[slurm-users] how to locate the problem when slurm failed to restrict gpu usage of user jobs

[taleintervenor at sjtu.edu.cn]

Hi, all:

 

We found a problem that slurm job with argument such as --gres gpu:1 didn't
be restricted with gpu usage, user still can see all gpu card on allocated
nodes.

Our gpu node has 4 cards with their gres.conf to be:

> cat /etc/slurm/gres.conf

Name=gpu Type=NVlink_A100_40GB File=/dev/nvidia0 CPUs=0-15

Name=gpu Type=NVlink_A100_40GB File=/dev/nvidia1 CPUs=16-31

Name=gpu Type=NVlink_A100_40GB File=/dev/nvidia2 CPUs=32-47

Name=gpu Type=NVlink_A100_40GB File=/dev/nvidia3 CPUs=48-63

 

And for test, we submit simple job batch like:

#!/bin/bash

#SBATCH --job-name=test

#SBATCH --partition=a100

#SBATCH --nodes=1

#SBATCH --ntasks=6

#SBATCH --gres=gpu:1

#SBATCH --reservation="gpu test"

hostname

nvidia-smi

echo end

 

Then in the out file the nvidia-smi showed all 4 gpu cards. But we expect to
see only 1 allocated gpu card.

 

Official document of slurm said it will set CUDA_VISIBLE_DEVICES env var to
restrict the gpu card available to user. But we didn't find such variable
exists in job environment. We only confirmed it do exist in prolog script
environment by adding debug command "echo $CUDA_VISIBLE_DEVICES" to slurm
prolog script.

 

So how do slurm co-operate with nvidia tools to make job user only see its
allocated gpu card? What is the requirement on nvidia gpu drivers, CUDA
toolkit or any other part to help slurm correctly restrict the gpu usage?

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20220323/2ee538a2/attachment.htm>