使用 slurm 在单个节点上使用 python 进行多处理

Question

I am trying to run some parallel code on a cluster. The cluster uses slurm and my code is in python. The code uses multiple cores when I run it on my own machine. However, when I try to run the code on the cluster it is extremely slow and does not appear to be using multiple cores.

Here is the relevant code from python:

from multiprocessing import Pool

Nz_i=range(1,13)

p=Pool()
p.map(Err_Calc,Nz_i)
p.close()
p.join()

the function Err_Calc is defined earlier on. I don't think its definition is relevant.

The SBATCH I am using to run the code on the cluster is the following:

#!/bin/bash
#SBATCH -N 1
#SBATCH -p RM-shared
#SBATCH --ntasks-per-node 13
#SBATCH -t 03:10:00

module load python/intel_2.7.14

python Err_vs_Nz_Cl.py 

The file Err_vs_Nz_Cl.py contains the code I showed above. I would expect this SBATCH to provide me with 13 cores, but the code seems to be using only 1 core or perhaps is slow for some other reason. Does anyone know what's going wrong?

Answer 1

This may be wrong (I'm a newbie to this), but what happens if you change the --ntasks-per-node 13 argument to --cpus-per-task 13 ? I think the docs say that you need to explicitly specify the number of cpus in this way, else it will only run with one cpu.

Source: https://slurm.schedmd.com/sbatch.html

Answer 2

Because you're running without srun -n (which is correct for multiprocessing with its process-based "threading" model) you need another way to tell it how many CPUs to use. This is done in your sbatch script with --ntasks=1 and --cpus-per-task=13 (or however many cores your node has).

Be cautious with multiprocessing . If you're not running on a whole node, or explicitly specifying the number of cores, it will attempt to run on all visible CPU cores, regardless of whether they've been allocated to you or not!