I'm currently embarking on another "teaching SLURM new tricks" project, and
won't be working on the semi-auto re-MIG for a while.

We've long had a "Level 3" (DUA-governed) multi-tenant environment for
people who have datasets that we can't administratively allow on our main
cluster. Until now its had very limited resources available to the tenants
(N x VMs, where more VMs provide more compute).

We're now starting a project to develop a system in which we have,
per-tenant, one VM to land on as a login node, and another to run
slurmctld/mariadbd/slurmdbd. Outside the tenant environment we'll have an
"uber-SLURM" cluster that fields requests from a broker process sitting
between it and the tenant SLURMs, and allocates resources for them,
making them available within the tenant's VLAN/subnet space. This is almost
like the uber-SLURM is a cloud resource manager, but rather than running
jobs in its cloud it's only plucking the resources from it to push down
into the tenant for a period of time. The tenant will have cloud node
resources defined in its SLURM configuration.

At least, that's the idea in broad strokes.

On Fri, Feb 6, 2026 at 6:55 PM Davide DelVento <[email protected]>
wrote:

> Thanks for sharing, this is very good to know.
> Having to drain the whole node is obviously not ideal, but still much
> better than not being able to do dynamic MIG at all, so perhaps I'll give
> it a try after I talk to the users to better understand their priorities.
> I'll keep you posted about what I find out, please do likewise if you end
> up playing with it.
> Thanks and have a great weekend
>
> On Fri, Feb 6, 2026 at 4:16 PM Fulcomer, Sam <[email protected]>
> wrote:
>
>> I actually spent a bit of time in the SLURM booth at SC discussing this
>> (and also frequently hanging out in their comfy chairs - easy times on the
>> bad hip).
>>
>> This is on the back burner for us. The basic problem is that SLURM
>> doesn't have a mechanism to drain a GPU; rather, the entire node has to be
>> drained to make changes. That's the easy description of the problem. There
>> may be ways to do it within the current capabilities of SLURM, but we
>> haven't picked up that effort in earnest, yet...
>>
>> We do find some occasional issues in nvml control of reconfiguring MIG on
>> multi-GPU systems, where our scripts occasionally fail on one of more GPUs,
>> and they need to be manually reconfigured after that (but that's something
>> in the nvidia driver, presumably). We either run nodes un-MIGed or split
>> into two nominally equal slices. We're currently defaulting to using MIG on
>> half of our B200s and all of our Max-Qs.
>>
>> Being able to drain a single GPU would obviously be great.
>>
>> On Fri, Feb 6, 2026 at 5:07 PM Davide DelVento via slurm-users <
>> [email protected]> wrote:
>>
>>> Aaron (or anyone else),
>>>
>>> Did you manage to get Dynamic MIG working in Slurm? I'm actually
>>> surprised that after these many years SchedMD has not implemented this
>>> feature yet, especially now that newer GPUs allow MIG repartitioning
>>> without being root. The only mention of this in their ticketing system is
>>> at https://support.schedmd.com/show_bug.cgi?id=11091#c8 (and subsequent
>>> c10) which say that it's not on their roadmap, but that was 5 years ago.
>>>
>>> I have heard that some users manage dynamic changes by draining nodes,
>>> running scripts to reconfigure MIG via nvidia-smi, bringing the node back
>>> and then submitting the job. Anybody here has tried that and with what
>>> success?
>>>
>>> I speculate that now that NVIDIA owns SchedMD perhaps this feature will
>>> be at a higher priority, but maybe not? Anybody knows anything about it and
>>> is not bound by an NDA to keep mum?
>>>
>>> Thanks
>>>
>>>
>>> On Wed, Nov 22, 2023 at 1:22 PM Davide DelVento <
>>> [email protected]> wrote:
>>>
>>>> I assume you mean the sentence about dynamic MIG at
>>>> https://slurm.schedmd.com/gres.html#MIG_Management
>>>> Could it be supported? I think so, but only if one of their paying
>>>> customers (that could be you) asks for it.
>>>>
>>>> On Wed, Nov 22, 2023 at 11:24 AM Aaron Kollmann <
>>>> [email protected]> wrote:
>>>>
>>>>> Hello All,
>>>>>
>>>>> I am currently working in a research project and we are trying to find
>>>>> out whether we can use NVIDIAs multi-instance GPU (MIG) dynamically in
>>>>> SLURM.
>>>>>
>>>>> For instance:
>>>>>
>>>>> - a user requests a job and wants a GPU but none is available
>>>>>
>>>>> - now SLURM will reconfigure a MIG GPU to create a partition (e.g.
>>>>> 1g.5gb) which becomes available and allocated immediately
>>>>>
>>>>> I can already reconfigure MIG + SLURM within a few seconds to start
>>>>> jobs on newly partitioned resources, but Jobs get killed when I restart
>>>>> slurmd on nodes with a changed MIG config. (see script example below)
>>>>>
>>>>> *Do you think it is possible to develop a plugin or change SLURM to
>>>>> the extent that dynamic MIG will be supported one day? *
>>>>>
>>>>> (The website says it is not supported)
>>>>>
>>>>>
>>>>>
>>>>> Best
>>>>>
>>>>> - Aaron
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> #!/usr/bin/bash
>>>>>
>>>>> # Generate Start Config
>>>>> killall slurmd
>>>>> killall slurmctld
>>>>> nvidia-smi mig -dci
>>>>> nvidia-smi mig -dgi
>>>>> nvidia-smi mig -cgi 19,14,5 -i 0 -C
>>>>> nvidia-smi mig -cgi 0 -i 1 -C
>>>>> cp -f ./slurm-19145-0.conf /etc/slurm/slurm.conf
>>>>> slurmd -c
>>>>> slurmctld -c
>>>>> sleep 5
>>>>>
>>>>> # Start a running and a pending job (the first job gets killed by
>>>>> slurm)
>>>>> srun -w gx06 -c 2 --mem 1G --gres=gpu:a100_1g.5gb:1 sleep 300 &
>>>>> srun -w gx06 -c 2 --mem 1G --gres=gpu:a100_1g.5gb:1 sleep 300 &
>>>>> sleep 5
>>>>>
>>>>> # Simulate MIG Config Change
>>>>> nvidia-smi mig -i 1 -dci
>>>>> nvidia-smi mig -i 1 -dgi
>>>>> nvidia-smi mig -cgi 19,14,5 -i 1 -C
>>>>> cp -f ./slurm-2x19145.conf /etc/slurm/slurm.conf
>>>>> killall slurmd
>>>>> killall slurmctld
>>>>> slurmd
>>>>> slurmctld
>>>>>
>>>>
>>> --
>>> slurm-users mailing list -- [email protected]
>>> To unsubscribe send an email to [email protected]
>>>
>>
-- 
slurm-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to