[slurm-users] Help with binding GPUs to sockets (NVlink, P2P)

Luis Altenkort altenkort at physik.uni-bielefeld.de
Fri Jun 28 09:29:00 UTC 2019

thanks for the answer. We actually had this already set up correctly, I 
simply forgot to add #SBATCH --sockets-per-node=1 to my script. Now 
--gpus-per-socket works!

Am 28.06.19 um 09:27 schrieb Daniel Vecerka:
> Hi,
>  I'm not sure how it works in 19.0.5, but with 18.x  it's possible to 
> specify CPU affinity in the file  /etc/slurm/gres.conf
> Name=gpu Type=v100 File=/dev/nvidia0 CPUs=0-17,36-53
> Name=gpu Type=v100 File=/dev/nvidia1 CPUs=0-17,36-53
> Name=gpu Type=v100 File=/dev/nvidia2 CPUs=18-35,54-71
> Name=gpu Type=v100 File=/dev/nvidia3 CPUs=18-35,54-71
> The CPUs number you can get with command:
> nvidia-smi topo -m
>         GPU0    GPU1    GPU2    GPU3    mlx5_0  CPU Affinity
> GPU0     X      NV2     NV2     NV2     NODE    0-17,36-53
> GPU1    NV2      X      NV2     NV2     NODE    0-17,36-53
> GPU2    NV2     NV2      X      NV2     SYS     18-35,54-71
> GPU3    NV2     NV2     NV2      X      SYS     18-35,54-71
> mlx5_0  NODE    NODE    SYS     SYS      X
> Best regards, Daniel
> On 27.06.2019 15:17, Luis Altenkort wrote:
>> Hello everyone,
>> I have several nodes with 2 sockets each and 4 GPUs per Socket (i.e. 
>> 8 GPUs per bode). I now want to tell SLURM that GPUs with device ID 
>> 0,1,2,3 are connected to socket 0 and GPUs 4,5,6,7 are connected to 
>> socket 1. I want to do this in order to be able to use the new 
>> command --gpus-per-socket. All GPUs on one socket are directly linked 
>> via NVlink and use P2P communication. In the end I want to be able to 
>> run multi-gpu jobs on GPUs that are all on one socket (and not 
>> distributed accross sockets or nodes). What do I have to change in my 
>> slurm.conf and how would I submit jobs? Like this?:
>> #!/bin/bash
>> #SBATCH --job-name=test
>> #SBATCH --partition=volta
>> #SBATCH --gpus-per-socket=4
>> #SBATCH --ntasks=1
>> #SBATCH --cpus-per-task=1
>> Slurm is on version 19.05.0.
>> Thanks in advance!

More information about the slurm-users mailing list