clarifies your doubts. Please let me know if it answers your questions.

4

MIG will not improve the performance of a single simulation, it is can be useful for ensemble simulations, but even there the benefits are modest and MPS is more important; see this post from a few years ago: