- slurm-users - lists.schedmd.com

GPU fallen of the bus
by Ratnasamy, Fritz 27 May '26

27 May '26

Hello, We are noticing that some of the gpus on a specific node have "fallen of the bus". We would like to remove this specific gpu from the slurm scheduler. For example, let's say GPU0 has fallen off the bus, we would need the rest of the GPU1-8 to be available and make GPU0 not able to be allocated. How can we achieve that? I have read about blacklist on the slurm forum but it seems there is no satisfying solution. Best, *Fritz Ratnasamy*Data Scientist Information Technology

3 2

Slurm version 26.05 is now available
by Marshall Garey 26 May '26

26 May '26

We are pleased to announce the availability of Slurm 26.05. The release notes summarizing the new features, and including links to the corresponding documentation, can be found at: https://slurm.schedmd.com/release_notes.html A more extensive list of changes are available in the CHANGELOG: https://github.com/SchedMD/slurm/blob/slurm-26.05/CHANGELOG/slurm-26.05.md The Slurm documentation has also been updated to the 26.05 release: https://slurm.schedmd.com Slurm can be downloaded from: https://github.com/SchedMD/slurm/releases -- Marshall Garey Release Management, Support, and Development SchedMD LLC - Commercial Slurm Development and Support

1 0

Core Specialization Feature
by Dan Healy 21 May '26

21 May '26

I'm curious how to report to our users that the number of available CPU cores on a given node are *(CoresPerSocket * SocketsPerBoard * ThreadsPerCore) - CpuSpecCount*. In a basic example, if the node has 32 cores and CpuSpecCount=1, then there's actually 31 cores available. If a user runs `srun -n32 -N1 bash` he/she will be denied for resources. Then, If a user runs `sinfo -O cpus,nodes` shows there's 32 cores available and the user is confused. How can the user clearly identify that in the above case, the exact number of available CPU cores available are 31 and *not* 32? -- Thanks, Daniel Healy

2 2

cgroup error in SLURM version 25.11.2
by bradford.kerr＠srs.gov 19 May '26

19 May '26

I'm trying to install and run SLURM on a stand-alone (no network access) cluster of four PCs running Red Hat Linux 9. On the first compute node of the cluster, I can not get past the error message "cannot create cgroup context for cgroup/v2". It appears that file cgroup_v2.so does not exist. How do I get past this error?

1 0

We are pleased to announce the availability of Slurm versions 25.11.6 and 25.05.8.
by Marshall Garey 14 May '26

14 May '26

We are pleased to announce the availability of Slurm versions 25.11.6 and 25.05.8. Fixes in these releases primarily cover various stability issues when reconfiguring or shutting down and improved support for newer libraries and compilers. The 25.11.6 release also includes a performance fix for scheduling with topology/block and a regression from 25.11.4 that could cause a "No records archived...but some found" message when archiving job data. The full list of changes are available in the CHANGELOG files for each version: https://github.com/SchedMD/slurm/blob/slurm-25.11/CHANGELOG/slurm-25.11.md https://github.com/SchedMD/slurm/blob/slurm-25.05/CHANGELOG/slurm-25.05.md Slurm can be downloaded from: https://github.com/SchedMD/slurm/releases -- Marshall Garey Senior System Software Engineer NVIDIA

1 0

inconsistent 'Requested node configuration is not available' error
by Paul Raines 14 May '26

14 May '26

In my cluster right now we have: $ sinfo ... pubgpu-req up 7-00:00:00 7 mix cobra,l40s-[01,03],luisa,rtx-03,sevilla,shelob pubgpu-req up 7-00:00:00 5 idle fiber,glaurung,gothmog,l40s-02,leo The following works fine: $ srun -p pubgpu-req -A sysadm --nodelist=cobra,l40s-02 --gres=gpu:1 -N 1 \ --ntasks-per-node=1 --mem=1G --time=1:00:00 --cpus-per-task=4 --pty /bin/bash srun: tres_per_node => gres/gpu:1 cobra[0]:~$ exit exit However just change the order of the nodelist and you get $ srun -p pubgpu-req -A sysadm --nodelist=l40s-02,cobra --gres=gpu:1 -N 1 \ --ntasks-per-node=1 --mem=1G --time=1:00:00 --cpus-per-task=4 --pty /bin/bash srun: tres_per_node => gres/gpu:1 srun: error: Unable to create step for job 8255390: Requested node configuration is not available More experimentation with this and it appears nodelist HAS to be in alphabetical order $ srun -p pubgpu-req -A sysadm --nodelist=l40s-02,leo,fiber \ --gres=gpu:1 -N 1 --ntasks-per-node=1 --mem=1G --time=1:00:00 \ --cpus-per-task=4 --pty /bin/bash srun: tres_per_node => gres/gpu:1 srun: error: Unable to create step for job 8255401: Requested node configuration is not available $ srun -p pubgpu-req -A sysadm --nodelist=fiber,l40s-02,leo \ --gres=gpu:1 -N 1 --ntasks-per-node=1 --mem=1G --time=1:00:00 \ --cpus-per-task=4 --pty /bin/bash srun: tres_per_node => gres/gpu:1 fiber[0]:~$ Surely there is no good reason for this? --------------------------------------------------------------- Paul Raines http://help.nmr.mgh.harvard.edu MGH/MIT/HMS Athinoula A. Martinos Center for Biomedical Imaging 149 (2301) 13th Street Charlestown, MA 02129 USA The information in this e-mail is intended only for the person to whom it is addressed. If you believe this e-mail was sent to you in error and the e-mail contains patient information, please contact the Mass General Brigham Compliance HelpLine at https://www.massgeneralbrigham.org/complianceline <https://www.massgeneralbrigham.org/complianceline> . Please note that this e-mail is not secure (encrypted). If you do not wish to continue communication over unencrypted e-mail, please notify the sender of this message immediately. Continuing to send or respond to e-mail after receiving this message means you understand and accept this risk and wish to continue to communicate over unencrypted e-mail.

2 1

SLURM config option to not tie up a host completely.
by Ron Gould 08 May '26

08 May '26

I have a user asking about a SLURM config setting that would allow a job to not tie up a node's unused resources, thereby allowing another job to run concurrently. User: On our cluster, our Fluent GPU jobs each use only 1 of the 2 GPUs on GPUServer[1,2]. Currently the node is fully allocated to whichever job lands on it first, leaving the second GPU, 38 CPU cores, and 270+ GB of RAM idle. Has anyone dealt with this? What options would facilitate this? Are there any gotchas or pitfalls?

4 3

Moving a Slurm database from one slurm database host to another
by James A Allsopp 08 May '26

08 May '26

Hi, We'd like to move our Slurm database from its existing machine running mariadb-server-5.5.68 to another running mariadb-server-10.5.29-3, both of which are running slurm 24.11.7. We dumped the database successfully from the old server, reimported it into the new server using mysqldump and *mysql < db_dump.sql*. The new slurmdbd service was then started. The clusters have different names so there were a few inconsistencies, but when we started the new slurm cluster, we could see the jobs according to sacct (using the -M option) matched those of the original cluster, but none of the associated job steps were shown. I manually logged into the new database and could see the jobs in the job_table but it looked like the foreign key link had been broken in the step table as there were entries in that table. Am I missing some information about how to export and import properly? I was wondering if the import had triggered auto-increment on keys and that was breaking things? Ideally, we'd like to change the name of the cluster as well. Is that just a matter of altering the prefix on the relevant tables, which is something I can do by hand on the database dump before reimporting? For instance, changing the table <old_name>_event_table to <new_cluster_name>_event_table. Thanks James -- Dr James Allsopp | Research Platforms team IT Services | University of Sheffield I do not work Tuesday and Wednesday 3-5pm. I also send emails at odd times, but respond at your convenience.

2 1

Slurm release candidate version 26.05.0rc1 is available for testing
by Tim Wickberg 07 May '26

07 May '26

We are pleased to announce the availability of Slurm release candidate 26.05.0rc1. To highlight some new features coming in 26.05: - New "srun --async" step mode that submits step processes to stepmgr to queue and eventually launch. This avoids issues with keeping a large number of srun processes backgrounded to queue up future step-based workflows. (This was previously described as "mini-batch" in the Slurm roadmap presentations.) - New topology/ring and topology/torus3d topology plugins. These implement a single-dimensional and three-dimensional topology respectively. - Slurm REST API - Add support for creating/updating/removing partitions, and for viewing the active slurmctld and slurmdbd configurations. - Dynamic Memory Resizing. A job can now release memory (and have the cgroup limits updated) using 'scontrol update' while running. A new "sbatch --mem-update=<margin>@<delay>" option can also automatically reduce the memory limit to the current usage plus a given margin percentage after a specified time. - Add topology-based sorting for node ranks when using dynamic nodes with the topology plugins. This is also available generically for topology/flat (no topology) with a new alpha_step_rank option. - Add an optimized single-node path through the scheduling logic for increased performance. - Expanded the openmetrics (Prometheus) nodes, partitions, and jobs endpoints with gpu allocation statistics. This is the first release candidate of the upcoming 26.05 release series, and represents the end of development for this release, and a finalization of the RPC and state file formats. If any issues are identified with this release candidate, please report them through https://support.schedmd.com against the 26.05.x version and we will address them before the first production 26.05.0 release is made. Please note that the release candidates are not intended for production use. A preview of the updated documentation can be found at https://slurm.schedmd.com/archive/slurm-master/ . Slurm can be downloaded from https://github.com/SchedMD/slurm/releases. The changelog for 26.05 can be found here: https://github.com/SchedMD/slurm/blob/master/CHANGELOG/slurm-26.05.md -- Tim Wickberg Director, System Software, Slurm + Slinky, NVIDIA

1 0

cgroup/v2 regression in 25.11.2 — memory.peak read disabled after first task
by Emyr James 06 May '26

06 May '26

Dear all, Heads-up for anyone running 25.11.2 with cgroup/v2 + jobacct_gather/cgroup: there's a regression in the recent series that effectively disables memory.peak accounting for every task after the first one per slurmstepd. The commit is f2d05c08e6 ("cgroup/v2 - Remove static qualifier from variable", ticket 23646, cherry-picked from e1835a1602). It dropped static from memory_peak_interface in cgroup_p_task_get_acct_data while leaving interfaces_checked static. The function now looks like: bool memory_peak_interface = false; static bool interfaces_checked = false; ... if (!interfaces_checked) { memory_peak_interface = cgroup_p_has_feature(CG_MEMCG_PEAK); interfaces_checked = true; } ... if (memory_peak_interface) { common_cgroup_get_param(..., "memory.peak", ...); } interfaces_checked persists across calls; memory_peak_interface no longer does. So the feature-probe block runs exactly once. From the second call onward memory_peak_interface is a fresh stack-local set to false, and the memory.peak read in the second if is skipped. stats->memory_peak ends up INFINITE64 for every task except the first one ever sampled. User-visible effect: MaxRSS in sacct stops reflecting the kernel-tracked peak and falls back to polling memory.current — which, at typical jobacct_gather intervals, routinely misses short-lived peaks. Which is of course exactly the problem memory.peak was added to solve (see the 2024 thread that prompted the original implementation: https://groups.google.com/g/slurm-users/c/BX1_4bztrJA) One-line fix is to put static back on memory_peak_interface. Or, arguably more robust, drop the interfaces_checked cache entirely and run the feature probe on every call — the cost is one access()-equivalent per accounting poll, basically free. he original v2 implementation in ba34862a40 was great, thanks for that — this looks like a small static-qualifier slip on the way through a cleanup. Cheers, Emyr James Head of Scientific IT CRG - Centre for Genomic Regulation

1 1

slurmdbd and slurmctld prevent alma9 login
by Pols, Maarten 06 May '26

06 May '26

Dear Community, Our Slurm cluster has been running without any issues for several months on version 25.05.3. Last Friday, we experienced a power outage which required us to restart the server. After the restart, we were unable to log in to the master node. Eventually, we managed to access the system via a safe mode workaround. Through a process of elimination, we identified the slurmdbd and slurmctld services as the root cause of the issue. Would you happen to have any idea what might have caused this behavior? We have since upgraded to version 25.11.5, which appears to be running smoothly. However, we would still like to understand the underlying cause of the problem. Thank you in advance for your help. Kind regards, Maarten

3 3

Jobs canceling when nodes become unreachable – need guidance
by Pharthiphan Asokan 05 May '26

05 May '26

Hi, We’re seeing an issue where jobs submitted via salloc are automatically cancelled when a compute node becomes temporarily unreachable. Our goal is to keep jobs pending or requeued instead of being cancelled outright when a node drops offline briefly Slurm sometimes cancels the job rather than requeuing it when the node is marked DOWN/ DRAIN/ DRAINING*. Is there a recommended configuration or additional parameter that ensures jobs remain pending/requeued until the node returns, rather than being cancelled? Any insights or examples from similar setups would be greatly appreciated. Regards, Pharthiphan

3 2

How to track solver usage in Slurm
by fernando.seguro＠rtx.com 23 Apr '26

23 Apr '26

I don't see an easy way of tracking this in slurm. I tried using the --comment="slover" directive, but this gets mapped to the "derived_es" column in the database, which is reserved for job info on errors, from what I read. This would make a good enhancement request. Where would I submit that? Something like: SBATCH --solver=fluent or SBATCH --solver=LSDyna This would be very helpful for tracking solver usage without having to use the job-name. Great for reporting. Please let me know if someone else has requested this previously. If anyone has figured out a way to do it, please share. thanks

4 3

Can't clear the REASON
by Berg, Stephen P CIV USN NRL DET SSC MS (USA) 17 Apr '26

17 Apr '26

Just upgraded my grid to 25.11 a couple days ago. Got two nodes that were down for a bit and when I set them to resume and set the reason="" they are still showing up with the reason I had set while they were down. Nothing I do seems to unset that reason flag for these two nodes. I've set the nodes to down with a reason and then back to idle with the blank reason but it never goes back to "none" like the other nodes. Whatever the last reason set was persists. The behavior seems different since I updated to 25.11. Was running the 20.11 available from the epel-8 repos previously.

2 4

Jobs aborting after slurmctld reload on Intel nodes - AMD unaffected
by Pharthiphan Asokan 16 Apr '26

16 Apr '26

Hi team, We’re observing job aborts on Intel-based nodes immediately after a slurmctld reload. AMD nodes remain stable and jobs continue unaffected. No system or Slurm configuration changes were made before the issue started. Error observed: error: Aborting JobID=1288 due to change in socket/core configuration of allocated nodes Relevant node configuration (Intel node example): NodeName=smc-h4-u19 CPUs=128 Boards=1 SocketsPerBoard=2 CoresPerSocket=64 ThreadsPerCore=1 RealMemory=510976 State=UNKNOWN Feature=model_SSG-1228-B From logs: error: valid_job_resources: smc-h4-u19 sockets:2:2, cores 64,32 error: Node configuration differs from hardware: CPUs=128:128(hw) Boards=1:1(hw) SocketsPerBoard=2:2(hw) What we’ve verified: * No changes in BIOS, firmware, or hardware topology * No edits to slurm.conf or slurm_nodes.conf * Reload (scontrol reconfigure) triggers job aborts only on Intel nodes * AMD nodes remain intact through reloads

3 2

We are pleased to announce the availability of Slurm version 25.11.5
by Marshall Garey 14 Apr '26

14 Apr '26

We are pleased to announce the availability of Slurm version 25.11.5. Some of the changes in 25.11.0 are highlighted below: * Fix a regression in 25.11.0 that could cause daemons to hang after reaching conmgr_max_connections. * Fix a regression added in 25.05.0 concerning how the slurmctld inherits /run/slurmctld/sack.socket when using AuthType=auth/slurm to prevent clients that connected during a reconfigure from hanging indefinitely. * Fix a race in tree forwarding that could cause a crash during shutdown * Fix several bugs in the namespace/linux plugin that caused instability The full list of changes are available in the CHANGELOG file: https://github.com/SchedMD/slurm/blob/slurm-25.11/CHANGELOG/slurm-25.11.md Slurm can be downloaded from: https://www.schedmd.com/download-slurm/ -- Marshall Garey Senior System Software Engineer NVIDIA

1 0

Why my job can't start (backfill reservation issue)
by Massimo Sgaravatto 14 Apr '26

14 Apr '26

Dear all I (try to) manage a slurm cluster composed by some CPU-only nodes and some worker nodes which have also GPUs: NodeName=cld-ter-[01-06] Sockets=2 CoresPerSocket=96 ThreadsPerCore=2 RealMemory=1536000 State=UNKNOWN NodeName=cld-ter-gpu-[01-05] Sockets=2 CoresPerSocket=96 ThreadsPerCore=2 Gres=gpu:nvidia-h100:4 RealMemory=1536000 State=UNKNOWN The GPU nodes are exposed through multiple partitions: PartitionName=gpus Nodes=cld-ter-gpu-[01-02] State=UP PriorityTier=20 PartitionName=sparch Nodes=cld-ter-gpu-03 AllowAccounts=sparch,operators QoS=sparch State=UP PriorityTier=20 PartitionName=geant4 Nodes=cld-ter-gpu-03 AllowAccounts=geant4,operators QoS=geant4 State=UP PriorityTier=20 PartitionName=enipred Nodes=cld-ter-gpu-04 AllowAccounts=enipred,operators QoS=enipred State=UP PriorityTier=20 PartitionName=enipiml Nodes=cld-ter-gpu-05 AllowAccounts=enipiml,operators QoS=enipiml State=UP PriorityTier=20 We also set a partition to allow cpu-only jobs on the GPU nodes, but these jobs should be preempted (killed and requeued) if jobs submitted to partitions with higher priorities require those resources: PreemptType=preempt/partition_prio PreemptMode=REQUEUE PartitionName=onlycpus-opp Nodes=cld-ter-gpu-[01-05],cld-dfa-gpu-06,btc-dfa-gpu-02 State=UP PriorityTier=10 Now, I don't understand why this job [*] submitted on the onlycpus-opp partition can't start running e.g. on the cld-ter-gpu-01, since it has a lot of free resources: [sgaravat@cld-ter-ui-01 ~]$ scontrol show node cld-ter-gpu-01 NodeName=cld-ter-gpu-01 Arch=x86_64 CoresPerSocket=96 CPUAlloc=8 CPUEfctv=384 CPUTot=384 CPULoad=5.93 AvailableFeatures=(null) ActiveFeatures=(null) Gres=gpu:nvidia-h100:4 NodeAddr=cld-ter-gpu-01 NodeHostName=cld-ter-gpu-01 Version=25.11.3 OS=Linux 5.14.0-611.45.1.el9_7.x86_64 #1 SMP PREEMPT_DYNAMIC Wed Apr 1 05:56:53 EDT 2026 RealMemory=1536000 AllocMem=560000 FreeMem=1192357 Sockets=2 Boards=1 State=MIXED+PLANNED ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A Partitions=gpus,onlycpus-opp BootTime=2026-04-09T10:39:35 SlurmdStartTime=2026-04-09T10:40:01 LastBusyTime=2026-04-09T11:54:46 ResumeAfterTime=None CfgTRES=cpu=384,mem=1500G,billing=839,gres/gpu=4,gres/gpu:nvidia-h100=4 AllocTRES=cpu=8,mem=560000M,gres/gpu=4,gres/gpu:nvidia-h100=4 CurrentWatts=0 AveWatts=0 I guess the "MIXED+PLANNED" is the answer, but as far as I can see only a job (283469) is planned for this worker node: sgaravat@cld-ter-ui-01 ~]$ squeue --start | grep ter-gpu-01 JOBID PARTITION NAME USER ST START_TIME NODES SCHEDNODES NODELIST(REASON) 283469 gpus vllm-pod ciangott PD 2026-04-13T14:31:40 1 cld-ter-gpu-01 (Resources) But job 283469 doesn't require too many resources [**], so the 2 jobs could run together. Why job 283534 can't start ? Any hints ? Thanks, Massimo [*] [sgaravat@cld-ter-ui-01 ~]$ scontrol show job=283534 JobId=283534 JobName=myscript.sh UserId=sgaravat(5008) GroupId=tbadmin(5001) MCS_label=N/A Priority=542954 Nice=0 Account=operators QOS=normal JobState=RUNNING Reason=None Dependency=(null) Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0 RunTime=00:00:41 TimeLimit=1-00:00:00 TimeMin=N/A SubmitTime=2026-04-13T11:10:13 EligibleTime=2026-04-13T11:10:13 AccrueTime=2026-04-13T11:10:13 StartTime=2026-04-13T11:58:39 EndTime=2026-04-14T11:58:39 Deadline=N/A PreemptEligibleTime=2026-04-13T11:58:39 PreemptTime=None SuspendTime=None SecsPreSuspend=0 LastSchedEval=2026-04-13T11:58:39 Scheduler=Backfill Partition=onlycpus-opp AllocNode:Sid=cld-ter-ui-01:3035857 ReqNodeList=(null) ExcNodeList=(null) NodeList=btc-dfa-gpu-02 BatchHost=btc-dfa-gpu-02 NumNodes=1 NumCPUs=1 NumTasks=1 CPUs/Task=1 ReqB:S:C:T=0:0:*:* ReqTRES=cpu=1,mem=100G,node=1,billing=26 AllocTRES=cpu=1,mem=100G,node=1,billing=26 Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=* MinCPUsNode=1 MinMemoryNode=100G MinTmpDiskNode=0 Features=(null) DelayBoot=00:00:00 OverSubscribe=OK Contiguous=0 Licenses=(null) LicensesAlloc=(null) Network=(null) Command=/shared/home/sgaravat/myscript.sh SubmitLine=sbatch myscript.sh WorkDir=/shared/home/sgaravat StdErr=/shared/home/sgaravat/JOB-myscript.sh.283534.4294967294.err StdIn=/dev/null StdOut=/shared/home/sgaravat/JOB-myscript.sh.283534.4294967294.out MailUser=massimo.sgaravatto(a)pd.infn.it MailType=INVALID_DEPEND,BEGIN,END,FAIL,REQUEUE,STAGE_OUT [**] sgaravat@cld-ter-ui-01 ~]$ scontrol show job=283469 JobId=283469 JobName=vllm-pod UserId=ciangott(6054) GroupId=tbuser(6000) MCS_label=N/A Priority=499703 Nice=0 Account=cms QOS=normal JobState=PENDING Reason=Resources Dependency=(null) Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0 RunTime=00:00:00 TimeLimit=1-00:00:00 TimeMin=N/A SubmitTime=2026-04-13T06:48:37 EligibleTime=2026-04-13T06:48:37 AccrueTime=2026-04-13T06:48:37 StartTime=2026-04-13T14:31:40 EndTime=2026-04-14T14:31:40 Deadline=N/A SuspendTime=None SecsPreSuspend=0 LastSchedEval=2026-04-13T11:59:48 Scheduler=Main Partition=gpus AllocNode:Sid=cld-ter-ui-01:3015801 ReqNodeList=(null) ExcNodeList=(null) NodeList= SchedNodeList=cld-ter-gpu-01 NumNodes=1-1 NumCPUs=32 NumTasks=1 CPUs/Task=32 ReqB:S:C:T=0:0:*:* ReqTRES=cpu=32,mem=190734M,node=1,billing=118,gres/gpu=2,gres/gpu:nvidia-h100=2 AllocTRES=(null) Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=* MinCPUsNode=32 MinMemoryNode=190734M MinTmpDiskNode=0 Features=(null) DelayBoot=00:00:00 OverSubscribe=OK Contiguous=0 Licenses=(null) LicensesAlloc=(null) Network=(null) Command=.interlink/jobs/default-0c0257f8-d1ea-4135-a602-96c229ce8516/job.slurm SubmitLine=sbatch .interlink/jobs/default-0c0257f8-d1ea-4135-a602-96c229ce8516/job.slurm WorkDir=/shared/home/ciangott StdErr= StdIn=/dev/null StdOut=/shared/home/ciangott/.interlink/jobs/default-0c0257f8-d1ea-4135-a602-96c229ce8516/job.out TresPerNode=gres/gpu:nvidia-h100:2 TresPerTask=cpu=32

5 10

pam_slurm_adopt
by Ratnasamy, Fritz 09 Apr '26

09 Apr '26

Hi, We have blocked ssh to nodes where users do not have a running job. However, when there is a running job, a user can ssh to that node and use more than the slurm resources allocated on that node. Am I missing something? I usually request an interactive session of 2GB. I then open another terminal and ssh to the node where the interactive session runs. I allocated a 4GB R variable on the node where I just ssh but this does not error out. Additionally, if I end my interactive session, the ssh connection still lives which I believe should not happen. Best, *Fritz Ratnasamy*Data Scientist Information Technology

7 6

How to delete my defaultwckey ?
by Faraz Hussain 07 Apr '26

07 Apr '26

I want every submitted job to have some value for the wckey, i.e: #SBATCH --wckey=myproject I made the appropriate changes to slurm.conf and slurmdb.conf and it works great. I can track how many hours people are using with those wckeys. But now I want to make it mandatory to use a wckey. To do that I need to delete the default wckey associated with the user's account. I tried doing it as follows, but it still lets me submit jobs without a wckey. It probably thinks I have an "empty" default wckey. sacctmgr mod user fhussa set defaultwckey= [root@mas01 ~]# sacctmgr list user fhussa format=user,defaultwckey User Def WCKey ---------- ---------- fhussa

1 0

Difference between sreport and sacct
by hany.ibrahim＠kit.edu 07 Apr '26

07 Apr '26

Hello, I am trying to understand how sreport work. So I run sreport and obtain th estatistics about speciic project. On teh other side, I collected the jobs related to this account using sacct. When I try to sum the numbers, sometimes it don't sum up. I can't see any reservations. This is the sreport I used: sreport --tres cpu cluster AccountUtilizationByUser Start=2021-01-01 account=<account> -t hours -np This is jobs I collected by using sacct: sudo -u slurm sacct -a -S "2021-01-01" -E NOW --format=JobID%100,User,Account%100,State,AllocCPUs,CPUTime%100,CPUTIMERAW%100,ElapsedRaw%100,AllocTres%100,AllocNodes,NodeList,NNodes --noheader --parsable2 | grep <account> I tried to exclude the steps from the main Jobs, so I summed only the main jobs whatever its state (canceled, completed, timeout, fail..). However the sreport occasionally is larger. What I am missing exactly? Warm regards, Hany.

1 0

Shared queue: how to request full node with all resources?
by Dustin Lang 31 Mar '26

31 Mar '26

Hi, With a partition with "OverSubscribe=FORCE" set, is there a way to request all the node resources? I see "--mem=0" does that for memory. But I do not see an option to request all the CPUs and GRES/TRES such as GPUs. I tried "--nodes=1 --ntasks=1 --cpus-per-task=0", but "--cpus-per-task=0" does not do the same thing as "--mem=0". In other words, is it possible to have a queue where nodes can be shared between jobs, but have a simple way for an sbatch script to request the full node with all its memory, cpus, and other resources? We have a heterogeneous cluster, so telling users (or jupyterhub scripts) to list exactly the resources they want doesn't really work. thanks, dustin

3 4

How do you handle GPU node failures during long jobs?
by Antonio Jose Alonso-Stepanov 26 Mar '26

26 Mar '26

Hi all, I'm a Stanford CS student looking into how sites handle GPU node failures during long-running jobs. A couple questions: When a GPU node goes down mid-job, do most sites use Slurm's requeue or --no-kill to handle it, or is it mostly manual drain and resubmit? Is anyone using HealthCheckProgram to catch GPU issues (like ECC errors via DCGM), or do you handle GPU health monitoring outside of Slurm? Curious what's worked and what hasn't. Thanks. Antonio

3 2

_node_config_validate: gres/gpu: Count changed on node (0 != 2)
by Xaver Stiensmeier 23 Mar '26

23 Mar '26

Hey Slurm-users list, while our regular gpu nodes are working fine, our on demand gpu nodes have a weird issue. They power up, I can ssh onto them and execute nvidia-smi on them without issue, but they are marked invalid and slurmctld logs _node_config_validate: gres/gpu: Count changed on node (0 != 2) however, node show shows that the gpus are recognized and the gres.conf are stored on the worker nodes as expected and the node entries in the slurm.conf are fine, too: # slurm.conf NodeName=my_worker_node SocketsPerBoard=16 CoresPerSocket=1 RealMemory=64075 MemSpecLimit=4000 State=CLOUD Gres=gpu:L4:2 # openstack # gres.conf on my_worker_node ubuntu@my_node:~$ cat /etc/slurm/gres.conf # GRES CONFIG Name=gpu Type=L4 File=/dev/nvidia0 Name=gpu Type=L4 File=/dev/nvidia1 Thankful for any ideas and debugging ideas. Best, Xaver PS: By executing: sudo scontrol update NodeName=$(bibiname 0) Gres= sudo scontrol reconfigure sudo scontrol update NodeName=$(bibiname 0) state=RESUME reason=None the node can be resumed. However, this is not a real solution.

2 3

We are pleased to announce the availability of Slurm versions 25.11.4 and 25.05.7.
by Marshall Garey 12 Mar '26

12 Mar '26

We are pleased to announce the availability of Slurm versions 25.11.4 and 25.05.7. Changes in both versions include preventing SIGHUP from terminating slurmrestd. Changes in 25.11.4 include fixing several issues with the switch/nvidia_imex plugin, several memory leaks, a race that could cause jobs to be stuck in the completing state or be requeued and held if the --get-user-env option was used, flex reservations taking nodes from other reservations that do not request full nodes, and a variety of other issues. Changes in 25.05.7 include fixing the gpu/nrt plugin, and fixing some other regressions and stability issues. The full list of changes are available in the CHANGELOG files for each version: https://github.com/SchedMD/slurm/blob/slurm-25.11/CHANGELOG/slurm-25.11.md https://github.com/SchedMD/slurm/blob/slurm-25.05/CHANGELOG/slurm-25.05.md Slurm can be downloaded from: https://www.schedmd.com/download-slurm/ -- Marshall Garey Senior System Software Engineer NVIDIA

1 0

Server with two GPUs sharing each GPU to different partitions
by Gestió Servidors 09 Mar '26

09 Mar '26

Hello, First of all, sorry if my question is about something easy, but in my environment, this is the first time I need to do this. I have one server that have 2 identical GPUs. I have two different partitions: one is accessible from some allocation nodes and the other from other allocation nodes. My question is if it's possible to share one GPU with one partition and the other GPU with the other partition. For example, something like this: * gres.conf: NodeName=gpu-node AutoDetect=off Name=gpu Type=RTX5070 File=/dev/nvidia0 Cores=0-127 NodeName=gpu-node AutoDetect=off Name=gpu Type=RTX5070 File=/dev/nvidia1 Cores=128-255 * slurm.conf: NodeName=gpu-node CPUs=256 SocketsPerBoard=2 CoresPerSocket=64 ThreadsPerCore=2 RealMemory=260000 TmpDisk=47000 Gres= RTX5070:2 PartitionName=gpu-int.q Nodes=gpu-node OverSubscribe=No State=UP AllocNodes=submit-node-1 PriorityTier=3 GraceTime=900 PartitionName=gpu-ext.q Nodes=gpu-node OverSubscribe=No State=UP AllocNodes=submit-node-1 PriorityTier=3 GraceTime=900 With this configuration, if simultaneously two different users tries to execute in one of the RTX5070 (one in gpu-int.q and the other in gpu-ext.q), could both of them execute with no problems? If user from gpu-int.q uses device #0 and user from gpu-ext.q uses device #1, could SLURM work without problems if a second simultaneously exection assigns device #1 to user from gpu-int.q and device #0 to user from gpu-ext.q? Thanks.

1 0