Need Help Running DeePMD-Kit Quick Start Tutorial on HPC Using SLURM

[agvent]

Hello,

I am attempting to run the DeePMD-Kit Quick Start Tutorial on an HPC. My HPC uses SLURM, so I need to create a SLURM script. I cannot just run "dp" in the command line as shown in the tutorial. I am seeking help duplicating the results of the tutorial to ensure I have the proper setup on my HPC. I am currently running into issues where the tutorial case is running, but is running significantly slower than expected. In particular, I am seeking clarification with the following aspects of my SLURM script so I can properly replicate the tutorial case (I cannot find this information in the documentation or on the discussion board):

- How many processors should I use? How many processors per CPU should I use? Do I need to request memory? If so, how much memory should I request? and Anything else I need to know to properly format my SLURM script.

Here is what I have done so far:

1. I have successfully installed DeePMD-Kit and all required dependencies into a new python environment using miniconda and conda forge. I am using the CPU version of DeePMD-Kit, (Version 3.0.1).
2. I have downloaded the example case into my home directory and placed it into a new folder. That is the data I am using to run the tutorial.

I am currently using 32 processors on a single node. When I submit my job, it runs, but it takes significantly longer than the example case.

Here are my results:

As you can see this is significantly slower than the approximately 20 second wall time shown in the tutorial for the same part.

Here is the beginning output of my case for reference:

here is my SLURM script for reference:

#!/bin/bash

#SBATCH --ntasks=32
#SBATCH --nodes=1
#SBATCH --time=00:30:00
#SBATCH --job-name="Deepmd Tutorial Case Run 6"
#SBATCH --mem=500GB

echo "working directory = "$SLURM_SUBMIT_DIR
 
source /home/user/miniconda3/bin/activate deepmd_cpu

mpirun -np 32 dp train input.json

echo "All Done!"

Thank you in advance for any help that can be provided, it is much appreciated :)

[njzjz]

1. I am using the CPU version of DeePMD-Kit

The tutorial uses a GPU.

[agvent]

Thank you for the quick response. There appears to be some confusion on my end because based on this output included with the tutorial on the documentation, it appears to state that the build variant is the CPU version which is why I thought the tutorial used the CPU only version. Please see the screenshot below.

I apologize if I had misinterpreted the tutorial. My main purpose of this question is to determine what I need to include in my SLURM script (i.e. how many tasks?, how many CPUS per task? do I need to set environment variables?, etc) so that I can properly replicate the tutorial case on my HPC which requires the use of SLURM and to ensure I have DeePMD-Kit installed properly.

Thank you again for your time, it is very much appreciated.

[njzjz]

Sorry I didn't realize it.

[njzjz]

You shouldn't use the MPI when using the dp train on CPUs of a single machine.