<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=Windows-1252">
</head>
<body>
<style type="text/css" style="display:none;"><!-- P {margin-top:0;margin-bottom:0;} --></style>
<div id="divtagdefaultwrapper" style="font-size:12pt;color:#000000;font-family:Calibri,Helvetica,sans-serif;" dir="ltr">
<p>Hi,</p>
<p><br>
</p>
<p>if by "share the GPU" you mean exclusive allocation to a single job then, I believe, you are missing cgroup configuration for isolating access to the GPU.</p>
<p><br>
</p>
<p>Below the relevant parts (I believe) of our configuration.</p>
<p><br>
</p>
<p>There also is a way of time- and space-slice GPUs but I guess you should get things setup without slicing.<br>
</p>
<p><br>
</p>
<p>I hope this helps.</p>
<p><br>
</p>
<p>Manuel<br>
</p>
<p><br>
</p>
<p></p>
<div>==> /etc/slurm/cgroup.conf <==<br>
# https://bugs.schedmd.com/show_bug.cgi?id=3701<br>
CgroupMountpoint="/sys/fs/cgroup"<br>
</div>
<div>CgroupAutomount=yes<br>
</div>
<div>AllowedDevicesFile="/etc/slurm/cgroup_allowed_devices_file.conf"<br>
<br>
==> /etc/slurm/cgroup_allowed_devices_file.conf <==<br>
/dev/null<br>
/dev/urandom<br>
/dev/zero<br>
/dev/sda*<br>
/dev/cpu/*/*<br>
/dev/pts/*<br>
/dev/nvidia*<br>
<br>
</div>
<p>==> /etc/slurm/slurm.conf <==</p>
<p></p>
<div>ProctrackType=proctrack/cgroup<br>
<br>
</div>
<div># Memory is enforced via cgroups, so we should not do this here by [*]<br>
#<br>
# /etc/slurm/cgroup.conf: ConstrainRAMSpace=yes<br>
#<br>
# [*] https://bugs.schedmd.com/show_bug.cgi?id=5262<br>
JobAcctGatherParams=NoOverMemoryKill<br>
<br>
</div>
<div>TaskPlugin=task/cgroup<br>
</div>
<div><br>
</div>
JobAcctGatherType=jobacct_gather/cgroup<br>
<p></p>
<p></p>
<p><br>
</p>
<div id="Signature">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div style="font-family:Tahoma; font-size:13px">
<div><font size="2"><span style="font-size:10pt">-- <br>
Dr. Manuel Holtgrewe, Dipl.-Inform.<br>
Bioinformatician<br>
Core Unit Bioinformatics – CUBI<br>
Berlin Institute of Health / Max Delbrück Center for Molecular Medicine in the Helmholtz Association / Charité – Universitätsmedizin Berlin<br>
<br>
Visiting Address: Invalidenstr. 80, 3rd Floor, Room 03 028, 10117 Berlin<br>
Postal Address: Chariteplatz 1, 10117 Berlin<br>
<br>
E-Mail: manuel.holtgrewe@bihealth.de<br>
Phone: +49 30 450 543 607<br>
Fax: +49 30 450 7 543 901<br>
Web: cubi.bihealth.org www.bihealth.org www.mdc-berlin.de www.charite.de</span></font><font size="2"><span style="font-size:10pt">
</span></font></div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
<hr style="display:inline-block;width:98%" tabindex="-1">
<div id="divRplyFwdMsg" dir="ltr"><font face="Calibri, sans-serif" style="font-size:11pt" color="#000000"><b>From:</b> slurm-users <slurm-users-bounces@lists.schedmd.com> on behalf of Analabha Roy <hariseldon99@gmail.com><br>
<b>Sent:</b> Wednesday, February 1, 2023 6:12:40 PM<br>
<b>To:</b> slurm-users@lists.schedmd.com<br>
<b>Subject:</b> [ext] [slurm-users] Enforce gpu usage limits (with GRES?)</font>
<div> </div>
</div>
<div>
<div dir="ltr">Hi,
<div><br>
</div>
<div>I'm new to slurm, so I apologize in advance if my question seems basic.</div>
<div><br>
</div>
<div>I just purchased a single node 'cluster' consisting of one 64-core cpu and an nvidia rtx5k gpu (Turing architecture, I think). The vendor supplied it with ubuntu 20.04 and slurm-wlm 19.05.5. Now I'm trying to adjust the config to suit the needs of my department.</div>
<div><br>
</div>
<div>I'm trying to bone up on GRES scheduling by reading <a href="https://slurm.schedmd.com/gres.html">
this manual page</a>, but am confused about some things.</div>
<div><br>
</div>
<div>My slurm.conf file has the following lines put in it by the vendor:</div>
<div><br>
</div>
<div>###################</div>
# COMPUTE NODES<br>
GresTypes=gpu<br>
NodeName=shavak-DIT400TR-55L CPUs=64 SocketsPerBoard=2 CoresPerSocket=32 ThreadsPerCore=1 RealMemory=95311 Gres=gpu:1<br>
#PartitionName=debug Nodes=ALL Default=YES MaxTime=INFINITE State=UP<br>
<br>
PartitionName=CPU Nodes=ALL Default=Yes MaxTime=INFINITE State=UP<br>
<br>
PartitionName=GPU Nodes=ALL Default=NO MaxTime=INFINITE State=UP<br>
<div>#####################</div>
<div><br>
</div>
<div>So they created two partitions that are essentially identical. Secondly, they put just the following line in gres.conf:</div>
<div><br>
</div>
<div>###################<br>
</div>
<div>NodeName=shavak-DIT400TR-55L Name=gpu File=/dev/nvidia0<br>
</div>
<div>###################<br>
</div>
<div><br>
</div>
<div>That's all. However, this configuration does not appear to constrain anyone in any manner. As a regular user, I can still use srun or sbatch to start GPU jobs from the "CPU partition," and nvidia-smi says that a simple
<a href="https://cupy.dev/">cupy</a> script that multiplies matrices and starts as an sbatch job in the CPU partition can access the gpu just fine. Note that the environment variable "CUDA_VISIBLE_DEVICES" does not appear to be set in any job step. I tested
this by starting an interactive srun shell in both CPU and GPU partition and running ''echo $CUDA_VISIBLE_DEVICES" and got bupkis for both.<br>
</div>
<div><br>
</div>
<div><br>
</div>
<div>What I need to do is constrain jobs to using chunks of GPU Cores/RAM so that multiple jobs can share the GPU. </div>
<div><br>
</div>
<div>As I understand from the gres manpage, simply adding "AutoDetect=nvml" (NVML should be installed with the NVIDIA HPC SDK, right? I installed it with apt-get...) in gres.conf should allow Slurm to detect the GPU's internal specifications automatically.
Is that all, or do I need to config an mps GRES as well? Will that succeed in jailing out the GPU from jobs that don't mention any gres parameters (perhaps by setting CUDA_VISIBLE_DEVICES), or is there any additional config for that? Do I really need that
extra "GPU" partition that the vendor put in for any of this, or is there a way to bind GRES resources to a particular partition in such a way that simply launching jobs in that partition will be enough?</div>
<div><br>
</div>
<div>Thanks for your attention.</div>
<div>Regards</div>
<div>AR</div>
<div><br>
</div>
<div><br>
</div>
<div><br>
</div>
<div><br>
</div>
<div><br>
</div>
<div><br>
</div>
<div><br>
</div>
<div><br>
</div>
<div><br>
</div>
<div> </div>
<div><br>
</div>
<div><br clear="all">
<div><br>
</div>
-- <br>
<div dir="ltr" class="gmail_signature" data-smartmail="gmail_signature">
<div dir="ltr">
<div>
<div dir="ltr">
<div>Analabha Roy<br>
</div>
<div>Assistant Professor</div>
<div><a href="http://www.buruniv.ac.in/academics/department/physics" target="_blank">Department of Physics</a></div>
<div><a href="http://www.buruniv.ac.in/" target="_blank">The University of Burdwan</a></div>
<div>Golapbag Campus, Barddhaman 713104</div>
<div>West Bengal, India</div>
<div>Emails: <a href="mailto:daneel@utexas.edu" target="_blank">daneel@utexas.edu</a>,
<a href="mailto:aroy@phys.buruniv.ac.in" target="_blank">aroy@phys.buruniv.ac.in</a>,
<a href="mailto:hariseldon99@gmail.com" target="_blank">hariseldon99@gmail.com</a><br>
<div><font face="tahoma, sans-serif">Webpage: <a href="http://www.ph.utexas.edu/~daneel/" target="_blank">
http://www.ph.utexas.edu/~daneel/</a></font></div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</div>
</body>
</html>