The se_a descriptor cannot utilize all cores when trained on the cpu

[gengxingze]

Summary

When training the dp_sea model on a 64core cpu-only node, the cpu utilization is only about 1800%. How to set it up to further increase the cpu utilization?

DeePMD-kit Version

2.2.10

Backend and its version

v2

Python Version, CUDA Version, GCC Version, LAMMPS Version, etc

python=3.10.13

Details

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
34016 gengxz 20 0 16.1g 636216 102264 S 1779 0.2 1268:54 python

[njzjz]

the cpu utilization is only about 180%.

Isn't it 1800%?

[gengxingze]

the cpu utilization is only about 180%.

Isn't it 1800%?

I apologize for the correction, but 1800% is still only 18 cores working.

[njzjz]

In the top program, you can type 1 to see how much each process is utilized.

How do you set the threads?

[gengxingze]

In the top program, you can type 1 to see how much each process is utilized.

How do you set the threads?
#!/bin/bash
#SBATCH --job-name=deepmd
#SBATCH --partition=cpu
#SBATCH --output=job.%j.out
#SBATCH --error=job.%j.err
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=64

begin=date +%s
echo "======== Job starts at date +'%Y-%m-%d %T' ======== "

module load mkl mpi compiler
conda activate deepmd

dp train input.json
echo "======== Job ends at date +'%Y-%m-%d %T' ======== "
final=date +%s
total=$(($final-$begin))
day=$((total/86400))
hour=$((total%86400/3600))
minute=$((total%3600/60))
second=$((total%60))
echo totaltime: $day days, ${hour}:${minute}:${second} | tee -a out

[njzjz]

From the top I can see that all cores have been used.