Hi Ewan, Thank you for the response. Exactly the source of information I was looking for. The 'slurm-mig-discovery' tool is perfect.
Cheers Tim ________________________________ From: slurm-users <[email protected]> on behalf of Ewan Roche <[email protected]> Sent: Wednesday, 21 April 2021 10:12 To: Slurm User Community List <[email protected]> Cc: [email protected] <[email protected]> Subject: Re: [slurm-users] {Suspected Spam?} SLURM A100 CAUTION: This email originated outside the UCT network. Do not click any links or open attachments unless you know and trust the source. Hi Tim, we have MIG configured and integrated with Slurm using the slurm-mig-discovery tools: https://gitlab.com/nvidia/hpc/slurm-mig-discovery<https://gitlab.com/nvidia/hpc/slurm-mig-discovery> The mig-parted tool is great for setting up MIG itself: https://github.com/NVIDIA/mig-parted<https://github.com/NVIDIA/mig-parted> Once setup MIG instances work fine with Slurm although the output from nvidia-smi is a little different as one sees both GPUs - the “visible device” is the MIG instance:: $ salloc -p interactive -n 1 -c 8 --gres=gpu:1 salloc: Granted job allocation 5235 salloc: Waiting for resource configuration salloc: Nodes gpu001 are ready for job $ env | grep CUDA CUDA_VISIBLE_DEVICES=0 $ nvidia-smi -L GPU 0: A100-PCIE-40GB (UUID: GPU-c1976541-7b00-3f9f-f557-a17f45b879e9) MIG 3g.20gb Device 0: (UUID: MIG-GPU-c1976541-7b00-3f9f-f557-a17f45b879e9/1/0) GPU 1: A100-PCIE-40GB (UUID: GPU-83f9ff5b-09c3-8de1-b3eb-adaadb1cda9f) The caveats are that MIG and the slurm integration is rather static for the moment so it’s not really possible to dynamically change the profiles. The other slight issue is that all combinations of MIG instances waste some compute or memory capacity. We have divided each A100 into two 3g.20gb devices so all the memory is used but 1/7 of the compute capacity is lost. Thanks Ewan Roche Division Calcul et Soutien à la Recherche UNIL | Université de Lausanne On 21 Apr 2021, at 09:14, Timothy Carr <[email protected]<mailto:[email protected]>> wrote: Dear Community, Trust everyone is well and keeping safe? We are considering the purchase of nodes with the Nvidia A100 GPUs and enabling the MIG feature which allows for the creation of instance resource profiles. The creation of these profiles seems to be straightforward as per the documentation. Have any of you had the opportunity to implement the A100 MIG with SLURM and have you found any caveats you are willing to share? Kind Regards -- Tim Disclaimer - University of Cape Town This email is subject to UCT policies and email disclaimer published on our website at http://www.uct.ac.za/main/email-disclaimer<http://www.uct.ac.za/main/email-disclaimer> or obtainable from +27 21 650 9111. If this email is not related to the business of UCT, it is sent by the sender in an individual capacity. Please report security incidents or abuse via https://csirt.uct.ac.za/page/report-an-incident.php<https://csirt.uct.ac.za/page/report-an-incident.php>. Disclaimer - University of Cape Town This email is subject to UCT policies and email disclaimer published on our website at http://www.uct.ac.za/main/email-disclaimer or obtainable from +27 21 650 9111. If this email is not related to the business of UCT, it is sent by the sender in an individual capacity. Please report security incidents or abuse via https://csirt.uct.ac.za/page/report-an-incident.php.
