Slow performance running as a Slurm job

[ricalessandri]

I have modified the alanine-dipeptide metadynamics example to try to bias the system only along 1 CV (instead of the 2 dihedrals used in the example). Namely, after chatting with @pabloferz , I modified the following 2 lines:

PySAGES/examples/openmm/metad/alanine-dipeptide.py

Line 101 in 8ede30b

cvs = [DihedralAngle([4, 6, 8, 14]), DihedralAngle([6, 8, 14, 16])]

PySAGES/examples/openmm/metad/alanine-dipeptide.py

Line 111 in 8ede30b

grid = pysages.Grid(lower=(-pi, -pi), upper=(pi, pi), shape=(50, 50), periodic=True)

To, respectively:

cvs = [DihedralAngle([4, 6, 8, 14])]

grid = pysages.Grid(lower=(-pi, ), upper=(pi, ), shape=(50, ), periodic=True)

However, doing this brings down the performance a lot. For 1e5 steps (0.2 ns), in the 2D case of the example you get:

Completed the simulation in 94.5706 seconds.

While if I use that 1D case I get:

Completed the simulation in 878.7742 seconds.

which is almost a factor of 10 slower (independently of midway3 hardware, etc.).
Do you know what's up / what I am missing?
Have you tried anything along these lines perhaps @sivadasetty ?

Update(s):

• @sivadasetty sees "normal" performance (i.e., 1D faster than 2D) on midway2
• I tried also setting grid to None:

  method = Metadynamics(cvs, height, sigma, stride, ngauss, deltaT=deltaT, kB=kB, grid=None)
  

  which has no impact, i.e., the 2x CV case still takes around 90 sec and the 1x CV case still around 900 sec (for 1e5 steps).

[ricalessandri]

Well, it turns out this is a matter of having the following line:

#SBATCH --cpus-per-gpu=2

in the .sbatch script.
For future reference, this is what one should have the submission script:

#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-gpu=2
#SBATCH --partition=gpu