Lammps with multi GPU

[Lewis-YL]

Hello,

When I tried lammps with either:
mpirun -np 2 lmp -i in.lammps
or
horovodrun -np 2 lmp -i in.lammps

It creates 2 job tasks on 2 GPU and it runs at the same speed as with 1 GPU.
Here is a screen shot.

Do you know how to run a lammps on multi GPU?

[njzjz]

This should be expected if your GPU utilization is low. You may need to know the working mode of GPUs...

[Lewis-YL]

I am using 2 RTX3080Ti, deepmd v2.1.3 with cuda 11.6

Can you tell more about how to run lammps with multi GPU?

I think the speed should be 2X or at least 1.5X.

However, the program just created 2 same jobs on 2 GPUs and did not increase the speed at all.

When the GPU was working, the utilization was always 90%. The mode was always P2 under working and P8 idle.
Is this normal?

[njzjz]

How many atoms are there in the system? On your screenshot, I see that only ~400 MB of GPU memory is used by a single lmp process...

[njzjz]

Note that utilization in nvidia-smi doesn't mean how many cores are used actively. It's "percent of time over the past sample period during which one or more kernels was executing on the GPU".

[Lewis-YL]

I uploaded the new screenshot of when running the command:

horovodrun -np 2 lmp -i in.lammps

I have 104 atoms in my system.

You can see in the screenshot that there are 2 same runs on the 2 GPUs (id: 77634, 77635).

[njzjz]

The system of 104 atoms are too small to utilize all GPU cores.

You may firstly test on a single GPU card. Increase the number of atoms and see how the timing scales.