- slurm-users - lists.schedmd.com

Slurm suspend preemption not working
by Verma, Nischey (HPC ENG,RAL,LSCI) 15 Mar '24

15 Mar '24

Hi All, we are trying to implement preemption in one of our partitions so we can run priority jobs on it and suspend the ones running on the partition and resume once the priority job is done. We have read through the Slurm documentation and did the configuration, but somehow we can not make it work. We tried other preemption like cancel which works fine, but Suspend isn't working. The jobs with higher priority which should suspend other jobs stays as pending and waiting for resources. We are using a QOS to assign priority to the job and also configuring the QOS so it can preempt certain other QOSs. this is the output of our scontrol show config : AccountingStorageBackupHost = (null) AccountingStorageEnforce = associations,limits,qos AccountingStorageHost = -Some-server- AccountingStorageExternalHost = (null) AccountingStorageParameters = (null) AccountingStoragePort = 6819 AccountingStorageTRES = cpu,mem,energy,node,billing,fs/disk,vmem,pages AccountingStorageType = accounting_storage/slurmdbd AccountingStorageUser = N/A AccountingStoreFlags = job_comment,job_env,job_extra,job_script AcctGatherEnergyType = (null) AcctGatherFilesystemType = (null) AcctGatherInterconnectType = (null) AcctGatherNodeFreq = 0 sec AcctGatherProfileType = (null) AllowSpecResourcesUsage = No AuthAltTypes = (null) AuthAltParameters = (null) AuthInfo = (null) AuthType = auth/munge BatchStartTimeout = 10 sec BcastExclude = /lib,/usr/lib,/lib64,/usr/lib64 BcastParameters = (null) BOOT_TIME = 2024-03-15T13:32:05 BurstBufferType = (null) CliFilterPlugins = (null) ClusterName = cluster CommunicationParameters = (null) CompleteWait = 0 sec CoreSpecPlugin = (null) CpuFreqDef = Unknown CpuFreqGovernors = OnDemand,Performance,UserSpace CredType = cred/munge DebugFlags = (null) DefMemPerCPU = 3500 DependencyParameters = (null) DisableRootJobs = Yes EioTimeout = 60 EnforcePartLimits = ALL Epilog = (null) EpilogMsgTime = 2000 usec EpilogSlurmctld = (null) ExtSensorsType = (null) ExtSensorsFreq = 0 sec FairShareDampeningFactor = 1 FederationParameters = (null) FirstJobId = 1 GetEnvTimeout = 2 sec GresTypes = gpu GpuFreqDef = (null) GroupUpdateForce = 1 GroupUpdateTime = 600 sec HASH_VAL = Match HealthCheckInterval = 300 sec HealthCheckNodeState = ANY HealthCheckProgram = /usr/sbin/nhc InactiveLimit = 0 sec InteractiveStepOptions = --interactive --preserve-env --pty $SHELL JobAcctGatherFrequency = 30 JobAcctGatherType = jobacct_gather/linux JobAcctGatherParams = (null) JobCompHost = localhost JobCompLoc = (null) JobCompParams = (null) JobCompPort = 0 JobCompType = (null) JobCompUser = root JobContainerType = (null) JobDefaults = (null) JobFileAppend = 0 JobRequeue = 1 JobSubmitPlugins = lua KillOnBadExit = 0 KillWait = 30 sec LaunchParameters = (null) Licenses = (null) LogTimeFormat = iso8601_ms MailDomain = (null) MailProg = /bin/mail MaxArraySize = 1001 MaxBatchRequeue = 5 MaxDBDMsgs = 20024 MaxJobCount = 10000 MaxJobId = 67043328 MaxMemPerNode = UNLIMITED MaxNodeCount = 6 MaxStepCount = 40000 MaxTasksPerNode = 512 MCSPlugin = (null) MCSParameters = (null) MessageTimeout = 10 sec MinJobAge = 300 sec MpiDefault = pmix_v2 MpiParams = (null) NEXT_JOB_ID = 3626 NodeFeaturesPlugins = (null) OverTimeLimit = 0 min PluginDir = /usr/lib64/slurm PlugStackConfig = (null) PowerParameters = (null) PowerPlugin = (null) PreemptMode = GANG,SUSPEND PreemptParameters = (null) PreemptType = preempt/qos PreemptExemptTime = 00:00:00 PrEpParameters = (null) PrEpPlugins = prep/script PriorityParameters = (null) PrioritySiteFactorParameters = (null) PrioritySiteFactorPlugin = (null) PriorityDecayHalfLife = 7-00:00:00 PriorityCalcPeriod = 00:05:00 PriorityFavorSmall = No PriorityFlags = PriorityMaxAge = 7-00:00:00 PriorityUsageResetPeriod = NONE PriorityType = priority/multifactor PriorityWeightAge = 0 PriorityWeightAssoc = 0 PriorityWeightFairShare = 0 PriorityWeightJobSize = 0 PriorityWeightPartition = 0 PriorityWeightQOS = 5000 PriorityWeightTRES = (null) PrivateData = none ProctrackType = proctrack/cgroup Prolog = (null) PrologEpilogTimeout = 65534 PrologSlurmctld = (null) PrologFlags = (null) PropagatePrioProcess = 0 PropagateResourceLimits = ALL PropagateResourceLimitsExcept = (null) RebootProgram = /etc/slurm/slurmupdate.sh ReconfigFlags = (null) RequeueExit = (null) RequeueExitHold = (null) ResumeFailProgram = (null) ResumeProgram = (null) ResumeRate = 300 nodes/min ResumeTimeout = 60 sec ResvEpilog = (null) ResvOverRun = 0 min ResvProlog = (null) ReturnToService = 1 SchedulerParameters = bf_max_job_user=2 SchedulerTimeSlice = 30 sec SchedulerType = sched/backfill ScronParameters = (null) SelectType = select/cons_tres SelectTypeParameters = CR_CORE_MEMORY SlurmUser = slurm(888) SlurmctldAddr = (null) SlurmctldDebug = info SlurmctldHost[0] = -some-server- SlurmctldLogFile = /var/log/slurm/slurmctld.log SlurmctldPort = 6817 SlurmctldSyslogDebug = (null) SlurmctldPrimaryOffProg = (null) SlurmctldPrimaryOnProg = (null) SlurmctldTimeout = 120 sec SlurmctldParameters = (null) SlurmdDebug = info SlurmdLogFile = /var/log/slurm/slurmd.log SlurmdParameters = (null) SlurmdPidFile = /var/run/slurmd.pid SlurmdPort = 6818 SlurmdSpoolDir = /var/spool/slurm SlurmdSyslogDebug = (null) SlurmdTimeout = 300 sec SlurmdUser = root(0) SlurmSchedLogFile = (null) SlurmSchedLogLevel = 0 SlurmctldPidFile = /var/run/slurmctld.pid SLURM_CONF = /etc/slurm/slurm.conf SLURM_VERSION = 23.11.1 SrunEpilog = (null) SrunPortRange = 0-0 SrunProlog = (null) StateSaveLocation = /var/spool/slurmctld SuspendExcNodes = (null) SuspendExcParts = (null) SuspendExcStates = (null) SuspendProgram = (null) SuspendRate = 60 nodes/min SuspendTime = INFINITE SuspendTimeout = 30 sec SwitchParameters = (null) SwitchType = (null) TaskEpilog = (null) TaskPlugin = task/cgroup,task/affinity TaskPluginParam = none TaskProlog = (null) TCPTimeout = 2 sec TmpFS = /tmp TopologyParam = (null) TopologyPlugin = topology/default TrackWCKey = No TreeWidth = 16 UsePam = No UnkillableStepProgram = (null) UnkillableStepTimeout = 60 sec VSizeFactor = 0 percent WaitTime = 0 sec X11Parameters = (null) Cgroup Support Configuration: AllowedRAMSpace = 100.0% AllowedSwapSpace = 0.0% CgroupMountpoint = /sys/fs/cgroup CgroupPlugin = autodetect ConstrainCores = yes ConstrainDevices = yes ConstrainRAMSpace = yes ConstrainSwapSpace = no EnableControllers = no IgnoreSystemd = no IgnoreSystemdOnFailure = no MaxRAMPercent = 100.0% MaxSwapPercent = 100.0% MemorySwappiness = (null) MinRAMSpace = 30 MB MPI Plugins Configuration: PMIxCliTmpDirBase = (null) PMIxCollFence = (null) PMIxDebug = 0 PMIxDirectConn = yes PMIxDirectConnEarly = no PMIxDirectConnUCX = no PMIxDirectSameArch = no PMIxEnv = (null) PMIxFenceBarrier = no PMIxNetDevicesUCX = (null) PMIxTimeout = 300 PMIxTlsUCX = (null) the partition is configured like: PartitionName=test-partition AllowGroups=sysadmin,users AllowAccounts=ALL AllowQos=ALL AllocNodes=ALL Default=NO QoS=N/A DefaultTime=NONE DisableRootJobs=YES ExclusiveUser=NO GraceTime=0 Hidden=NO MaxNodes=UNLIMITED MaxTime=UNLIMITED MinNodes=0 LLN=NO MaxCPUsPerNode=UNLIMITED MaxCPUsPerSocket=UNLIMITED Nodes=vserv-[275-277] PriorityJobFactor=1 PriorityTier=100 RootOnly=NO ReqResv=NO OverSubscribe=FORCE:1 OverTimeLimit=NONE PreemptMode=GANG,SUSPEND State=UP TotalCPUs=32 TotalNodes=4 SelectTypeParameters=NONE JobDefaults=(null) DefMemPerNode=UNLIMITED MaxMemPerNode=UNLIMITED TRES=cpu=32,mem=304933M,node=4,billing=32 Our QOS looks like this: | Name | Priority | GraceTime | Preempt | PreemptMode | Flags | UsageFactor | MaxTRESPU | MaxJobsPU | MaxSubmitPU |-----------|----------|-----------|-----------|-------------|-----------------|-------------|-----------------|-----------|----------- | normal | 50 | 00:00:00 | | cluster | | 1.000000 | | 20 | 50 | preempter | 1000 | 00:00:00 | preempted | gang,suspe+| | 1.000000 | | | | preempted | 0 | 00:00:00 | | gang,suspe+ | OverPartQOS | 1.000000 | cpu=100,node=10 | | I can provide more configs if needed. Do you guys see anything strange? Or any property to be set? This is the state of the queue: |JOBID | QOS | ST | TIME | NODELIST(REASON) | PARTITION | PRIORITY | | ---------------- ----------- --------- ----------- ---------------------- ----------------- ----------| | 3629 | preempted | PD | 0:00 | (Resources) | test-partition | 1 | | 3627 | preempted | R | 0:20 | vserv-276 | test-partition | 1 | | 3628 | preempted | R | 0:20 | vserv-277 | test-partition | 1 | | 3626 | preempted | R | 0:27 | vserv-275 | test-partition | 1 | | 3630 | preempter | PD | 0:00 | (Resources) | test-partition | 5000 | any advice is welcomed. Regards, Nischey Verma Nischey Verma

3 2

Issues with Slurm 23.11.1
by Fokke Dijkstra 15 Mar '24

15 Mar '24

Dear all, Since the upgrade from Slurm 22.05 to 23.11.1 we are having problems with the communication between the slurmctld and slurmd processes. We are running a cluster with 183 nodes and almost 19000 cores. Unfortunately some nodes are in a different network preventing full internode communication. A network topology and setting TopologyParam RouteTree have been used to make sure no slurmd communication happens between nodes on different networks. In the new Slurm version we see the following issues, which did not appear in 22.05: 1. slurmd processes acquire many network connections in CLOSE-WAIT (or CLOSE_WAIT depending on the tool used) causing the processes to hang, when trying to restart slurmd. When checking for CLOSE-WAIT processes we see the following behaviour: Recv-Q Send-Q Local Address:Port Peer Address:Port Process 1 0 10.5.2.40:6818 10.5.0.43:58572 users:(("slurmd",pid=1930095,fd=72),("slurmd",pid=1930067,fd=72)) 1 0 10.5.2.40:6818 10.5.0.43:58284 users:(("slurmd",pid=1930095,fd=8),("slurmd",pid=1930067,fd=8)) 1 0 10.5.2.40:6818 10.5.0.43:58186 users:(("slurmd",pid=1930095,fd=22),("slurmd",pid=1930067,fd=22)) 1 0 10.5.2.40:6818 10.5.0.43:58592 users:(("slurmd",pid=1930095,fd=76),("slurmd",pid=1930067,fd=76)) 1 0 10.5.2.40:6818 10.5.0.43:58338 users:(("slurmd",pid=1930095,fd=19),("slurmd",pid=1930067,fd=19)) 1 0 10.5.2.40:6818 10.5.0.43:58568 users:(("slurmd",pid=1930095,fd=68),("slurmd",pid=1930067,fd=68)) 1 0 10.5.2.40:6818 10.5.0.43:58472 users:(("slurmd",pid=1930095,fd=69),("slurmd",pid=1930067,fd=69)) 1 0 10.5.2.40:6818 10.5.0.43:58486 users:(("slurmd",pid=1930095,fd=38),("slurmd",pid=1930067,fd=38)) 1 0 10.5.2.40:6818 10.5.0.43:58316 users:(("slurmd",pid=1930095,fd=29),("slurmd",pid=1930067,fd=29)) The first IP address is that of the compute node, the second that of the node running slurmctld. The nodes can communicate using these IP addresses just fine. 2. slurmd cannot be properly restarted [2024-01-18T10:45:26.589] slurmd version 23.11.1 started [2024-01-18T10:45:26.593] error: Error binding slurm stream socket: Address already in use [2024-01-18T10:45:26.593] fatal: Unable to bind listen port (6818): Address already in use This is probably because of the processes being in CLOSE-WAIT, which can only be killed using signal -9. 3. We see jobs stuck in completing CG state, probably due to communication issues between slurmctld and slurmd. The slurmctld sends repeated kill requests but those do not seem to be acknowledged by the client. This happens more often in large job arrays, or generally when many jobs start at the same time. However, this could be just a biased observation (i.e., it is more noticeable on large job arrays because there are more jobs to fail in the first place). 4. Since the new version we also see messages like: [2024-01-17T09:58:48.589] error: Failed to kill program loading user environment [2024-01-17T09:58:48.590] error: Failed to load current user environment variables [2024-01-17T09:58:48.590] error: _get_user_env: Unable to get user's local environment, running only with passed environment The effect of this is that the users run with the wrong environment and can’t load the modules for the software that is needed by their jobs. This leads to many job failures. The issue appears to be somewhat similar to the one described at: https://bugs.schedmd.com/show_bug.cgi?id=18561 In that case the site downgraded the slurmd clients to 22.05 which got rid of the problems. We’ve now downgraded the slurmd on the compute nodes to 23.02.7 which also seems to be a workaround for the issue. Does anyone know of a better solution? Kind regards, Fokke Dijkstra -- Fokke Dijkstra <f.dijkstra(a)rug.nl> <f.dijkstra(a)rug.nl> Team High Performance Computing Center for Information Technology, University of Groningen Postbus 11044, 9700 CA Groningen, The Netherlands

2 3

Overhead of multiple concurrent job steps
by Mehta, Kshitij 15 Mar '24

15 Mar '24

Hello, We have a use case in which we need to launch multiple concurrently running MPI applications inside a job allocation. Most supercomputing facilities limit the number of concurrent job steps as they incur an overhead with the global Slurm scheduler. Some frameworks, such as the Flux framework from LLNL, claim to mitigate this issue by starting an instance of their own scheduler inside an allocation, which then acts as the resource manager for the compute nodes in the allocation. Out of curiosity, I was wondering if there is a fundamental reason behind having a single global scheduler that the srun launch commands must contact to launch job steps. Perhaps it was overkill to develop a ‘hierarhical’ design in which Slurm launches a local job daemon for every allocation that manages resources for that allocation? I would appreciate your insight in understanding more about Slurm’s core design. Thanks and regards, Kshitij Mehta Oak Ridge National Laboratory

1 0

Jobs being denied for GrpCpuLimit despite having enough resource
by Simon Andrews 14 Mar '24

14 Mar '24

Our cluster has developed a strange intermittent behaviour where jobs are being put into a pending state because they aren't passing the AssocGrpCpuLimit, even though the user submitting has enough cpus for the job to run. For example: $ squeue -o "%.6i %.9P %.8j %.8u %.2t %.10M %.7m %.7c %.20R" JOBID PARTITION NAME USER ST TIME MIN_MEM MIN_CPU NODELIST(REASON) 799 normal hostname andrewss PD 0:00 2G 5 (AssocGrpCpuLimit) ..so the job isn't running, and it's the only job in the queue, but: $ sacctmgr list associations part=normal user=andrewss format=Account,User,Partition,Share,GrpTRES Account User Partition Share GrpTRES ---------- ---------- ---------- --------- ------------- andrewss andrewss normal 1 cpu=5 That user has a limit of 5 CPUs so the job should run. The weird thing is that this effect is intermittent. A job can hang and the queue will stall for ages but will then suddenly start working and you can submit several jobs and they all work, until one fails again. The cluster has active nodes and plenty of resource: $ sinfo PARTITION AVAIL TIMELIMIT NODES STATE NODELIST normal* up infinite 2 idle compute-0-[6-7] interactive up 1-12:00:00 3 idle compute-1-[0-1,3] The slurmctld log just says: [2024-03-14T16:21:41.275] _slurm_rpc_submit_batch_job: JobId=799 InitPrio=4294901720 usec=259 Whilst it's in this state I can run other jobs with core requests of up to 4 and they work, but not 5. It's like slurm is adding one CPU to the request and then denying it. I'm sure I'm missing something fundamental but would appreciate it if someone could point out what it is! Thanks Simon.

3 3

We're Live! Check out the new SchedMD.com now!
by Victoria Hobson 13 Mar '24

13 Mar '24

Our website has gone through some much needed change and we'd love for you to explore it! The new SchedMD.com is equipped with the latest information about Slurm, your favorite workload manager, and details about SchedMD services, support, and training offerings. Toggle through our Industries pages (https://www.schedmd.com/slurm-industries/) to learn more about how Slurm can service your specific site needs. Why Slurm? (https://www.schedmd.com/slurm/why-slurm/) gives you all the basics around our market-leading scheduler and SchedMD Services (https://www.schedmd.com/slurm-support/our-services/) addresses all the ways we can help you optimize your site. These new web pages also feature access to our Documentation Site, Bug Site, and Installation Guide. Browse our Events tab to see where we'll be when, and be sure to register for our Slurm User Group (SLUG) in Oslo, Norway this fall! (https://www.schedmd.com/about-schedmd/events/) SchedMD.com, your one stop shop for all things Slurm. Check it out now! -- Victoria Hobson SchedMD LLC Vice President of Marketing

2 1

Slurm Utilities
by Paul Edmon 13 Mar '24

13 Mar '24

Just wanted to share some slurm utilities that we've written at Harvard FASRC that maybe useful to the community. seff-account: https://github.com/fasrc/seff-account Creates job statistics summaries for users and accounts similar to what seff and seff-array does. showq: https://github.com/fasrc/slurm_showq A slurm version of the Moab showq command lsload: https://github.com/fasrc/lsload A slurm version of the LSF lsload command scalc: https://github.com/fasrc/scalc A calculator for various fairshare related things spart: https://github.com/fasrc/spart A simplified output for slurm partition information stdg: https://github.com/fasrc/stdg Slurm test deck generator prometheus-slurm-exporter: https://github.com/fasrc/prometheus-slurm-exporter Slurm exporters for prometheus Hopefully people find these useful. Pull requests are always appreciated. -Paul Edmon-

1 0

Re: SLURM in K8s, any advice?
by LEAVY Alan 13 Mar '24

13 Mar '24

I'm a little late to this party but would love to establish contact with others using slurm in Kubernetes. I recently joined a research institute in Vienna (IIASA) and I'm getting to grips with slurm and Kubernetes (my previous role was data engineering / fintech). My current setup sounds like what Urban described in this thread, back in Nov 22. It has some rough edges though. Right now, I'm trying to upgrade to slurm-23.11.4 in Ubuntu 23.10 containers. I'm having trouble with the cgroup/v2 plugin. Are you still using slurm on K8s Urban? How did your installation work out Hans? Would either of you be willing to share your experiences? Regards, Alan.

4 3

Shards distribution over multiple GPUs
by Francois Broquedis 12 Mar '24

12 Mar '24

Hi all, We're trying to enable sharding on our compute cluster. On this cluster: - ensicompute-1 comes with 1 NVIDIA V100 GPU ; - ensicompute-13 comes with 3 NVIDIA A40 GPUs ; - all other nodes (for now, ensicompute-11 and ensicompute-12, but several others will come) come with 3 NVIDIA RTX 6000 GPUs. To enable sharding, I followed these steps: 1. [slurm.conf] Add "shard" to GresTypes ; 2. [slurm.conf] Add "shard:N" to Gres for each node. For testing purposes, I have set N to 9, so each GPU can execute up to 3 jobs concurrently: NodeName=ensicompute-[11-12] Gres=gpu:Quadro:3,shard:9 CPUs=40 RealMemory=128520 Sockets=2 CoresPerSocket=10 ThreadsPerCore=2 State=UNKNOWN Feature=gpu,ht 3. [gres.conf] Declare the shards after the definition of the GPUs GRES. For step 3, I tried different things, leading to different outcomes: a. Define a global number of shards, for the entire host: Name=shard Count=9 ==> This way, sharding seems to work ok, but all the jobs are executed on GPU#0. If running 12 jobs for example, 9 of them are assigned to GPU#0 and start executing, while 3 of them remain in a pending state. No job is assigned to GPU#1 or GPU#2. b. Define a per-GPU number of shards, associated to the device file representing the GPU: Name=shard Count=3 File=/dev/nvidia0 Name=shard Count=3 File=/dev/nvidia1 Name=shard Count=3 File=/dev/nvidia2 ==> In this case, the slurmd service fails to start on the compute node. The error message found in /var/log/slurmd.log is "fatal: Invalid GRES record for shard, count does not match File value". c. Don't define anything about shards in gres.conf. ==> Same behavior than in a.: all jobs are executed on GPU#0. I attach to this message the full content of the slurm.conf and gres.conf files. What is the proper way to configure sharding in a cluster with several GPUs per node? Is there a way to specify how many shards should be allocated to each GPU? Cheers, François === slurm.conf === # slurm.conf file generated by configurator.html. # Put this file on all nodes of your cluster. # See the slurm.conf man page for more information. # ClusterName=ensimag SlurmctldHost=nash ProctrackType=proctrack/cgroup SlurmctldPidFile=/var/run/slurmctld.pid SlurmctldPort=6817 SlurmdPidFile=/var/run/slurmd.pid SlurmdPort=6818 SlurmdSpoolDir=/var/spool/slurmd SlurmUser=slurm #SlurmdUser=root StateSaveLocation=/var/spool/slurmctld TaskPlugin=task/affinity,task/cgroup ReturnToService=2 # # # TIMERS InactiveLimit=0 KillWait=30 MinJobAge=300 SlurmctldTimeout=120 SlurmdTimeout=300 Waittime=0 # # # SCHEDULING SchedulerType=sched/backfill SelectType=select/cons_tres # # # LOGGING AND ACCOUNTING JobCompType=jobcomp/none JobAcctGatherFrequency=30 SlurmctldDebug=info SlurmctldLogFile=/var/log/slurmctld.log SlurmdDebug=info SlurmdLogFile=/var/log/slurmd.log # # # COMPUTE NODES GresTypes=gpu,shard NodeName=ensicompute-1 Gres=gpu:Tesla:1,shard:3 CPUs=40 RealMemory=128520 Sockets=2 CoresPerSocket=10 ThreadsPerCore=2 State=UNKNOWN Feature=gpu,ht NodeName=ensicompute-13 Gres=gpu:A40:3,shard:9 CPUs=40 RealMemory=128520 Sockets=2 CoresPerSocket=10 ThreadsPerCore=2 State=UNKNOWN Feature=gpu,ht NodeName=ensicompute-[11-12] Gres=gpu:Quadro:3,shard:9 CPUs=40 RealMemory=128520 Sockets=2 CoresPerSocket=10 ThreadsPerCore=2 State=UNKNOWN Feature=gpu,ht PartitionName=compute Nodes=ALL Default=YES MaxTime=INFINITE State=UP === gres.conf === AutoDetect=off # ensicompute-1 NodeName=ensicompute-1 Name=gpu Type=Tesla File=/dev/nvidia0 NodeName=ensicompute-1 Name=shard Count=3 File=/dev/nvidia0 # ensicompute-11 NodeName=ensicompute-11 Name=gpu Type=Quadro File=/dev/nvidia0 NodeName=ensicompute-11 Name=gpu Type=Quadro File=/dev/nvidia1 NodeName=ensicompute-11 Name=gpu Type=Quadro File=/dev/nvidia2 NodeName=ensicompute-11 Name=shard Count=3 File=/dev/nvidia0 NodeName=ensicompute-11 Name=shard Count=3 File=/dev/nvidia1 NodeName=ensicompute-11 Name=shard Count=3 File=/dev/nvidia2 # ensicompute-12 NodeName=ensicompute-12 Name=gpu Type=Quadro File=/dev/nvidia0 NodeName=ensicompute-12 Name=gpu Type=Quadro File=/dev/nvidia1 NodeName=ensicompute-12 Name=gpu Type=Quadro File=/dev/nvidia2 NodeName=ensicompute-12 Name=shard Count=3 File=/dev/nvidia0 NodeName=ensicompute-12 Name=shard Count=3 File=/dev/nvidia1 NodeName=ensicompute-12 Name=shard Count=3 File=/dev/nvidia2 # ensicompute-13 NodeName=ensicompute-13 Name=gpu Type=A40 File=/dev/nvidia0 NodeName=ensicompute-13 Name=gpu Type=A40 File=/dev/nvidia1 NodeName=ensicompute-13 Name=gpu Type=A40 File=/dev/nvidia2 NodeName=ensicompute-13 Name=shard Count=3 File=/dev/nvidia0 NodeName=ensicompute-13 Name=shard Count=3 File=/dev/nvidia1 NodeName=ensicompute-13 Name=shard Count=3 File=/dev/nvidia2 -- François Broquedis, Ingénieur Service Informatique Grenoble INP - Ensimag, bureau E208 681 rue de la Passerelle BP 72, 38402 Saint Martin d'Hères CEDEX Tél.: +33 (0)4 76 82 72 78

1 0

Re: user_name attribute missing in job response from slurm rest api method jobs
by Acartürk, Ayhan 11 Mar '24

11 Mar '24

Sorry... my previous message should have been a reply to the post by Filip Holka from March, 25th 2022: https://lists.schedmd.com/pipermail/slurm-users/2022-March/008543.html Best wishes, Ayhan Acartürk

1 0

Re: user_name attribute missing in job response from slurm rest api method jobs
by Acartürk, Ayhan 11 Mar '24

11 Mar '24

Hello together, we have the same problem here. All jobs delivered by the api endpoint /slurm/v0.0.36/jobs have a user_id attached but no user_name. Currently, we solve this issue by an extra call for each job_id to the api endpoint /slurmdb/v.0.0.36/job/{job_id}. Any idea where to search for a possible solution is highly appreciated. Best wishes, Ayhan Acartürk

1 0

Problem building slurm with PMIx
by Patrick Begou 11 Mar '24

11 Mar '24

Hi ! I manage a small CentOS8 cluster using slurm slurm-20.11.7-1 and OpenMPI built from sources. - I know this OS is not maintained any more and I need to negotiate downtime to reinstall - I know Slurm 20.11.7 has security issue (I've built it from source some years ago with rpmbuild -ta --with mysql --with hwloc slurm-20.11.7.tar.bz) and I should update. All was running fine until I add a GPU Node and Nvidia sdk. This SDK provides an openMPI3 implementation GPU aware but I'm unable to launch an intranode parallel job with it using srun: -------------------------------------------------------------------------- The application appears to have been direct launched using "srun", but OMPI was not built with SLURM's PMI support and therefore cannot execute. There are several options for building PMI support under SLURM, depending upon the SLURM version you are using: version 16.05 or later: you can use SLURM's PMIx support. This requires that you configure and build SLURM --with-pmix. Versions earlier than 16.05: you must use either SLURM's PMI-1 or PMI-2 support. SLURM builds PMI-1 by default, or you can manually install PMI-2. You must then build Open MPI using --with-pmi pointing to the SLURM PMI library location. Please configure as appropriate and try again. -------------------------------------------------------------------------- I check with "srun --mpi=list" and got no pmx. srun: MPI types are... srun: pmi2 srun: cray_shasta srun: none So I decide to build the rpms from slurm-20.11.9.tar.bz2 as I had done previously for 20.11.7 and update. I've first installed pmix-2.1.1-1 from src as I had no pmix-devel rpm in my local CentOS8 repo: rpm --rebuild pmix-2.1.1-1.el8.src.rpm dnf install pmix-devel-2.1.1-1.el8.x86_64.rpm pmix-2.1.1-1.el8.x86_64.rpm Then build Slurm from slurm-20.11.9.tar.bz2 (just changing python3 to python38 in the spec file) rpmbuild -ta --with mysql --with hwloc --with pmix slurm-20.11.9.tar.bz2 And the try to install these package on the GPU node dnf install slurm-slurmd-20.11.9-1.el8.x86_64.rpm slurm-20.11.9-1.el8.x86_64.rpm slurm-devel-20.11.9-1.el8.x86_64.rpm slurm-libpmi-20.11.9-1.el8.x86_64.rpm But I get this strange error: Error: Problem: conflicting requests - nothing provides pmix = 20.11.9 needed by slurm-slurmd-20.11.9-1.el8.x86_64 (try to add '--skip-broken' to skip uninstallable packages or '--nobest' to use not only best candidate packages) Why this request on PMIX with the slurm version number ? Am I wrong somewhere ? Thanks for your help Patrick

2 1

FairShare value is always 0.
by Zacarias Benta 07 Mar '24

07 Mar '24

Hi guys, We've just setup our new cluster and are facing some issues regading fairshare calculation. Our slurm directive regarding priority calculation are defines as follows: PriorityType=priority/multifactor PriorityFlags=MAX_TRES PriorityDecayHalfLife=14-0 PriorityFavorSmall=NO PriorityMaxAge=14-0 PriorityWeightAge=1000 PriorityWeightJobSize=1000 PriorityWeightPartition=10000000 PriorityWeightQOS=10000000 PriorityWeightTRES=CPU=2000,Mem=4000 PriorityWeightFairshare=100000 The partition we are submitinh out jobs to is setup as follows: PartitionName=mypartPriority=1000TRESBillingWeights="CPU=1.0,Mem=0.25G"Default=YESMaxTime=96:0:0DefMemPerCPU=5333Nodes=node[001-036] MaxNodes=20 Whenever we take a look at the fairshare value using sshare -l we see the following output: AccountUserRawSharesNormSharesRawUsageNormUsageEffectvUsageFairShareLevelFSGrpTRESMinsTRESRunMins ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- root10.0000002687245970.0000000.000000cpu=1098201,mem=5856709132,en+ rootroot10.10000000.0000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group110.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group210.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group310.1000002687245970.0000000.000000cpu=1098201,mem=5856709132,en+ group410.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group510.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group610.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group710.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group810.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ group910.10000000.0000000.000000cpu=0,mem=0,energy=0,node=0,b+ We think it is really weird that the FairShare value is 0 for the root account and "NULL" for all other groups, even the one who had the greatest raw usage. While taking a look at the data for our users we see the following: AccountUserRawSharesNormSharesRawUsageEffectvUsageFairShare ------------------------------------------------------------------------------------- root10.0000002689837210.000000 rootroot10.10000000.0000000.000000 group310.1000002689837210.000000 group3user110.090909121093740.0000000.000000 group3user210.09090900.0000000.000000 group3user310.09090900.0000000.000000 group3user410.09090900.0000000.000000 group3user510.09090900.0000000.000000 group3user610.09090900.0000000.000000 group3user710.09090900.0000000.000000 group3user810.0909092088245970.0000000.000000 group3user910.09090900.0000000.000000 group3user1010.09090900.0000000.000000 group3user1110.090909480497500.0000000.000000 group410.10000000.000000 group4user1310.0000004994520.0000000.000000 group510.10000000.000000 group5user1410.00000015396030.0000000.000000 This is a weird behavior, since user1, user8, user11, user13 and user14 are the ones who have more RawUsage and the FairShare value is the same for all of them, including the users that have no yet submited any job. We also noticed that in the slurmctld log there is the fillowing error message that appears with some regularity [2024-03-07T16:38:13.260] error: _append_list_to_array: unable to append NULL list to assoc list. [2024-03-07T16:38:13.260] error: _calc_tree_fs: unable to calculate fairshare on empty tree The error above looks like it is coming from: https://github.com/SchedMD/slurm/blob/b11bf689b270f1f5dfe4b0cd54c4fa84b4af3… Are we missing any setting on slurm.conf? This is kind of strange, because we have another cluster with pretty much the same configuration and the FairShare is calculated without any problems. Any help would be appreciated. -- Cumprimentos / Best Regards, Zacarias Benta LIP/INCD @ UMINHO ---------------------------------------------- / Use linux, and may the source be with you. / ---------------------------------------------- \ __ -=(o '. '.-.\ /| \\ '| || _\_):,_

1 0

Slurm management of dual-node server trays?
by Ole Holm Nielsen 07 Mar '24

07 Mar '24

We're in the process of installing some racks with Lenovo SD665 V3 [1] water-cooled servers. A Lenovo DW612S chassis contains 6 1U trays with 2 SD665 V3 servers mounted side-by-side in each tray. Lenovo delivers SD665 V3 servers including water-cooled NVIDIA InfiniBand "SharedIO" adapters [2] so that one node is the Primary including a PCIe adapter, and the other is Auxiliary with just a cable to the Primary's adapter. Obviously, servicing 2 "Siamese twin" Slurm nodes requires a bit of care and planning. What is worse is that when the Primary node is rebooted or powered down, the Auxiliary node will lose its Infiniband connection and may have a PCIe fault or an NMI as documented in [3]. And when nodes are powered up, the Primary must have completed POST before the Auxiliary gets started. I wonder how to best deal with power failures? It seems that when Slurm jobs are running on Auxiliary nodes, these jobs are going to crash when the possibly unrelated Primary node goes down. This looks like a pretty bad system design on the part of Lenovo :-( The goal was apparently to same some money on IB adapters and having fewer IB cables. Question: Do any Slurm sites out there already have experiences with Lenovo "Siamese twin" nodes with SharedIO IB? Have you developed some operational strategies, for example dealing with node pairs as a single entity for job scheduling? Thanks for sharing any ideas and insights! Ole [1] https://lenovopress.lenovo.com/lp1612-lenovo-thinksystem-sd665-v3-server [2] https://lenovopress.lenovo.com/lp1693-thinksystem-nvidia-connectx-7-ndr200-… [3] https://support.lenovo.com/us/en/solutions/ht510888-thinksystem-sd650-and-c… -- Ole Holm Nielsen PhD, Senior HPC Officer Department of Physics, Technical University of Denmark

3 3

Re: Lua script
by Paul Raines 06 Mar '24

06 Mar '24

Alternativelly consider setting EnforcePartLimits=ALL in slurm.conf -- Paul Raines (http://help.nmr.mgh.harvard.edu) 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.

1 0

Re: Lua script
by Gestió Servidors 06 Mar '24

06 Mar '24

And how can I reject the job inside the lua script?

2 1

Lua script
by Gestió Servidors 06 Mar '24

06 Mar '24

Hello, I'm writing a small lua script that for modify "TimeLimit" of a submited job if user has configured a TimeLimit bigger that configured in the partition. So, is TimeLimit for partition is, for example, 4 hours (04:00:00) and user submit his/her job with a TimeLimit of 5 hours, lua script modify this submit and forces "TimeLimit" to 4 hours for avoiding that job remains at queue in state "pending" with reason "Partition Time Limit...". I have written these lines: function slurm_job_submit(job_desc, part_list, submit_uid) --[[ my id is 1008; for testing purposes, lua only applies to me --]] if (job_desc.user_id == 1008) then slurm.log_info("Submit job by me") --[[ I check partition --]] if (job_desc['partition'] == "nodo.q") then --[[ if (job_desc['time_limit'] > 14400) job_desc['time_limit'] = 14400 end end end return slurm.SUCCESS end However, if I submit a job with TimeLimit of 5 hours, lua script doesn't modify submit and job remains "pending"... What am I doing wrong? Thanks.

2 1

Slurm billback and sreport
by Chip Seraphine 05 Mar '24

05 Mar '24

Hello, I am attempting to implement a billback model and finding myself stymied by the way that sreport handles job arrays. Basically, when a user submits a large array, their usage includes time that jobs in the back of the array spend waiting their turn. (My #1 user in “sreport user topusage” shows more “used” cpu*minutes than the cluster physically _has_ during that interval.) However, jobs that are idle pending resources are simply regarded as pending; as a result, a “polite” user who submits an array of 1000 jobs running N at a time is penalized over a user who just dumps 1000 loose jobs into the queue. This incentives my users to do exactly what I do not want! Has anyone tried to bill their users based on the results of sreport? If so, how did you work around this problem? What did you use to determine the # of CPU*Minutes that a user actually allocated on during a given interval? -- Chip Seraphine Grid Operations For support please use help-grid in email or slack. This e-mail and any attachments may contain information that is confidential and proprietary and otherwise protected from disclosure. If you are not the intended recipient of this e-mail, do not read, duplicate or redistribute it by any means. Please immediately delete it and any attachments and notify the sender that you have received it by mistake. Unintended recipients are prohibited from taking action on the basis of information in this e-mail or any attachments. The DRW Companies make no representations that this e-mail or any attachments are free of computer viruses or other defects.

3 3

Is SWAP memory mandatory for SLURM
by John Joseph 04 Mar '24

04 Mar '24

Dear All, Good morning I do have a 4 node SLURM instance up and running. Like to know if I disable the SWAP memory, will it effect the SLURM performance Is SWAP a mandatory requirement, I have each node more RAM, if my phsicall RAM is more, is there any need for the SWAP thanks Joseph John

4 3

What is "eligible time"
by Oren Shani 04 Mar '24

04 Mar '24

Hi All, Can someone shed some light on the exact meaning of the "eligible time" attribute in the slurm accounter, or to be more exact on the time between when the jobs becomes "eligible" to the time it actually starts to run? I understand that "eligible time" refers to the time in which the job becomes a candidate for running and that one case in which it will still not run, is that it waits for resources to become available. But are there any other conditions for a job to keep waiting at this point? For example, is fair-share scheduling considered only at this point? Many thanks, Oren

1 0

Question about IB and Ethernet networks
by Dan Healy 04 Mar '24

04 Mar '24

Hi Fellow Slurm Users, This question is not slurm-specific, but it might develop into that. My question relates to understanding how *typical* HPCs are designed in terms of networking. To start, is it typical for there to be a high speed Ethernet *and* Infiniband networks (meaning separate switches, NICs)? I know you can easily setup IP over IB, but is IB usually fully reserved for MPI messages? I’m tempted to spec all new HPCs with only a high speed (200Gbps) IB network, and use IPoIB for all slurm comms with compute nodes. I plan on using BeeGFS for the file system with RDMA. Just looking for some feedback, please. Is this OK? Is there a better way? If yes, please share why it’s better. Thanks, Daniel Healy

6 6

How to get usage data for a QOS
by thomas.hartmann＠th-ht.de 01 Mar '24

01 Mar '24

Hi, so, I figured out that I can give some users priority access for a specific amount of TRES by creating a qos with the GrpTRESMins property and the DenyOnLimit,NoDecay flags. This works nicely. However, I would like to know, how much of this has already been consumed and I have not yet found a way to do this. Like: How can I get the amount of TRES/TRES minutes consumed for a certain QOS? Thanks a lot! Thomas

3 2

MinTRES in QoS and power saving
by Stefan Walter 01 Mar '24

01 Mar '24

I am running a GPU cluster where nodes are mostly off to save on electricity. I have run into the problem that if I set 'MinTRES=gres/gpu=1' in the QoS for user-account associations, waking up nodes on-demand stops working for these users. Jobs are allocated on all running nodes but if a user starts a job that would require a node to be woken up, it remains pending with 'QOSMinGRES'. In the verbose logs, such a job will show this: debug3: TRES Weight: cpu = 1.000000 * 0.000000 = 0.000000 debug3: TRES Weight: mem = 6000.000000 * 0.000000 = 0.000000 debug3: TRES Weight: energy = 0.000000 * 0.000000 = 0.000000 debug3: TRES Weight: node = 1.000000 * 0.000000 = 0.000000 debug3: TRES Weight: fs/disk = 0.000000 * 0.000000 = 0.000000 debug3: TRES Weight: vmem = 0.000000 * 0.000000 = 0.000000 debug3: TRES Weight: pages = 0.000000 * 0.000000 = 0.000000 debug3: TRES Weight: gres/gpu = 0.000000 * 1.000000 = 0.000000 debug3: TRES Weight: gres/gpu:nvidia_geforce_gtx_1080ti = 0.000000 * 0.000000 = 0.000000 debug3: TRES Weight: gres/gpumem = 0.000000 * 0.000000 = 0.000000 debug3: TRES Weight: gres/gpuutil = 0.000000 * 0.000000 = 0.000000 debug3: TRES Weighted: SUM(TRES) = 0.000000 debug2: JobId=4256 is being held, QOS local818-test min tres(gres/gpu) per job request 0 exceeds min tres limit 1 The job was started with '-G 1' or '--gres=gpu:1' but notably the TRES value for 'gres/gpu' becomes 0. For jobs that fit on the running nodes the TRES weight remains correct: debug3: TRES Weight: gres/gpu = 1.000000 * 1.000000 = 1.000000 The quick fix was to unset MinTRES for all QoS but I still would force user in some accounts to use GPUs because that is what is billed for them. Is this by design? I was not successful to find any hint that this is expected behavior. Thanks, Stefan

1 0

pty jobs are killed when another job on the same node terminates
by Thomas Hartmann 01 Mar '24

01 Mar '24

Hi, when I start an interactive job like this: srun --pty --mem=3G -c2 bash And then I schedule and run other jobs (can be interactive or non interactive) and one of these jobs that runs on the same node terminates, the interactive job gets killed with this message: srun: error: node01.abc.at: task 0: Killed I attached our slurm config. Does anybody have an idea what is going on here or where I could look to debug? I'm quite new to slurm, so I don't know all the places to look... Thanks a lot in advance! Thomas

4 3

GPU shards not exclusive
by Reed Dier 29 Feb '24

29 Feb '24

I seem to have run into an edge case where I’m able to oversubscribe a specific subset of GPUs on one host in particular. Slurm 22.05.8 Ubuntu 20.04 cgroups v1 (ProctrackType=proctrack/cgroup) It seems to be partly a corner case with a couple of caveats. This host has 2 different GPU types in the same host. All GPUs are configured with 6 shards/gpu. > NodeName=gpu03 CPUs=88 RealMemory=768000 Sockets=2 CoresPerSocket=22 ThreadsPerCore=2 State=UNKNOWN Feature=avx512 Gres=gpu:rtx:2,gpu:p40:6,shard:48 My gres.conf is auto-nvml, with nothing else in the file. > AutoDetect=nvml Schedule and select from slurm.conf > SchedulerType=sched/backfill > SelectType=select/cons_tres > SelectTypeParameters=CR_CPU_Memory > GresTypes=gpu,shard Here’s a copy of slurmd’s log starting up with debug logging: https://pastebin.com/fbriTAZD <https://pastebin.com/fbriTAZD> I checked there to make sure that it was allocating an even 6 shards per GPU as intended, and it did. The only thing odd about the rtx cards is that they have an NVLink bridge between them, but I wouldn’t expect that to have any impact with respect to shards. So far it only appears to manifest if anything is using the named gpu (--gres gpu:rtx:1), but not as a general gres (--gres gpu:1). It also doesn’t occur with any count of named p40 gres either, only with the rtx gres. The below was as “high” as I was able to capture it oversubscribed in very quick testing. You can see there are 2 rtx and 43 shards allocated, which should be an impossible number, if we correctly assume an rtx==6 shards, and with 48 total shards, that should reduce the available pool of shards to 36 shards. > NodeName=gpu03 Arch=x86_64 CoresPerSocket=22 > CPUAlloc=88 CPUEfctv=88 CPUTot=88 CPULoad=0.27 > AvailableFeatures=avx512 > ActiveFeatures=avx512 > Gres=gpu:p40:6(S:0),gpu:rtx:2(S:0),shard:p40:36(S:0),shard:rtx:12(S:0) > NodeAddr=gpu03 NodeHostName=gpu03 Version=22.05.8 > OS=Linux 5.4.0-164-generic #181-Ubuntu SMP Fri Sep 1 13:41:22 UTC 2023 > RealMemory=768000 AllocMem=45056 FreeMem=563310 Sockets=2 Boards=1 > State=ALLOCATED+RESERVED ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A > Partitions=gpu,gpu-prod > BootTime=2024-02-02T14:17:07 SlurmdStartTime=2024-02-14T16:28:17 > LastBusyTime=2024-02-14T16:38:03 > CfgTRES=cpu=88,mem=750G,billing=88,gres/gpu=8,gres/gpu:p40=6,gres/gpu:rtx=2,gres/shard=48 > AllocTRES=cpu=88,mem=44G,gres/gpu=2,gres/gpu:rtx=2,gres/shard=43 > CapWatts=n/a > CurrentWatts=0 AveWatts=0 > ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s The above is 1 rtx:2 jobs, and 43 shard:1 jobs, which is as many as I can seem to push through with the given cpu cores on the system, I expect if there were 98-100 cores I could completely oversubscribe it. If I do (1) rtx:2 + (24) shard:2 it actually limits it correctly to 36 shards, which is odd because the case that caused us to discover this was when 2 rtx:1 jobs were running for multiple days, and somehow shards got scheduled on those gpus. > AllocTRES=cpu=40,mem=20G,gres/gpu=2,gres/gpu:rtx=2,gres/shard=36 If I flip it around to shards before rtx, then I’m unable to oversubscribe it. If I add any P40 jobs before the rtx and then shard jobs, also unable to oversubscribe. This sounded familiar to Bug 16484 <https://bugs.schedmd.com/show_bug.cgi?id=16484> although they claimed it was resolved by using auto-nvml which would preclude this node. Here’s my very simple bash script that just srun's a sleep 20 so that it runs long enough to observe it. > #!/bin/bash > #rtx 2 > for((i=1;i<=1;++i)) ; > do > srun --gres gpu:rtx:2 --reservation=shardtest -J "RTX $i" sleep 20 & > done > sleep 1 > #shard 48 > for((i=1;i<=43;++i)) ; > do > srun --gres shard:1 --reservation=shardtest -J "shard $i" sleep 20 & > done > sleep 5 ; scontrol show node gpu03 Curious if anyone has any ideas beyond try a newer release? Thanks, Reed

2 2

sbatch and cgroup v2
by Dietmar Rieder 29 Feb '24

29 Feb '24

Hi, I'm new to slrum, but maybe someone can help me: I'm trying to restrict the CPU usage to the actually requested/allocated resources using cgroup v2. For this I made the following settings in slurmd.conf: ProctrackType=proctrack/cgroup TaskPlugin=task/cgroup,task/affinity And in cgroup.conf CgroupPlugin=cgroup/v2 CgroupAutomount=yes ConstrainCores=yes ConstrainRAMSpace=yes ConstrainDevices=yes AllowedRAMSpace=98 cgroup v2 seems to be active on the compute node: # mount | grep cgroup cgroup2 on /sys/fs/cgroup type cgroup2 (rw,nosuid,nodev,noexec,relatime,nsdelegate,memory_recursiveprot) # cat /sys/fs/cgroup/cgroup.subtree_control cpuset cpu io memory pids # cat /sys/fs/cgroup/system.slice/cgroup.subtree_control cpuset cpu io memory pids Now, when I use sbatch to submit the following test script, the python script which is started from the batch script is utilizing all CPUs (96) at 100% on the allocated node, although I only ask for 4 cpus (--cpus-per-task=4). I'd expect that the task can not use more that these 4. #!/bin/bash #SBATCH --output=/local/users/appadmin/test-%j.log #SBATCH --job-name=test #SBATCH --chdir=/local/users/appadmin #SBATCH --cpus-per-task=4 #SBATCH --ntasks=1 #SBATCH --nodes=1 #SBATCH --mem=64gb #SBATCH --time=4:00:00 #SBATCH --partition=standard #SBATCH --gpus=0 #SBATCH --export #SBATCH --get-user-env=L export PATH=/usr/local/bioinf/jupyterhub/bin:/usr/local/bioinf/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/usr/local/bioinf/miniforge/condabin source .bashrc conda activate test python test.py The python code in test.py is the following using the cpu_load_generator package from [1]: #!/usr/bin/env python import sys from cpu_load_generator import load_single_core, load_all_cores, from_profile load_all_cores(duration_s=120, target_load=1) # generates load on all cores Interestingly, when I use srun to launch an interactive job, and run the python script manually, I see with top that only 4 cpus are running at 100%. And I also python errors thrown when the script tries to start the 5th process (which makes sense): File "/usr/local/bioinf/miniforge/envs/test/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap self.run() File "/usr/local/bioinf/miniforge/envs/test/lib/python3.12/multiprocessing/process.py", line 108, in run self._target(*self._args, **self._kwargs) File "/usr/local/bioinf/miniforge/envs/test/lib/python3.12/site-packages/cpu_load_generator/_interface.py", line 24, in load_single_core process.cpu_affinity([core_num]) File "/usr/local/bioinf/miniforge/envs/test/lib/python3.12/site-packages/psutil/__init__.py", line 867, in cpu_affinity self._proc.cpu_affinity_set(list(set(cpus))) File "/usr/local/bioinf/miniforge/envs/test/lib/python3.12/site-packages/psutil/_pslinux.py", line 1714, in wrapper return fun(self, *args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/local/bioinf/miniforge/envs/test/lib/python3.12/site-packages/psutil/_pslinux.py", line 2213, in cpu_affinity_set cext.proc_cpu_affinity_set(self.pid, cpus) OSError: [Errno 22] Invalid argument What am I missing, why are the CPU resources not restricted when I use sbatch? Thanks for any input or hint Dietmar [1]: https://pypi.org/project/cpu-load-generator/

3 7

slurmdbd error - Symbol `slurm_conf' has different size in shared object
by Miriam Olmi 29 Feb '24

29 Feb '24

Hi all, I am having some issue with the new version of slurm 23.11.0-1. I had already installed and configured slurm 23.02.3-1 on my cluster and all the services were active and running properly. Following the instructions of the official SLURM webpage, for the moment I upgrated only the slurmdbd service. In principle the cluster should be able to work properly if the slurmdbd has a higher version with respect to slurmctld and slurmd. Unfortunately the slurmdbd service fails to start with the following status: slurmdbd.service - Slurm DBD accounting daemon Loaded: loaded (/usr/lib/systemd/system/slurmdbd.service; enabled; vendor preset: disabled) Active: failed (Result: core-dump) since Wed 2024-02-28 10:05:53 CET; 9min ago Process: 534938 ExecStart=/usr/sbin/slurmdbd -D -s $SLURMDBD_OPTIONS (code=dumped, signal=SEGV) Main PID: 534938 (code=dumped, signal=SEGV) Feb 28 10:05:53 slurm-db systemd[1]: Started Slurm DBD accounting daemon. Feb 28 10:05:53 slurm-db slurmdbd[534938]: /usr/sbin/slurmdbd: Symbol `slurm_conf' has different size in shared object, consider re-linking Feb 28 10:05:53 slurm-db systemd[1]: slurmdbd.service: Main process exited, code=dumped, status=11/SEGV Feb 28 10:05:53 slurm-db systemd[1]: slurmdbd.service: Failed with result 'core-dump'. Can anyone help me? Thanks in advance, Miriam

2 4

salloc+srun vs just srun
by wdennis＠nec-labs.com 29 Feb '24

29 Feb '24

Hi list, In our institution, our instructions to users who want to spawn an interactive job (for us, a bash shell) have always been to do "srun ..." from the login node, which has always been working well for us. But when we had a recent Slurm training, the SchedMD folks advised us to use "salloc" and then "srun" to do interactive jobs. I tried this today, "salloc" gave me a shell on a server, the same as srun does, but then when I tried to "srun [programname]" it hung there with no output. Of course when I tried "srun [programname] &" it spawned the background job, and gave me back a prompt. Either time I had to Ctrl-C the running srun job, and got no output other than the srun/slurmstepd termination output. I think I read somewhere that directly invoking srun creates an allocation; why then would I want to do an initial salloc, and then srun? (i the case that I want a foreground program, such as a bash shell) I have surveyed some other institution's Slurm interactive jobs documentation for users, I see both examples of advice to run srun directly, or salloc and then srun. Please help me to understand how this is intended to work, and if we are "doing it wrong" :) Thanks, Will

3 4

canonical way to run longer shell/bash interactive job (instead of srun inside of screen/tmux at front-end)?
by Josef Dvoracek 28 Feb '24

28 Feb '24

What is the recommended way to run longer interactive job at your systems? Our how-to includes starting screen at front-end node and running srun with bash/zsh inside, but that indeed brings dependency between login node (with screen) and the compute node job. On systems with multiple front-ends users need to remember the login node where they have their screen session.. Are you anybody using something more advanced and still understandable by casual user of HPC? (I know Open On Demand, but often the use of native console has certain benefits. ) cheers josef

7 9

Enforcing relative resource restrictions in submission script
by Matthew R. Baney 28 Feb '24

28 Feb '24

Hello Slurm users, I'm trying to write a check in our job_submit.lua script that enforces relative resource requirements such as disallowing more than 4 CPUs or 48GB of memory per GPU. The QOS itself has a MaxTRESPerJob of cpu=32,gres/gpu=8,mem=384G (roughly one full node), but we're looking to prevent jobs from "stranding" GPUs, e.g., a 32 CPU/384GB memory job with only 1 GPU. I might be missing something obvious, but the rabbit hole I'm going down at the moment is trying to check all of the different ways job arguments could be set in the job descriptor. i.e., the following should all be disallowed: srun --gres=gpu:1 --mem=49G ... (tres_per_node, mem_per_node set in the descriptor) srun --gpus=1 --mem-per-gpu=49G ... (tres_per_job, mem_per_tres) srun --gres=gpu:1 --ntasks-per-gpu=5 ... (tres_per_node, num_tasks, ntasks_per_tres) srun --gpus=1 --ntasks=2 --mem-per-cpu=25G ... (tres_per_job, num_tasks, mem_per_cpu) ... Essentially what I'm looking for is a way to access the ReqTRES string from the job record before it exists, and then run some logic against that i.e., if (CPU count / GPU count) > 4 or (mem count / GPU count) > 48G, error out. Is something like this possible? Thanks, Matthew -- Matthew Baney Assistant Director of Computational Systems mbaney(a)umd.edu | (301) 405-6756 University of Maryland Institute for Advanced Computer Studies 3154 Brendan Iribe Center 8125 Paint Branch Dr. College Park, MD 20742

2 1

User-facing documentation on shard use
by wdennis＠nec-labs.com 28 Feb '24

28 Feb '24

Hello list, We have just enabled "gres/shard" in order to enable sharing of GPUs on our cluster. I am now looking for examples of user-facing documentation on this feature. If anyone has something, and can send a URL or other example, I'd appreciate it. Thanks, Will

1 0

Partition, Qos Limits & Scheduling of large jobs
by Muck, Katrin 28 Feb '24

28 Feb '24

Hi everyone! I read the slurm documentation about qos, resource limits, scheduling and priority now multiple times and even looked into the slurm source but I'm still not sure if I got everything correctly, so this is why I decided to ask here ... The problem: we see the effect that sometimes larger jobs with e.g. 16 gpus in our (small) gpu queue get delayed and shifted back without any reason that is apparent to us. small jobs that only use e.g. 1 or 2 gpus get scheduled much quicker even though they have a runtime of 3 days ... What we want to do: - We have a number of nodes with 2x gpus that are usable by the users of our cluster - Some of these nodes belong to so called 'private projects'. Private projects have higher priority than other projects. Attached to that is a contingent of nodes & "guaranteed" nodes e.g. they could have a contingent of 4 nodes (8 gpus) and e.g. 2 "guaranteed" nodes (4 gpus) - Guaranteed nodes are nodes that should always be kept idle for the private project, so users of the private project can immediately schedule work on those nodes - The other nodes are shared with other projects in general if they are not "in use" How we are currently doing this (it has history): Lets assume we have 50 nodes and 100 gpus. - We have a single partition for all gpu nodes (e.g. 50 nodes) - Private projects have private queues with a very high priority and a gres limit of the number of gpus they reserved (e.g. 10 nodes -> 20 gpus) - Normal projects only have access to the public queue and schedule work there. - This public queue has an upper gres limit of "total number of gpus" - "guaranteed gpus of all private projects" (e.g. 50 - 10 nodes -> 40 nodes -> 80 gpus). Regarding the scheduler, we currently use the following settings: SchedulerType=sched/backfill SelectType=select/cons_tres SelectTypeParameters=CR_CORE_MEMORY,CR_CORE_DEFAULT_DIST_BLOCK,CR_ONE_TASK_PER_CORE SchedulerParameters=defer,max_sched_time=4,default_queue_depth=1000,partition_job_depth=500,enable_user_top,bf_max_job_user=20,bf_interval=120,bf_window=4320,bf_resolution=1800,bf_continue Partition/queue depth is deliberately set high at the moment to avoid problems with jobs not even being examined. The problem in more detail: One of the last jobs (16 gpus needed) we diagnosed had a approximate start time that was beyond all end times of running/scheduled jobs - like jobs on feb 22 would release more than enough gpus so the job could have been immediately scheduled afterwards, but the start time was still feb 23. Priority wise the job had the highest priority of all pending jobs for the partition. When we turned on scheduler debugging and increased log levels, we observed the following messages for this job: JobId=xxxx being held, if allowed the job request will exceed QOS xxxxx group max tres(gres/gpu) limit yy with already used yy + requested 16 followed by sched: JobId=2796696 delayed for accounting policy So to us this meant that the scheduler was always hitting the qos limits, which makes sense because the usage is always very high in the gpu queue and thus the job wasn't scheduled ... At first we were worried, that this meant that "held"/"delayed" jobs like this would never actually get scheduled when contention is high enough e.g. small jobs getting backfilled in and thus qos limits stay at max for a long time. But for some reason we could not determine the job eventually got scheduled at one point and then ran at the scheduled start time. Open Questions: - why it couldn't be scheduled in the first place. initially we thought (from the source code i looked into) the "delayed for accounting policy" prevents further scheduling in general, but since it was scheduled this assumption must be wrong? - why it was scheduled at some point. when it was scheduled, contention was still high and the qos limits definitely still applied - how we could modify the current setup so that the scheduling of larger jobs becomes "better" and more reproducible/explainable Apart from all of this I'm also asking myself if there is maybe a better way to setup a system that works the way we want? This got a bit long but I hope its clear enough :) Kind regards, Katrin

1 0

"Optimal" slurm configuration
by Max Grönke 26 Feb '24

26 Feb '24

Hello! In our current cluster the workflows are quite diverse (bunch of large, long (24-72h) jobs; medium size <4h job; and many small 1 node jobs). The current priority is fair share only (averaged on a ~few months timescale). For the new setup we would like to (1) discourage the 1 node jobs [especially the long ones] (2) have a fair use of the machine for the different cases. In particular, we found that the waiting time / runtime ratio is very large for the ~1-4h jobs. For this reason, I have been trying to read up on strategies in terms of partition / priority politics but could not find any good documentation in this regard (as https://slurm.schedmd.com/documentation.html is more focused on the technical aspects). Does anyone know any good references for this? In particular, we were thinking about (a) somehow have a long-term as well as shorter term fairshare count, i.e., lowering the priority if a user has something running already or (b) introduce a "small" partition for the <4h jobs with higher priority but we're unsure if this will block all the larger jobs to run... Any help or reading material is highly appreciated. Many thanks, Max

2 1

FAQ describing how to hold a job ignores scontrol subcommands specifically for that purpose
by urbanjost 25 Feb '24

25 Feb '24

There are scontrol subcommands uhold/hold/release/requeuehold that are ignored when describing how to place a job on hold in FAQ 21; and it is never explained why the method described therein is the best method, it just states it is. Does anyone know why the FAQ method is better than using the subcommands? Is it because the PRIORITY and/or NICE values are not altered (maybe)? The question is also about Running but the answer is just about Starting and not Suspending which is not quite as clear (I think "running" should be "starting" to make that clear; and/or how to suspend should be described as well). If the answer is not clear to anyone, I might turn this into a request for clarification in the Slurm bugzilla as a documentation change request but wanted to see if this was already clear to anyone and I am missing something. From FAQ: 21. How can I temporarily prevent a job from running (e.g. place it into a hold state)? The easiest way to do this is to change a job's earliest begin time (optionally set at job submit time using the --begin option). The example below places a job into hold state (preventing its initiation for 30 days) and later permitting it to start now. <METHOD I> $ scontrol update JobId=1234 StartTime=now+30days ... later ... $ scontrol update JobId=1234 StartTime=now Note: Empirically in METHOD I the JobId can be a <job_list> , which I initially thought required single JobIDs. No explanation is given on why METHOD I is best; and there are other methods that seem more intuitive. I wonder what is undesirable about the following method which I have been using -- using the scontrol(1) subcommands hold/uhold/release/requeuehold. <METHOD II> $ scontrol hold <job_list> # advantage to administrator as user cannot change $ scontrol uhold <job_list> $ scontrol release <job_list> Examples: $ scontrol uhold jobname=JOB_NAME $ scontrol uhold '[100-200],300,500' Using uhold the "Reason" changes to something easily identifying the job is being held, as "Reason=None" became "Reason=JobHeldUser which seems better that Method I in that regard. The downside might be PRIORITY changed to zero and then went to a very large value when released? Another method appears to be that setting PRIORITY to zero also places jobs in hold. <METHOD III> $ scontrol update jobid=373 Priority=0 $ scontrol release jobid=373 # sets to a very high value $ scontrol update jobid=373 Priority=11111 # put back to lower desired value Once lowered, does an optional setting prevent a user from raising PRIORITY(?) The manual says Only the Slurm administrator or root can increase job's priority. At least on my machine the "release" buts the priority to a very high value, and a regular user can lower the value back to the (probably) lower original value. I did not see it happening but there are some statements in the documentation that make me think not only PRIORITY but perhaps the NICE value might be changed by METHOD II and METHOD III, although I could not get the NICE value to be inadvertently changed. Sent with [Proton Mail](https://proton.me/) secure email.

1 1

slurm-23.11.3-1 with X11 and zram causing permission errors: error: _forkexec_slurmstepd: slurmstepd failed to send return code got 0: Resource temporarily unavailable; Requeue of JobId=530 returned an error: Only batch jobs are accepted or processed
by Robert Kudyba 24 Feb '24

24 Feb '24

We switched over from using systemctl for tmp.mount and change to zram, e.g., modprobe zram echo 20GB > /sys/block/zram0/disksize mkfs.xfs /dev/zram0 mount -o discard /dev/zram0 /tmp srun with --x11 was working before changing this. We're on RHEL 9. slurmctld logs show this whenever --x11 is used with srun: [2024-02-23T20:22:43.442] [529.extern] error: setup_x11_forward: failed to create temporary XAUTHORITY file: Permission denied [2024-02-23T20:22:43.442] [529.extern] error: x11 port forwarding setup failed [2024-02-23T20:22:43.442] error: _forkexec_slurmstepd: slurmstepd failed to send return code got 0: Resource temporarily unavailable [2024-02-23T20:22:43.443] Could not launch job 529 and not able to requeue it, cancelling job [2024-02-23T20:26:15.881] [530.extern] error: setup_x11_forward: failed to create temporary XAUTHORITY file: Permission denied [2024-02-23T20:26:15.881] [530.extern] error: x11 port forwarding setup failed [2024-02-23T20:26:15.882] error: _forkexec_slurmstepd: slurmstepd failed to send return code got 0: Resource temporarily unavailable [2024-02-23T20:26:15.883] Could not launch job 530 and not able to requeue it, cancelling job slurmd log entries from a node: [2024-02-23T20:26:15.859] sched: _slurm_rpc_allocate_resources JobId=530 NodeList=2402-node005 usec=1800 [2024-02-23T20:26:15.882] _slurm_rpc_requeue: Requeue of JobId=530 returned an error: Only batch jobs are accepted or processed [2024-02-23T20:26:15.883] _slurm_rpc_kill_job: REQUEST_KILL_JOB JobId=530 uid 0 [2024-02-23T20:26:15.962] _slurm_rpc_complete_job_allocation: JobId=530 error Job/step already completing or completed srun -v --pty -t 0-4:00 --x11 --mem=10g srun: defined options srun: -------------------- -------------------- srun: account : me srun: mem : 10G srun: nodelist : our-node srun: pty : srun: time : 04:00:00 srun: verbose : 1 srun: x11 : all srun: -------------------- -------------------- srun: end of defined options srun: Waiting for resource configuration srun: error: Nodes our-node are still not ready srun: error: Something is wrong with the boot of the nodes. slurm.conf has PrologFlags=x11 set. /usr/bin/xauth is installed on each compute node. Is this a known issue with zram or is that just a red herring and there's something else wrong?

3 5

Slurm version 23.11.4 is now available
by Tim McMullan 22 Feb '24

22 Feb '24

We are pleased to announce the availability of Slurm version 23.11.4. The 23.11.4 release includes a number of fixes to stability and various bug fixes. Some notable changes include that VSZ is no longer being reported when using cgroup/v2 (this is not provided by the kernel), a warning has been added if using select/linear and tolology/tree together as this will not be supported in the next major release, and a backwards compatibility issue that caused jobs using --gpus to be rejected when submitted from 23.02 or 22.05. Slurm can be downloaded from https://www.schedmd.com/downloads.php . -Tim > * Changes in Slurm 23.11.4 > ========================== > -- Fix a memory leak when updating partition nodes. > -- Don't leave a partition around if it fails to create with scontrol. > -- Fix segfault when creating partition with bad node list from scontrol. > -- Fix preserving partition nodes on bad node list update from scontrol. > -- Fix assertion in developer mode on a failed message unpack. > -- Fix repeat POWER_DOWN requests making the nodes available for ping. > -- Fix rebuilding job alias_list on restart when nodes are still powering up. > -- Fix INVALID nodes running health check. > -- Fix cloud/future nodes not setting addresses on invalid registration. > -- scrun - Remove the requirement to set the SCRUN_WORKING_DIR environment > variable. This was a regression in 23.11. > -- Add warning for using select/linear with topology/tree. > This combination will not be supported in the next major version. > -- Fix health check program not being run after first pass of all nodes when > using MaxNodeCount. > -- sacct - Set process exit code to one for all errors. > -- Add SlurmctldParameters=disable_triggers option. > -- Fix issue running steps when the allocation requested an exclusive > allocation shards along with shards. > -- Fix cleaning up the sleep process and the cgroup of the extern step if > slurm_spank_task_post_fork returns an error. > -- slurm_completion - Add missing --gres-flags= options > multiple-tasks-per-sharing and one-task-per-sharing. > -- scrun - Avoid race condition that could cause outbound network > communications to incorrectly rejected with an incomplete packet error. > -- scrun - Gracefully handle kernel giving invalid expected number of incoming > bytes for a connection causing incoming packet corruption resulting in > connection getting closed. > -- srun - return 1 when a step lauch fails > -- scrun - Avoid race condition that could cause deadlock during shutdown. > -- Fix scontrol listpids to work under dynamic node scenarios. > -- Add --tres-bind to --help and --usage output. > -- Add --gres-flags=allow-task-sharing to allow GPUs to still be accessible > among all tasks when binding GPUs to specific tasks. > -- Fix issue with CUDA_VISIBLE_DEVICES showing the same MIG device for all > tasks when using MIGs with --tres-per-task or --gpus-per-task. > -- slurmctld - Prevent a potential hang during shutdown/reconfigure if the > association cache thread was previously shut down. > -- scrun - Avoid race condition that could cause scrun to hang during > shutdown when connections have pending events. > -- scrun - Avoid excessive polling of connections during shutdown that could > needlessly cause 100% CPU usage on a thread. > -- sbcast - Use user identity from broadcast credential instead of looking it > up locally on the node. > -- scontrol - Remove "abort" option handling. > -- Fix an error message referring to the wrong RPC. > -- Fix memory leak on error when creating dynamic nodes. > -- Fix a slurmctld segfault when a cloud/dynamic node changes hostname on > registration. > -- Prevent a slurmctld deadlock if the gpu plugin fails to load when > creating a node. > -- Change a slurmctld fatal() to an error() when attempting to create a > dynamic node with a global autodetect set in gres.conf. > -- Fix leaving node records on error when creating nodes with scontrol. > -- scrun/sackd - Avoid race condition where shutdown could deadlock. > -- Fix a regression in 23.02.5 that caused pam_slurm_adopt to fail when > the user has multiple jobs on a node. > -- Add GLOB_SILENCE flag that silences the error message which will display if > an include directive attempts to use the "*" wildcard. > -- Fix jobs getting rejected when submitting with --gpus option from older > versions of job submission commands (23.02 and older). > -- cgroup/v2 - Return 0 for VSZ. Kernel cgroups do not provide this metric. > -- scrun - Avoid race condition where outbound RPCs could be corrupted. > -- scrun - Avoid race condition that could cause a crash while compiled in > debug mode. > -- gpu/rsmi - Disable gpu usage statistics when not using ROCM 6.0.0+ > -- Fix stuck processes and incorrect environment when using --get-user-env. > -- Avoid segfault in the slurmdbd when TrackWCKey=no but you are still using > use WCKeys. > -- Fix ctld segfault with TopologyParam=RoutePart and no partition defined. > -- slurmctld - Fix missing --deadline handling for jobs not evaluated by the > schedulers (i.e. non-runnable, skipped for other reasons, etc.). > -- Demote some eio related logs from error to verbose in user commands. These > are not generally actionable by the user and are easilly generated by port > scanning a machine running srun. > -- Make sprio correctly print array tasks that have not yet been split out. > -- topology/block - Restrict the number of last-level blocks in any allocation. -- Tim McMullan Release Management, Support, and Development SchedMD LLC - Commercial Slurm Development and Support

1 0

RHEL 8.9+SLURM-23.11.3+MLNX_OFED_LINUX-23.10-1.1.9.0+ OpenMPI-5.0.2
by S L 20 Feb '24

20 Feb '24

Hi All, We're currently in the process of setting up SLURM on a RHEL 8.9 based cluster. Here's a summary of the steps we've taken so far: Installed MLNX OFED ConnectX-5.2. Compiled and installed PMiX and UCX. Compiled and installed Slurm with PMiX_v4 and UCX support. Compiled OpenMPI with SLURM, PMIx, libevent, and hwloc support. All compute nodes are reachable via the IB network. *Problem:* While hello world MPI jobs are working fine on multiple nodes, the jobs are not utilizing Infiniband. srun --mpi=pmix -N2 -n2 --ntasks-per-node=2 ./hello > log.out 2>&1 Output from srun --mpi=list: MPI plugin types are... none cray_shasta pmi2 pmix specific pmix plugin versions available: pmix_v4 Could someone please point me in the right direction on how to troubleshoot this issue? Thank you for your assistance. Sudhakar

2 1

URL for how to do for SLURM accounting setup
by John Joseph 20 Feb '24

20 Feb '24

Dear ALL, Good morning we were able to setup a test SLURM based system, with 4 nodes , Ubuntu 22.04 LTS and we were able to run COMSOL using "comsol batch" command Now we plan to have accounting https://slurm.schedmd.com/accounting.html Like to reach out and get guidance on any tutorial or how to do documentation on setting up accounting Appreciate your support Thanks Joseph John

2 2

CPU utilisation using two commands scontrol association and sreport makes a huge difference
by prachikakade.lit＠gmail.com 20 Feb '24

20 Feb '24

Dear Team, I created a small cluster of 3 nodes on my VM ware to work on the CPU utilization concept. I created a user name= hpcuser01, and allocated GrpTresMin=cpu=59000040 -> CPU minutes and gpu=0 Now, when I checked his utilization using scontrol association cmd # scontrol show ass user=hpcuser01 | grep -n15 hpcuser01 "" Output: QOS=hpcuser01_test(11) UsageRaw=5789263754.153907 GrpJobs=N(0) GrpJobsAccrue=N(5) GrpSubmitJobs=N(5) GrpWall=N(1188290.27) GrpTRES=cpu=N(0),mem=N(0),energy=N(0),node=N(0),billing=N(0),fs/disk=N(0),vmem=N(0),pages=N(0),gres/gpu=N(0) GrpTRESMins=cpu=59000040(59001152),mem=N(293429427523),energy=N(0),node=N(1993778),billing=N(96073808),fs/disk=N(0),vmem=N(0),pages=N(0),gres/gpu=1(0) GrpTRESRunMins=cpu=N(0),mem=N(0),energy=N(0),node=46080(0),billing=N(0),fs/disk=N(0),vmem=N(0),pages=N(0)"" Here we can see that GrpTRESMins=cpu=59000040(59001152) after converting in Hours ( value /60) it will be: 983334(983352) CPU hours This means 983352 CPU hours have been used by hpcuser01 but, when I checked his utilization using sreport cmd #sreport -t hour --tres=cpu,gres/gpu,billing cluster AccountUtilizationByUser user=hpcuser01 start=2021-01-01T00:00:00 end=2024-02-20T23:59:59 -P | grep -iw -e billing -e cpu -e gres/gpu Output: cpu|979403 billing|1594902 gres/gpu|0 here I found that his utilisation is : 979403 I could like to understand why there is so much of a difference using two cmds to find out utilisation. Almost the difference is 979403 - 983352 = 3949 Request you to help me out with this. Regards, Prachi

1 0

Slurm Power Saving Guide: Why doesnt slurm mark as failed when resumeProgram returns =/= 0
by Xaver Stiensmeier 19 Feb '24

19 Feb '24

Dear slurm-user list, I had cases where our resumeProgram failed due to temporary cloud timeouts. In that case the resumeProgram returns a value =/= 0. Why does Slurm still wait until resumeTimeout instead of just accepting the startup as failed which then should lead to a rescheduling of the job. Is there some way to achieve the described effect i.e. tell Slurm: "You can stop waiting, the node won't come alive." or am I missing the correct way how this should be handled in Slurm? Best regards, Xaver

1 0

slurmdbd 17.02: "cluster not registered" (but things work)
by Matthias Leopold 19 Feb '24

19 Feb '24

Hi, I need to take care of a 17.02 Slurm cluster (I'm preparing it for upgrades). I see that slurmdbd logs various "cluster not registered" messages at startup (DBD_CLUSTER_TRES,DBD_JOB_START,DBD_STEP_START), but I don't see a real problem. Accounting works. Do I have to worry? Can this be related to upper/lower case issues with ClusterName? For some reason ClusterName is all upper case in slurm.conf. According to docs (and my experience with Slurm 21.08) though, this should be OK. Thx for help Matthias

1 0

Recover Batch Script Error
by Jason Simms 16 Feb '24

16 Feb '24

Hello all, I've used the "scontrol write batch_script" command to output the job submission script from completed jobs in the past, but for some reason, no matter which job I specify, it tells me it is invalid. Any way to troubleshoot this? Alternatively, is there another way - even if a manual database query - to recover the job script, assuming it exists in the database? sacct --jobs=38960 JobID JobName Partition Account AllocCPUS State ExitCode ------------ ---------- ---------- ---------- ---------- ---------- -------- 38960 amr_run_v+ tsmith2lab tsmith2lab 72 COMPLETED 0:0 38960.batch batch tsmith2lab 40 COMPLETED 0:0 38960.extern extern tsmith2lab 72 COMPLETED 0:0 38960.0 hydra_pmi+ tsmith2lab 72 COMPLETED 0:0 scontrol write batch_script 38960 job script retrieval failed: Invalid job id specified Warmest regards, Jason -- *Jason L. Simms, Ph.D., M.P.H.* Manager of Research Computing Swarthmore College Information Technology Services (610) 328-8102 Schedule a meeting: https://calendly.com/jlsimms

4 3

Need help managing licence
by Sylvain MARET 16 Feb '24

16 Feb '24

Hello everyone ! Recently our users bought a cplex dynamic license and want to use it on our slurm cluster. I've installed the paid version of cplex within modules so authorized user can load it with a simple module load cplex/2111 command but I don't know how to manage and ensure slurm doesn't launch a job if 20 people are already running code with this license. How do you guys manage paid licenses on your cluster ? Any advice would be appreciated ! Regards, Sylvain Maret

2 1

MPI/PMIx Issues after 23.11 Update
by Oliver Smith 16 Feb '24

16 Feb '24

Hi all, We’re running a small slurm dev cluster on Ubuntu and are facing issues with MPI/PMIx after upgrading slurm from 23.02.5 to 23.11.3. The first job step to use MPI within a job fails roughly 80% of the time but following attempts to use MPI within the same job work fine. For the failing job step we see this error after hitting the MPI timeout: slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_reset_if_to: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:741: 0x55a03f8d7a90: collective timeout seq=0 slurmstepd: error: mpi/pmix_v4: pmixp_coll_log: hpc-d-msh-01a02 [1]: pmixp_coll.c:286: Dumping collective state slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:759: 0x55a03f8d7a90: COLL_FENCE_RING state seq=0 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:762: my peerid: 1:hpc-d-msh-01a02 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:769: neighbor id: next 0:hpc-d-msh-01a01, prev 0:hpc-d-msh-01a01 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:778: Context ptr=0x55a03f8d7b08, #0, in-use=0 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:778: Context ptr=0x55a03f8d7b40, #1, in-use=0 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:778: Context ptr=0x55a03f8d7b78, #2, in-use=1 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:787: seq=0 contribs: loc=1/prev=1/fwd=0 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:791: neighbor contribs [2]: slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:824: done contrib: hpc-d-msh-01a01 slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:826: wait contrib: - slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:828: status=PMIXP_COLL_RING_FINILIZE slurmstepd: error: mpi/pmix_v4: pmixp_coll_ring_log: hpc-d-msh-01a02 [1]: pmixp_coll_ring.c:831: buf (offset/size): 0/33362 [hpc-d-msh-01a01.tds.hpc.barf1.com:47652] pml_ucx.c:178 Error: Failed to receive UCX worker address: Not found (-13) [hpc-d-msh-01a01.tds.hpc.barf1.com:47652] pml_ucx.c:477 Error: Failed to resolve UCX endpoint for rank 31 [hpc-d-msh-01a01:47652] *** An error occurred in MPI_Send [hpc-d-msh-01a01:47652] *** reported by process [683360612,0] [hpc-d-msh-01a01:47652] *** on communicator MPI_COMM_WORLD [hpc-d-msh-01a01:47652] *** MPI_ERR_OTHER: known error not in list [hpc-d-msh-01a01:47652] *** MPI_ERRORS_ARE_FATAL (processes in this communicator will now abort, [hpc-d-msh-01a01:47652] *** and potentially your MPI job) slurmstepd: error: *** STEP 1890.0 ON hpc-d-msh-01a01 CANCELLED AT 2024-02-14T16:14:52 *** OpenMPI/PMIx versions have not changed and downgrading slurm to 23.02.5 seems to resolve the issue. We’d appreciate any pointers anyone might have. Thanks Oli This email comprises confidential information of Mercedes-Benz Grand Prix Limited ("MGP") unless it contains an explicit statement to the contrary made by an authorised representative of MGP. Contracts may only be concluded on behalf of MGP by its authorised signatories and not solely by email communication. No employee, agent, contractor, consultant and/or other representative of MGP is authorised to conclude any legally binding agreement on behalf of MGP by email alone without the express prior written confirmation of two authorised signatories of MGP. Mercedes-Benz Grand Prix Limited. Registered in England No. 787446. Registered Office at Mercedes-Benz Grand Prix Limited, Operations Centre, Brackley, Northants NN13 7BD. Note: The MGP Legal Department also acts on behalf of Mercedes-Benz Motorsport Limited ("MBM") and the above notice applies mutatis mutandis in respect of all email communications of MBM. MBM: Mercedes-Benz Motorsport Limited. Registered in England No. 13057973. Registered office at Mercedes-Benz Motorsport Limited, Lauda Drive, Brackley, Northants NN13 7BD. Please consider the environment before printing this email.

1 0

slurmctld: slurm_bufs_sendto(msg_type=SRUN_STEP_SIGNAL) failed: Connection reset by peer
by Rike-Benjamin Schuppner 15 Feb '24

15 Feb '24

Hi, I am getting the following error in the logs whenever I run a few srun jobs in a batch. Jan 25 11:24:03 slurmctl.XYZ slurmctld[272961]: slurmctld: debug: _send_timeout: Socket POLLERR: Connection reset by peer Jan 25 11:24:03 slurmctl.XYZ slurmctld[272961]: slurmctld: error: slurm_send_node_msg: [socket:[921897]] slurm_bufs_sendto(msg_type=SRUN_STEP_SIGNAL) failed: Connection reset by peer Jan 25 11:24:03 slurmctl.XYZ slurmctld[272961]: slurmctld: debug: laying out the 1 tasks on 1 hosts compute2 dist 1 The slurm version is 23.11.3 and an example sbatch file is: #!/bin/bash #SBATCH --job-name=slurm_test #SBATCH --mem=1gb #SBATCH --time=00:05:00 #SBATCH --output=slurm_test_%j.log pwd; hostname; date srun --exclusive -c1 -N1 -n1 bash -c "hostname ; sleep 10" & srun --exclusive -c1 -N1 -n1 bash -c "hostname ; sleep 10" & wait The more sruns I have in the script (&-backgrounded or not), the more the error shows up. Is there anything I could do to fix this error? Best /rike

2 1

Naive SLURM question: equivalent to LSF pre-exec
by Cutts, Tim 14 Feb '24

14 Feb '24

Hi, I apologise if I’ve failed to find this in the documentation (and am happy to be told to RTFM) but a recent issue for one of my users resulted in a question I couldn’t answer. LSF has a feature called a Pre-Exec where a script executes to check whether a node is ready to run a task. So, you can run arbitrary checks and go back to the queue if they fail. For example, if I have some automounted filesystems, and I want to be able to check for failure of the automounted, in an LSF world, I can do: bsub -E “test -f /nfs/someplace/file_I_know_exists” my_job.sh What’s the equivalent in SLURM? Thanks, Tim -- Tim Cutts Scientific Computing Platform Lead AstraZeneca Find out more about R&D IT Data, Analytics & AI and how we can support you by visiting our Service Catalogue<https://azcollaboration.sharepoint.com/sites/CMU993> | ________________________________ AstraZeneca UK Limited is a company incorporated in England and Wales with registered number:03674842 and its registered office at 1 Francis Crick Avenue, Cambridge Biomedical Campus, Cambridge, CB2 0AA. This e-mail and its attachments are intended for the above named recipient only and may contain confidential and privileged information. If they have come to you in error, you must not copy or show them to anyone; instead, please reply to this e-mail, highlighting the error to the sender and then immediately delete the message. For information about how AstraZeneca UK Limited and its affiliates may process information, personal data and monitor communications, please see our privacy notice at www.astrazeneca.com<https://www.astrazeneca.com>

3 2

job_submit.lua - uid in Docker cluster
by Loris Bennett 14 Feb '24

14 Feb '24

Hi, Having used https://github.com/giovtorres/slurm-docker-cluster successfully a couple of years ago to develop a job_submit.lua plugin, I am trying to do this again. However, the plugin which works on our current cluster (CentOS 7.9, Slurm 23.02.7) fails in the Docker cluster (Rocky 8.9, Slurm 23.02.7) with error: job_submit/lua: /etc/slurm/job_submit.lua: /usr/share/lua/5.3/posix/deprecated.lua:453: bad argument #1 to 'getpwuid' (int expected, got number) which occurs when submit_user = posix.getpasswd(submit_user.uid) The problems is that 'submit_user.uid' has the value 0.0 and is thus not an integer. The only user within the Docker cluster is 'root'. Has anyone come across this issue? Is it to do with the Docker environment or the difference in the OS versions (Lua 5.1.4 vs. 5.3.4, lua-posix 32 vs. 33.3.1)? Cheers, Loris -- Dr. Loris Bennett (Herr/Mr) FUB-IT (ex-ZEDAT), Freie Universität Berlin

1 0

Increasing SlurmdTimeout beyond 300 Seconds
by Andrew Baughan (ITCS - Staff) 12 Feb '24

12 Feb '24

Hi, We've been experiencing issues with network saturation on our older nodes caused by storage (GPFS) backups. This causes slurmctld to loose contact with slurmd on some compute nodes and for user jobs to be killed. While the longer term solution is to replace these and upgrade the network, I'm wondering if there are any ramifications, beyond nodes with genuine issues taking longer to get marked down, by increasing SlurmdTimeout. We've already applied a modest increase which has helped but not resolved the issue and wondering if we should push it further in the interim. Kind Regards Andy Baughan HPC Systems Developer

4 3

simple question, I guess… from a newbie sysadmin
by Richard Randriatoamanana 12 Feb '24

12 Feb '24

Hi, I am trying to help a sysadmin colleague (and to understand for myself) trying to configure a new slurm server and he struggles to understand if there is an alternative way to config slurm managing job policy submission per user without necessarily installing an accounting mariadb service. I said no but maybe I misunderstand slurm admin configuration. Thanks in advance for feedback. Bests, Richard — Sent from my rotary and brandy phone ! Apologies for typos and any spelling errors

2 1

Question on using a suffix with "squeue --Format=FORMAT"
by urbanjost 10 Feb '24

10 Feb '24

I do not see how to use the suffix for the --Format option of the squeue(1) command. The description does not indicate how to tell the end of the size and period components and the beginning of the suffix, but assuming the suffix must being with something other than a period or numeric character I found all suffix values produce a null separator. A wanted to use --Format somewhat like --format=%all, producing all fields. As the --Format did not have an equivalent descriptor, (ie. something like --Format=All ) it looked like it would be required to list all the allowed keywords with a suffix of "|", which from the description I assumed would be as simple as a longer version of something like thi: squeue --Format='STDOUT:|,STDERR:|,STDIN:|,WorkDir:|' Given the description: -O, --Format=<output_format> Specify the information to be displayed. Also see the -o, --format=<output_format> option described above (which supports greater flexibility in formatting, but does not support access to all fields because we ran out of letters). Requests a comma separated list of job information to be displayed. The format of each field is "type[:[.][size][suffix]]" size Minimum field size. If no size is specified, 20 char‐ acters will be allocated to print the information. . Indicates the output should be right justified and size must be specified. By default output is left justified. suffix Arbitrary string to append to the end of the field. Note that many of these type specifications are valid only for jobs while others are valid only for job steps. Valid type specifications include: Even something like this produces no error but the suffix is always a null string: squeue --Format='STDOUT:SUFFIX,STDERR:SUFFIXSTDIN:SUFFIXWorkDir:SUFFIX So if I want to generate output with squeue --Format=FORMAT with a vertical line character between the output fields what is the proper syntax? A test bash(1) script where I could not get the SUFFIX variable to work as expected ... #!/bin/bash #@(#) Build value for squeue --Format='FORMAT' that includes all available fields with a vertical bar separating each field GOODNAMES='' BADNAMES='' SUFFIX='SUFFIX' GOODNAMES='' BADNAMES='' SUFFIX='1000:|' GOODNAMES='' BADNAMES='' SUFFIX='|' for NAME in STDOUT Account AccrueTime admin_comment AllocNodes AllocSID ArrayJobID ArrayTaskID AssocID BatchFlag BatchHost BoardsPerNode BurstBuffer BurstBufferState Cluster ClusterFeature Command Comment Container ContainerID Contiguous Cores CoreSpec CPUFreq cpus-per-task cpus-per-tres Deadline DelayBal Licenses MaxCPUs MaxNodes MCSLabel mem-per-tres MinCpus MinMemory MinTime Mintition PendingTime PreemptTime Prefer Priority PriorityLong Profile QOS Reason Riable SiblingsViableRaw Sockets SPerBoard StartTime State StateCompact STDERR STres-per-node tres-per-socket tres-per-step tres-per-task UserID UserName Wait4Sw do if squeue --Format="$NAME" >/dev/null then GOODNAMES="$GOODNAMES${NAME}:${SUFFIX}," else BADNAMES="$BADNAMES${NAME}," fi done 2>&1 #FORMAT=$(tr --squeeze-repeats ' ' ' '<<<$GOODNAMES) FORMAT=${GOODNAMES%%,} # trim off trailing comma read r c < <(stty size) # get screen width for use with cut(1) # try it squeue --Format="$FORMAT"|column -t|cut -c 1-$c # echo it echo "$FORMAT" exit Sent with [Proton Mail](https://proton.me/) secure email.

1 0

question on sbatch --prefer
by Alan Stange 10 Feb '24

10 Feb '24

Hello all, I'm somewhat new to Slurm, but long time user of other batch systems. Assume we have a simple cluster of uniform racks of systems with no special resources, and our jobs are all single cpu tasks. Lets say I have a long running job in the cluster, which needs to spawn a helper process into the cluster. We have a strong preference for this helper to run on the same cluster node as the original job, but if that node is already scheduled full, then we want this new task to be scheduled on another systems without any delay. The problem I have is that the --nodelist doesn't solve this, and, as far as I can tell, there's no option with --prefer to specify a node name as a resource, without creating a gres for every hostname in the cluster. It seems like what I'm trying to do should be achievable, but having read though the documentation and searched the archives of this list, I'm not seeing a solution. I'm hoping someone here has some experience with this and can point me in the right direction. Sincerely, Alan

3 3

Compilation question
by Sylvain MARET 10 Feb '24

10 Feb '24

Hello everyone ! I'm trying to compile slurm 22.05.11 on Rocky linux 8.7 with freeipmi support I've seen the documentation so I've done the configure step : ./configure --with-pmix=$PMIXHOME --with-ucx=$UCXHOME --with-nvml=$NVMLHOME --prefix=$SLURMHOME --with-freeipmi=/usr but when I run make I end up with the following error : /bin/sh ../../../../../libtool --tag=CC --mode=link gcc -DNUMA_VERSION1_COMPATIBILITY -g -O2 -fno-omit-frame-pointer -pthread -ggdb3 -Wall -g -O1 -fno-strict-aliasing -export-dynamic -L/usr/lib64 -lhdf5_hl -lhdf5 -lsz -lz -ldl -lm -o sh5util sh5util.o -Wl,-rpath=/softs/batch/slurm/22.05.11/lib/slurm -L../../../../../src/api/.libs -lslurmfull -ldl ../libhdf5_api.la -lpthread -lm -lresolv libtool: link: gcc -DNUMA_VERSION1_COMPATIBILITY -g -O2 -fno-omit-frame-pointer -pthread -ggdb3 -Wall -g -O1 -fno-strict-aliasing -o .libs/sh5util sh5util.o -Wl,-rpath=/softs/batch/slurm/22.05.11/lib/slurm -Wl,--export-dynamic -L/usr/lib64 -L../../../../../src/api/.libs /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so ../.libs/libhdf5_api.a -lhdf5_hl -lhdf5 -lsz -lz -ldl -lpthread -lm -lresolv -pthread -Wl,-rpath -Wl,/softs/batch/slurm/22.05.11/lib/slurm sh5util.o:(.init_array+0x0): undefined reference to `__nv_init_env' sh5util.o:(.init_array+0x8): undefined reference to `__flushz' sh5util.o:(.init_array+0x10): undefined reference to `__daz' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_transfer_unique' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_sort_key_pairs' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_xstrchr' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_unsetenvp' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_sort' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_for_each' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `__builtin__pgi_isnanld' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_get_extra_conf_path' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `__blt_pgi_ctzll' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_running_in_slurmctld' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `__c_mcopy1' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `__blt_pgi_clzll' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_create' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_count' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `__builtin_va_gparg1' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_destroy_config_key_pair' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_xfree_ptr' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_getenvp' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_free_buf' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_get_log_level' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `__c_mset8' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_xstrdup_printf' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_delete_first' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_append' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_error' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_init_buf' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_log_var' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_s_p_parse_file' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_xstrdup' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_fatal' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `__blt_pgi_popcountll' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_xstrcmp' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_s_p_hashtbl_destroy' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_s_p_hashtbl_create' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_setenvpf' /softs/batch/slurm/slurm-22.05.11/src/api/.libs/libslurmfull.so: undefined reference to `slurm_list_destroy' collect2: error: ld returned 1 exit status What can I do to resolve these undefined reference error ? Regards, Sylvain Maret

3 2

Re: Selecting only a subset of GPU's from all available GPU's
by Minulakshmi S 10 Feb '24

10 Feb '24

Hi Loris, I have different kind of GPU's in same node , and I beleive feature applies to particular node , and cant be applied to only a few GPU's connected to a single node. On Mon, 18 Dec 2023, 12:14 Loris Bennett, <loris.bennett(a)fu-berlin.de> wrote: > Hi Minu, > > Minulakshmi S <minulakshmi.s(a)gmail.com> writes: > > > I'm submitting jobs to a cluster via SLURM scheduler, and let's say I > have access to 8 GPUs in my cluster in same node. They are GPUs of type > A,B,C,D,E,F,G,H. I > > would like to submit a job that requests the use of GPUs of type A or B > or C but NOT of type D/E/F/G/H. So I need some type of OR logic with the > --gres flag. > > > > Eg : When I request GPU of type A , I can do sbatch –gres=gpu:TypeA:1, > I need to input a subset of GPU’s and let slurm schedule job utilizing one > of the GPU > > from this allowed list. > > > > Regards > > Minu > > Assuming the GPUs within a node are all of the same type, could you define > a feature for each GPU type, assign the features to the appropriate > nodes, and then run the job with > > --contraint=gpu_TypeA --gres=gpu:1 > > ? > > This is obviously rather clunky and it would be much nicer if > multiple GPUs types passed to '--gres' were ORed. > > Cheers, > > Loris > > -- > Dr. Loris Bennett (Herr/Mr) > ZEDAT, Freie Universität Berlin >

1 0

Memory used per node
by Gerhard Strangar 09 Feb '24

09 Feb '24

Hello, I'm wondering if there's a way to tell how much memory my job is using per node. I'm doing #SBATCH -n 256 srun solver inputfile When I run sacct -o maxvmsize, the result apparently is the maxmimum VSZ of the largest solver process, not the maximum of the sum of them all (unlike when calling mpirun instead). When I sstat -o TresUsageInMax, I get the memory summed up over all nodes being used. Can I get the maximum VSZ per node? Gerhard

2 1

Could not find group with gid even when they exist
by Nic Lewis 09 Feb '24

09 Feb '24

After upgrading to version 23.11.3 we started to get slammed with the following log messages from slurmctld "error: validate_group: Could not find group with gid <id>" This spans a handful of groups and repeats constantly, drowning out just about everything else. Attempting to do a lookup on the group shows that they exist on the scheduler node, same for all the submission and compute nodes. As far as I can tell, slurm should be able to locate the group in question. Jobs submitted from users within those groups go through just fine. They get scheduled, run, and clean up no problem. I'm at a loss on where to look next.

1 1

Relative QOS limits in sacctmgr?
by Chip Seraphine 09 Feb '24

09 Feb '24

Hello, TL,DR: How does the relative QOS flag work? I have a QOS and I want it to be collectively restricted to 50% of the reachable cores in the cluster. I’ve been managing this by dividing my core count to 2 to get N, and doing ‘sacctmgr update qos foobar set MaxTRES=cpu=N’. That’s fine, except that N changes frequently as worker nodes come and go, so I have an hourly cronjob that enforces this. The problem with *that* is the values you back out with ‘sacctmgr show qos’ are often in different units than you input (e.g. gigs instead of megs of ram), which means I need to do some annoying parsing, will miss minor rounding errors, etc. (Using JSON solves that problem but introduces others that I do not want to deal with.) So while looking for a better way I noticed that QOSs support a flag called “relative”, which says “the QOS limits will be treated as percentages of a cluster/partition instead of absolutes”. Perfect! However, it’s not very clear to me what that means in practice. For example, with the “relative” flag set, the MaxTRES units still seem to work the same, expect metric suffices (“G”, ”M”, etc) and won’t accept periods. Is anyone actually using this flag, and if so can you tell me what attributes it affects, and in what way? -- Chip Seraphine Grid Operations For support please use help-grid in email or slack. This e-mail and any attachments may contain information that is confidential and proprietary and otherwise protected from disclosure. If you are not the intended recipient of this e-mail, do not read, duplicate or redistribute it by any means. Please immediately delete it and any attachments and notify the sender that you have received it by mistake. Unintended recipients are prohibited from taking action on the basis of information in this e-mail or any attachments. The DRW Companies make no representations that this e-mail or any attachments are free of computer viruses or other defects.

1 0

Is there a way to list allocated/unallocated resources defined in a QoS?
by Alastair Neil 09 Feb '24

09 Feb '24

Slurm version 23.02.07 If I have a QoS defined that has a set number of say GPU devices set in the GrpTRES. Is there an easy way to generate a list of how much of the defined quota is allocated or conversely un-allocated? e.g.: Name|Priority|GraceTime|Preempt|PreemptExemptTime|PreemptMode|Flags|UsageThres|UsageFactor|GrpTRES|GrpTRESMins|GrpTRESRunMins|GrpJobs|GrpSubmit|GrpWall|MaxTRES|MaxTRESPerNode|MaxTRESMins|MaxWall|MaxTRESPU|MaxJobsPU|MaxSubmitPU|MaxTRESPA|MaxJobsPA|MaxSubmitPA|MinTRES| normal|0|00:00:00|||cluster|||1.000000|||||||||||cpu=3000,gres/gpu=20||||||| dept1|1|00:00:00|||cluster|||1.000000|cpu=256,gres/gpu:1g.10gb=16,gres/gpu:2g.20gb=8,gres/gpu:3g.40gb=8,gres/gpu:a100.80gb=8||||||||||||||||| dept2|1|00:00:00|||cluster|||1.000000|cpu=256,gres/gpu:1g.10gb=0,gres/gpu:2g.20gb=0,gres/gpu:3g.40gb=0,gres/gpu:a100.80gb=16||||||||||||||||| So dept1 and dept2 qos are set on the same partition. How can a user with access to one or other see if there are available resources in the partition?

2 1

Errors upgrading to 23.11.0
by Xaver Stiensmeier 08 Feb '24

08 Feb '24

Dear slurm-user list, I got this error: Unable to start service slurmctld: Job for slurmctld.service failed because the control process exited with error code.\nSee \"systemctl status slurmctld.service\" and \"journalctl -xeu slurmctld.service\" for details. but in slurmctld.service I see nothing suspicious: slurmctld.service - Slurm controller daemon Loaded: loaded (/lib/systemd/system/slurmctld.service; enabled; vendor preset: enabled) Drop-In: /etc/systemd/system/slurmctld.service.d └─override.conf Active: active (running) since Wed 2024-02-07 15:50:56 UTC; 19min ago Main PID: 51552 (slurmctld) Tasks: 21 (limit: 9363) Memory: 10.4M CPU: 1min 16.088s CGroup: /system.slice/slurmctld.service ├─51552 /usr/sbin/slurmctld --systemd └─51553 "slurmctld: slurmscriptd" "" "" "" "" "" "" Feb 07 15:58:21 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: sched: _slurm_rpc_allocate_resources JobId=3 NodeList=(null) usec=959 Feb 07 15:58:23 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: _job_complete: JobId=3 WTERMSIG 2 Feb 07 15:58:23 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: _job_complete: JobId=3 cancelled by interactive user Feb 07 15:58:23 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: _job_complete: JobId=3 done Feb 07 15:58:23 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: _slurm_rpc_complete_job_allocation: JobId=3 error Job/step already completing or completed Feb 07 15:58:42 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: sched: _slurm_rpc_allocate_resources JobId=4 NodeList=cluster-master-2vt2bqh7ahec04c,cluster-worker-2vt2bqh7ahec04c-2 usec=512 Feb 07 16:06:04 cluster-master-2vt2bqh7ahec04c slurmctld[51553]: slurmctld: error: _run_script: JobId=0 resumeprog exit status 1:0 Feb 07 16:09:33 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: _job_complete: JobId=4 WTERMSIG 2 Feb 07 16:09:33 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: _job_complete: JobId=4 done Feb 07 16:09:33 cluster-master-2vt2bqh7ahec04c slurmctld[51552]: slurmctld: _slurm_rpc_complete_job_allocation: JobId=4 error Job/step already completing or completed I am unsure how to debug this further. It might be coming from a previous problem I tried to fix (basically a few deprecated keys in the configuration). I will try to restart the entire cluster with the added changes to rule out any follow up errors, but maybe it's something obvious a fellow list user can see. Best regards, Xaver

2 2

sinfo gresused shard count wrong/incomplete
by Reed Dier 07 Feb '24

07 Feb '24

I have a bash script that grabs current statistics from sinfo to ship into a time series database to use for Grafana dashboards. We recently began using shards with our gpus, and I’m seeing some unexpected behavior with the data reported from sinfo. > $ sinfo -h -O "NodeHost:5 ,GresUsed:100 ,Gres:100" | grep gpu03 > gpu03 gpu:p40:0(IDX:N/A),gpu:rtx:0(IDX:N/A),shard:p40:1(IDX:1),shard:rtx:0(IDX:N/A) gpu:p40:6(S:0),gpu:rtx:2(S:0),shard:p40:36(S:0),shard:rtx:12(S:0) > > $ scontrol show node gpu03 > NodeName=gpu03 Arch=x86_64 CoresPerSocket=22 > CPUAlloc=52 CPUEfctv=88 CPUTot=88 CPULoad=43.67 > AvailableFeatures=avx,avx512,largeMemory,matlab > ActiveFeatures=avx,avx512,largeMemory,matlab > Gres=gpu:p40:6(S:0),gpu:rtx:2(S:0),shard:p40:36(S:0),shard:rtx:12(S:0) > NodeAddr=gpu03 NodeHostName=gpu03 Version=22.05.8 > OS=Linux 5.4.0-164-generic #181-Ubuntu SMP Fri Sep 1 13:41:22 UTC 2023 > RealMemory=768000 AllocMem=425984 FreeMem=116850 Sockets=2 Boards=1 > State=MIXED ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A > Partitions=gpu > BootTime=2024-02-02T14:17:07 SlurmdStartTime=2024-02-05T12:33:50 > LastBusyTime=2024-02-07T09:48:50 > CfgTRES=cpu=88,mem=750G,billing=88,gres/gpu=8,gres/gpu:p40=6,gres/gpu:rtx=2,gres/shard=48 > AllocTRES=cpu=52,mem=416G,gres/shard=21 > CapWatts=n/a > CurrentWatts=0 AveWatts=0 > ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s Specifically, we can see there are 48 total shards on this specific node, 36 of type p40, 12 of type rtx. From control, I can see that 21 shards are in use, and from nvidia-smi I can see that the breakdown of shards per gpu index is: [0]p40:4/6 [1]p40:6/6 [2]rtx:4/6 [3]rtx:3/6 [4]p40:4/6 [5]p40:0/6 [6]p40:0/6 [7]p40:0/6 So, I can tell that what “GresUsed” is reporting is actually an entire gpu’s worth of shards used, and the index of that specific entire GPU consumed. However, what it doesn’t show is the amount of specific gres/shard actually used, or more succinctly, I can’t seem to extrapolate that there are 21 shards in use. Right now my script says there is 1 shard out of 48 used, because I’m scraping for the num of shards in use, which is reporting the value of entire GPUs used by shards, and anything less than 100% is rounded to 0.. So, given that I’m on 22.05.8, maybe its better in 23.02.X which I’m hoping to move to within the next month. However I can’t seem to find anything in the release notes for 23.02 or 23.11 that would imply that sinfo reports (my) expected value of the actual count of shards used, Does anyone have any ideas for how I might be able to achieve what I’m looking for using sinfo, or should I instead try to use the sinfo json parser which has a “tres_used” field, that doesn’t appear to be accessible outside of the json output? > { > "architecture": "x86_64", > "burstbuffer_network_address": "", > "boards": 1, > "boot_time": 1706901428, > "comment": "", > "cores": 22, > "cpu_binding": 0, > [SNIP] > "tres": "cpu=88,mem=750G,billing=88,gres/gpu=8,gres/gpu:p40=6,gres/gpu:rtx=2,gres/shard=48", > "slurmd_version": "22.05.8", > "alloc_memory": 284380, > "alloc_cpus": 62, > "idle_cpus": 26, > "tres_used": "cpu=62,mem=284380M,gres/gpu=7,gres/gpu:p40=5,gres/gpu:rtx=2,gres/shard=17", > "tres_weighted": 62 > }, Any ideas appreciated. Thanks, Reed

1 0

Re: Restricting local disk storage of jobs
by Tim Schneider 07 Feb '24

07 Feb '24

Hey Jeffrey, thanks for this suggestion! This is probably the way to go if one can find a way to access GRES in the prolog. I read somewhere that people were calling scontrol to get this information, but this seems a bit unclean. Anyway, if I find some time I will try it out. Best, Tim On 2/6/24 16:30, Jeffrey T Frey wrote: > Most of my ideas have revolved around creating file systems on-the-fly > as part of the job prolog and destroying them in the epilog. The > issue with that mechanism is that formatting a file system (e.g. > mkfs.<type>) can be time-consuming. E.g. formatting your local > scratch SSD as an LVM PV+VG and allocating per-job volumes, you'd > still need to run a e.g. mkfs.xfs and mount the new file system. > > > ZFS file system creation is much quicker (basically combines the LVM + > mkfs steps above) but I don't know of any clusters using ZFS to manage > local file systems on the compute nodes :-) > > > One /could/ leverage XFS project quotas. E.g. for Slurm job 2147483647: > > > *[root@r00n00 /]# mkdir /tmp-alloc/slurm-2147483647* > *[root@r00n00 /]# xfs_quota -x -c 'project -s -p > /tmp-alloc/slurm-2147483647 2147483647' /tmp-alloc* > Setting up project 2147483647 (path /tmp-alloc/slurm-2147483647)... > Processed 1 (/etc/projects and cmdline) paths for project > 2147483647 with recursion depth infinite (-1). > *[root@r00n00 /]# xfs_quota -x -c 'limit -p bhard=1g 2147483647' > /tmp-alloc* > *[root@r00n00 /]# cd /tmp-alloc/slurm-2147483647* > *[root@r00n00 slurm-2147483647]# dd if=/dev/zero of=zeroes bs=5M > count=1000* > dd: error writing ‘zeroes’: No space left on device > 205+0 records in > 204+0 records out > 1073741824 bytes (1.1 GB) copied, 2.92232 s, 367 MB/s > > : > > [root@r00n00 /]# rm -rf /tmp-alloc/*slurm-2147483647* > [root@r00n00 /]# *xfs_quota -x -c 'limit -p bhard=0 2147483647' > /tmp-alloc* > > > Since Slurm jobids max out at 0x03FFFFFF (and 2147483647 = 0x7FFFFFFF) > we have an easy on-demand project id to use on the file system. Slurm > tmpfs plugins have to do a mkdir to create the per-job directory, > adding two xfs_quota commands (which run in more or less O(1) time) > won't extend the prolog by much. Likewise, Slurm tmpfs plugins have to > scrub the directory at job cleanup, so adding another xfs_quota > command will not do much to change their epilog execution times. The > main question is "where does the tmpfs plugin find the quota limit for > the job?" > > > > > >> On Feb 6, 2024, at 08:39, Tim Schneider via slurm-users >> <slurm-users(a)lists.schedmd.com> wrote: >> >> Hi, >> >> In our SLURM cluster, we are using the job_container/tmpfs plugin to >> ensure that each user can use /tmp and it gets cleaned up after them. >> Currently, we are mapping /tmp into the nodes RAM, which means that >> the cgroups make sure that users can only use a certain amount of >> storage inside /tmp. >> >> Now we would like to use of the node's local SSD instead of its RAM >> to hold the files in /tmp. I have seen people define local storage as >> GRES, but I am wondering how to make sure that users do not exceed >> the storage space they requested in a job. Does anyone have an idea >> how to configure local storage as a proper tracked resource? >> >> Thanks a lot in advance! >> >> Best, >> >> Tim >> >> >> -- >> slurm-users mailing list -- slurm-users(a)lists.schedmd.com >> To unsubscribe send an email to slurm-users-leave(a)lists.schedmd.com >

2 1

scheme for protected GPU jobs from preemption
by Paul Raines 06 Feb '24

06 Feb '24

After using just Fairshare for over a year on our GPU cluster, we have decided it is not working for us for what we really want to achieve among our groups. We have decided to look at preemption. What we want is for users to NOT have a #job/GPU maximum (if they are only person on the cluster they should be able to use it all), but if another user comes to the "full" cluster they should immediately be able to run some jobs. Thus preemption is needed. In our scheme we want * users to have N protected GPU jobs that cannot be preempted where N is the number of GPUs allocated. * N may not be the same for all users. Some priviledged users get more. * jobs pending in the queue will have lower priority dependent on the number of GPUs allocated to running jobs. Maybe doable somehow with PriorityWeightJobSize though not sure how. * Jobs over N are subject to preemption (and requeued if --requeue is given) with shortest running jobs of the user with most unprotected GPUs preempted first. * another complication is we have a variety of different GPUs and users may ask for specific ones which can limit what unprotected GPU jobs are available for preemption My first attempt to do this in SLURM was to just create two partitions, GPU and GPU-req, with different PriorityTier values and the later partition have PreemptMode=REQUEUE. But N would be set by a MaxTRES on the first partition and be the same for everyone and we need it to be INDEPENDENT for each user. Also users would have to "think" about which partition to submit jobs to. And users want their longest running "unprotected" job to be able to be PROMOTED to a "protected" jobs automatically when a "protected" job finishes. However slurm does not allow running jobs to move between partitions. I am trying to figure out QOS pre-emption which might solve the independent N per user issue but I don't think it will solve the promotion issue. Any ideas how this scheme might be possible in SLURM? Otherwise I might have to write a complicated cron job that tries to do it all "outside" of SLURM issuing scontrol commands. --------------------------------------------------------------- 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.

1 0

Restricting local disk storage of jobs
by Tim Schneider 06 Feb '24

06 Feb '24

Hi, In our SLURM cluster, we are using the job_container/tmpfs plugin to ensure that each user can use /tmp and it gets cleaned up after them. Currently, we are mapping /tmp into the nodes RAM, which means that the cgroups make sure that users can only use a certain amount of storage inside /tmp. Now we would like to use of the node's local SSD instead of its RAM to hold the files in /tmp. I have seen people define local storage as GRES, but I am wondering how to make sure that users do not exceed the storage space they requested in a job. Does anyone have an idea how to configure local storage as a proper tracked resource? Thanks a lot in advance! Best, Tim

3 2

Re: [ext] Restricting local disk storage of jobs
by Tim Schneider 06 Feb '24

06 Feb '24

Hi Magnus, I understand. Thanks a lot for your suggestion. Best, Tim On 06.02.24 15:34, Hagdorn, Magnus Karl Moritz wrote: > Hi Tim, > in the end the InitScript didn't contain anything useful because > > slurmd: error: _parse_next_key: Parsing error at unrecognized key: > InitScript > > At this stage I gave up. This was with SLURM 23.02. My plan was to > setup the local scratch directory with XFS and then get the script to > apply a project quota, ie quota attached to the directory. > > I would start by checking if slurm recognises the InitScript option. > > Regards > magnus > > On Tue, 2024-02-06 at 15:24 +0100, Tim Schneider wrote: >> Hi Magnus, >> >> thanks for your reply! If you can, would you mind sharing the >> InitScript >> of your attempt at getting it to work? >> >> Best, >> >> Tim >> >> On 06.02.24 15:19, Hagdorn, Magnus Karl Moritz wrote: >>> Hi Tim, >>> we are using the container/tmpfs plugin to map /tmp to a local NVMe >>> drive which works great. I did consider setting up directory >>> quotas. I >>> thought the InitScript [1] option should do the trick. Alas, I >>> didn't >>> get it to work. If I remember correctly, slurm complained about the >>> option being present. In the end we recommend our users to make >>> exclusive use a node if they are going to use a lot of local >>> scratch >>> space. I don't think this happens very often if at all. >>> Regards >>> magnus >>> >>> [1] >>> https://slurm.schedmd.com/job_container.conf.html#OPT_InitScript >>> >>> >>> On Tue, 2024-02-06 at 14:39 +0100, Tim Schneider via slurm-users >>> wrote: >>>> Hi, >>>> >>>> In our SLURM cluster, we are using the job_container/tmpfs plugin >>>> to >>>> ensure that each user can use /tmp and it gets cleaned up after >>>> them. >>>> Currently, we are mapping /tmp into the nodes RAM, which means >>>> that >>>> the >>>> cgroups make sure that users can only use a certain amount of >>>> storage >>>> inside /tmp. >>>> >>>> Now we would like to use of the node's local SSD instead of its >>>> RAM >>>> to >>>> hold the files in /tmp. I have seen people define local storage >>>> as >>>> GRES, >>>> but I am wondering how to make sure that users do not exceed the >>>> storage >>>> space they requested in a job. Does anyone have an idea how to >>>> configure >>>> local storage as a proper tracked resource? >>>> >>>> Thanks a lot in advance! >>>> >>>> Best, >>>> >>>> Tim >>>> >>>>

1 0

Re: [ext] Restricting local disk storage of jobs
by Tim Schneider 06 Feb '24

06 Feb '24

Hi Magnus, thanks for your reply! If you can, would you mind sharing the InitScript of your attempt at getting it to work? Best, Tim On 06.02.24 15:19, Hagdorn, Magnus Karl Moritz wrote: > Hi Tim, > we are using the container/tmpfs plugin to map /tmp to a local NVMe > drive which works great. I did consider setting up directory quotas. I > thought the InitScript [1] option should do the trick. Alas, I didn't > get it to work. If I remember correctly, slurm complained about the > option being present. In the end we recommend our users to make > exclusive use a node if they are going to use a lot of local scratch > space. I don't think this happens very often if at all. > Regards > magnus > > [1] > https://slurm.schedmd.com/job_container.conf.html#OPT_InitScript > > > On Tue, 2024-02-06 at 14:39 +0100, Tim Schneider via slurm-users wrote: >> Hi, >> >> In our SLURM cluster, we are using the job_container/tmpfs plugin to >> ensure that each user can use /tmp and it gets cleaned up after them. >> Currently, we are mapping /tmp into the nodes RAM, which means that >> the >> cgroups make sure that users can only use a certain amount of storage >> inside /tmp. >> >> Now we would like to use of the node's local SSD instead of its RAM >> to >> hold the files in /tmp. I have seen people define local storage as >> GRES, >> but I am wondering how to make sure that users do not exceed the >> storage >> space they requested in a job. Does anyone have an idea how to >> configure >> local storage as a proper tracked resource? >> >> Thanks a lot in advance! >> >> Best, >> >> Tim >> >>

2 1

Starting a job after a file is created in previous job (dependency looking for soluton)
by Amjad Syed 06 Feb '24

06 Feb '24

Hello I have the following scenario: I need to submit a sequence of up to 400 jobs where the even jobs depend on the preceeding odd job to finish and every odd job depends on the presence of a file generated by the preceding even job (availability of the file for the first of those 400 jobs is guaranteed). If I just submit all those jobs via a loop using dependencies, then I end up with a lot of pending jobs who might later not even run because no output file has been produced by the preceding jobs. Is there a way to pause the submission loop until the required file has been generated so that at most two jobs are submitted at the same time? Here is a sample submission script showing what I want to achieve. for i in {1..200}; do FILE=GHM_paramset_${i}.dat # How can I pause the submission loop until the FILE has been created???? #if test -f "$FILE"; then jobid4=$(sbatch --parsable --dependency=afterok:$jobid3 job4_sub $i) jobid3=$(sbatch --parsable --dependency=afterok:$jobid4 job3_sub $i) #fi done Any help will be appreciated Amjad

3 2

SLURM configuration for LDAP users
by Richard Chang 05 Feb '24

05 Feb '24

Hi, I am a little new to this, so please pardon my ignorance. I have configured slurm in my cluster and it works fine with local users. But I am not able to get it working with LDAP/SSSD authentication. User logins using ssh are working fine. An LDAP user can login to the login, slurmctld and compute nodes, but when they try to submit jobs, slurmctld logs an error about invalid account or partition for user. Someone said we need to add the user manually into the database using the sacctmgr command. But I am not sure we need to do this for each and every LDAP user. Yes, it does work if we add the LDAP user manually using sacctmgr. But I am not convinced this manual way is the way to do. The documentation is not very clear about using LDAP accounts. Saw somewhere in the list about using UsePAM=1 and copying or creating a softlink for slurm PAM module under /etc/pam.d . But it didn't work for me. Saw somewhere else that we need to specifying LaunchParameters=enable_nss_slurm in the slurm.conf file and put slurm keyword in passwd/group entry in the /etc/nsswitch.conf file. Did these, but didn't help either. I am bereft of ideas at present. If anyone has real world experience and can advise, I will be grateful. Thank you, Richard

4 4

timeout of get_user_env does not obey time limit
by urbanjost 03 Feb '24

03 Feb '24

If I use the sbatch(1) option --export=NONE or wipe the environment with "env -i /usr/bin/sbatch ..." or use --export=NIL then the environment is not properly constructed and I see the message in the /var/log/*slurm* files: [2024-02-03T11:50:33.052] _get_user_env: get env for user jsu here [2024-02-03T11:52:33.152] timeout waiting for /bin/su to complete [2024-02-03T11:52:34.152] error: Failed to load current user environment variables [2024-02-03T11:52:34.153] error: _get_user_env: Unable to get user's local environment, running only with passed environment This occurs at 120 seconds not matter if I add --get-user-env=3600 or adjust many slurm.conf time-related parameters. It is easy to reproduce by adding "sleep 100" into a .cshrc file and sbatch(1) the file #!/bin/csh #SBATCH --export=NONE --propagate=NONE --get-user-env=3600L printenv HOME printenv USER printenv PATH env I have adjust MANY time-related limits in the slurm.conf file to no avail. When the system is unresponsive or heavily loaded or users have prologues that set up complex environments via module commands (which can be notoriously slow) the jobs are failing or producing errors. If I configure Slurm so that jobs that timeout requeue instead of running then a user with a slow login setup can submit a large number of jobs and basically close down a cluster because this option not only requeues jobs that fail but puts the node it occurred on in a DRAIN state. We see this as very dangerous as by defaults jobs proceed to execute even when their environment is not properly constructed. I can see that "slurmrestd getenv" and the procedure get_user_env(3c) are involved, but a preliminary scan of the code looked like the --get-user-env=NNNN value was being parsed, and I did not see a reason the setup always times out at 120 seconds (at least on my system). Does anyone know how to get the time allowed to get the default user environment to use the value on the --get-user-env option when no environment is being exported to a job? This is showing up sporadically and causing intermittent failures that are very confusing and disturbing to the users it occurs with. Sent with [Proton Mail](https://proton.me/) secure email.

1 0

selecting job-specific log messages
by urbanjost 03 Feb '24

03 Feb '24

log files use many strings to identify job, including not having a jobID in the message NUMBER=$SLURM_JOBID egrep "\.\<$NUMBER\>\] |\<$NUMBER\>\.batch|jobid \<$NUMBER\>|JObId=\<$NUMBER\>|job id \<$NUMBER\>|job\.\<$NUMBER\>|job \<$NUMBER\>|jobid \[\<$NUMBER\>\]|task_p_slurmd_batch_request: \<$NUMBER\>" /var/log/slurm* Even that misses cruciall data that does not even contain the jobid [2024-02-03T11:50:33.052] _get_user_env: get env for user jsu here [2024-02-03T11:52:33.152] timeout waiting for /bin/su to complete [2024-02-03T11:52:34.152] error: Failed to load current user environment variables [2024-02-03T11:52:34.153] error: _get_user_env: Unable to get user's local environment, running only with passed environment It would be very useful if all messages related to a job had a consistent string in them for grepping the log files; even better might be a command like "scontrol show jobid=NNNN log_messages But I could not find what I wanted (an easy way to find all daemon log messages related to a specific job). I would find it particularly useful if there were a way to automatically append such information to the stdout of the job at job termination so users would automatically get information about job failures or warnings. Is there such a feature available I have missed? Sent with [Proton Mail](https://proton.me/) secure email.

1 0

after upgrade to 23.11.1 nodes stuck in completion state
by Paul Raines 01 Feb '24

01 Feb '24

I finally had downtime on our cluster running 20.11.3 and decided to upgrade SLURM. All daemons were stopped on nodes and master. Rocky 8 Linux OS was updated but not changed configuration-wise in anyway. On the master, when I first installed 23.11.1 and tried to run slurmdbd -D -vvv at the command line, it balked as the more than 2 major version jump. I then installed 22.05.6 and run slurmdbd -D -vvv which took a long time but completed fine. I reinstalled 23.11.1 and did slurmdbd -D --vvv which went quick. I stopped it then ran slurmdbd normally with systemctl. When I ran slurmctld it complained about incompatable state files. I didn't care about loosing state so ran with the -i option to ignore. Seemed happy. I then started up slurmd on my various GPU and non-GPU nodes. I ran a test interactive job just to see if I would get a shell with expected SLURM environment on the nodes and it worked. Except on the GPU nodes I got errors about not being able to get/set GPU frequences. I removed GpuFreqDef=medium from slurm.conf and that went away. I had my people start running jobs last night. This morning I come in and see a lot of nodes are stuck in completing state. All these are GPU nodes. For example on node 'sevilla' squeue reports JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 3679433 lcna100 sjob_123 sg1526 CG 8:29 1 sevilla 3679432 lcna100 sjob_123 sg1526 CG 8:35 1 sevilla 3679431 lcna100 sjob_123 sg1526 CG 8:42 1 sevilla In the slurmd.log on sevilla I see every minute the lines: [2024-01-28T09:37:52.003] debug2: Start processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug2: Processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug: _rpc_terminate_job: uid = 1150 JobId=3679432 [2024-01-28T09:37:52.004] debug2: Start processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug2: Processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug: _rpc_terminate_job: uid = 1150 JobId=3679431 [2024-01-28T09:37:52.004] debug2: Finish processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug2: Start processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug2: Processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug: _rpc_terminate_job: uid = 1150 JobId=3679433 [2024-01-28T09:37:52.004] debug2: Finish processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T09:37:52.004] debug2: Finish processing RPC: REQUEST_TERMINATE_JOB Back at the time 3679433 was cancelled by the user I see [2024-01-28T00:27:48.658] debug2: Start processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T00:27:48.658] debug2: Processing RPC: REQUEST_TERMINATE_JOB [2024-01-28T00:27:48.658] debug: _rpc_terminate_job: uid = 1150 JobId=3679433 [2024-01-28T00:27:48.658] debug: credential for job 3679433 revoked [2024-01-28T00:27:48.658] debug: _rpc_terminate_job: sent SUCCESS for 3679433, waiting for prolog to finish [2024-01-28T00:27:48.658] debug: Waiting for job 3679433's prolog to complete On sevilla itself there are no processes currently running for user sg1526 so no sign of these jobs. There ARE processes running for user mu40 though [root@sevilla ~]# pgrep -f slurm_script | xargs -n1 pstree -p | grep ^slurm slurm_script(30691)---starter-suid(30764)-+-python3(30784)-+-{python3}(30817) slurm_script(30861)---starter-suid(30873)-+-python3(30888)-+-{python3}(30915) slurm_script(30963)---starter-suid(30975)-+-python3(30992)-+-{python3}(31020) [root@sevilla ~]# strings /proc/30691/cmdline /bin/sh /var/slurm/spool/d/job3679220/slurm_script /usr/bin/sacct -p --jobs=3679220 --format=JobID,User,ReqTRES,NodeList,Start,End,Elapsed,CPUTime,State,ExitCode JobID|User|ReqTRES|NodeList|Start|End|Elapsed|CPUTime|State|ExitCode| 3679220|mu40|billing=6,cpu=3,gres/gpu=1,mem=32G,node=1|sevilla|2024-01-27T22:33:27|2024-01-27T23:26:39|00:53:12|02:39:36|NODE_FAIL|1:0| 3679220.batch|||sevilla|2024-01-27T22:33:27|2024-01-27T23:26:39|00:53:12|02:39:36|CANCELLED|| 3679220.extern|||sevilla|2024-01-27T22:33:27|2024-01-27T23:26:39|00:53:12|02:39:36|CANCELLED|| The slurmd.log has at the time it started lines like: [2024-01-27T22:33:28.679] [3679220.batch] debug2: gpu/nvml: _get_gpuutil: Couldn't find pid 30691, probably hasn't started yet or has already finished [2024-01-27T22:33:58.092] [3679220.extern] debug2: gpu/nvml: _get_gpuutil: Couldn't find pid 30650, probably hasn't started yet or has already finished [2024-01-27T22:33:58.166] [3679220.batch] debug2: profile signaling type Task [2024-01-27T22:33:58.168] [3679220.batch] debug2: gpu/nvml: _get_gpuutil: Couldn't find pid 30691, probably hasn't started yet or has already finished [2024-01-27T22:33:58.170] [3679220.batch] debug2: gpu/nvml: _get_gpuutil: Couldn't find pid 30764, probably hasn't started yet or has already finished [2024-01-27T22:33:58.172] [3679220.batch] debug2: gpu/nvml: _get_gpuutil: Couldn't find pid 30784, probably hasn't started yet or has already finished ... [2024-01-27T22:46:39.018] debug: _step_connect: connect() failed for /var/slurm/spool/d/sevilla_3679220.4294967292: Connection refused [2024-01-27T22:46:39.018] debug: Cleaned up stray socket /var/slurm/spool/d/sevilla_3679220.4294967292 [2024-01-27T22:46:39.018] debug: _step_connect: connect() failed for /var/slurm/spool/d/sevilla_3679220.4294967291: Connection refused [2024-01-27T22:46:39.018] debug: Cleaned up stray socket /var/slurm/spool/d/sevilla_3679220.4294967291 [2024-01-27T22:46:39.018] _handle_stray_script: Purging vestigial job script /var/slurm/spool/d/job3679220/slurm_script then later at the time of termination [2024-01-27T23:26:39.042] debug: _rpc_terminate_job: uid = 1150 JobId=3679220 [2024-01-27T23:26:39.042] debug: credential for job 3679220 revoked [2024-01-27T23:26:39.042] debug2: No steps in jobid 3679220 to send signal 998 [2024-01-27T23:26:39.043] debug2: No steps in jobid 3679220 to send signal 18 [2024-01-27T23:26:39.043] debug2: No steps in jobid 3679220 to send signal 15 [2024-01-27T23:26:39.044] debug2: set revoke expiration for jobid 3679220 to 1706416119 UTS [2024-01-27T23:26:39.044] debug: Waiting for job 3679220's prolog to complete [2024-01-27T23:26:39.044] debug: Finished wait for job 3679220's prolog to complete [2024-01-27T23:26:39.068] debug: completed epilog for jobid 3679220 [2024-01-27T23:26:39.071] debug: JobId=3679220: sent epilog complete msg: rc = 0 So I have a node stuck in 'comp' state that claims to have jobs completing that are definite NOT running on the box BUT there are jobs running on the box that SLURM thinks are done --------------------------------------------------------------- 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 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.

4 7

Why is Slurm 20 the latest RPM in RHEL 8/Fedora repo?
by Robert Kudyba 31 Jan '24

31 Jan '24

According to these links: https://rpmfind.net/linux/rpm2html/search.php?query=slurm https://src.fedoraproject.org/rpms/slurm Why doesn't RHEL 8 get a newer version? Can someone update the repo maintainer Philip Kovacs <https://src.fedoraproject.org/user/pkfed> < pkfed(a)fedoraproject.org>? There was a ticket at https://bugzilla.redhat.com/show_bug.cgi?id=1912491 but no movement on RHEL 8.

5 4

Mailing list upgrade - slurm-users list paused
by Tim Wickberg 30 Jan '24

30 Jan '24

Hey folks - The mailing list will be offline for about an hour as we upgrade the host, upgrade the mailing list software, and change the mail configuration around. As part of these changes, the "From: " field will no longer be the original sender, but instead use the mailing list ID itself. This is to comply with DMARC sending options, and allow us to start DKIM signing messages to ensure deliverability once Google and Yahoo impose new policy changes in February. This is the last post on the current (mailman2) list. I'll send a welcome message on the upgraded (mailman3) list once finished, and when the list is open to new traffic again. - Tim -- Tim Wickberg Chief Technology Officer, SchedMD LLC Commercial Slurm Development and Support

1 1

Testing to SchedMD folks
by Tim Wickberg 30 Jan '24

30 Jan '24

Hey - this is hopefully a final test before the list is re-enabled for everyone. - Tim

1 0

Socket timed out - tuning
by Reed Dier 29 Jan '24

29 Jan '24

Hoping someone can help point me towards some tweaks to help prevent denial-of-service issues. > sbatch: error: Batch job submission failed: Socket timed out on send/recv operation Root cause is understood, issues with shared storage for the slurmctld’s was impacted, leading to an increase in write latency to the StateSaveLocation. Then with a large enough avalanche of job submissions, things the RPC’s would stack up and stop responding. I’ve been running well with some tweaks sourced from the “high-throughput” guide <https://slurm.schedmd.com/high_throughput.html>. > SchedulerParameters=max_rpc_cnt=400,\ > sched_min_interval=50000,\ > sched_max_job_start=300,\ > batch_sched_delay=6 > KillWait=30 > MessageTimeout=30 I’m assuming that I was running into batch_sched_delay because looking at sdiag after the fact, it was averaging .2s, and total time is 5.5h out of 16h8m18s at the time of the sdiag sample. > ******************************************************* > sdiag output at Thu Jan 25 11:08:18 2024 (1706198898) > Data since Wed Jan 24 19:00:00 2024 (1706140800) > ******************************************************* > REQUEST_SUBMIT_BATCH_JOB ( 4003) count:98400 ave_time:201442 total_time:19821991013 Currently on 22.05.8, but hoping to get to 23.02.7 soon™, and I think this could possible resolve the issue well enough if I’m reading it correctly from the release notes <https://slurm.schedmd.com/archive/slurm-23.02-latest/news.html>? > HIGHLIGHTS > ========== > -- slurmctld - Add new RPC rate limiting feature. This is enabled through > SlurmctldParameters=rl_enable, otherwise disabled by default. > rl_enable <https://slurm.schedmd.com/slurm.conf.html#OPT_rl_enable>Enable per-user RPC rate-limiting support. Client-commands will be told to back off and sleep for a second once the limit has been reached. This is implemented as a "token bucket", which permits a certain degree of "bursty" RPC load from an individual user before holding them to a steady-state RPC load established by the refill period and rate. But given that the hardware seems to be well over provisioned, CPU never drops below 5% idle, it feels like there is more room to squeeze some optimization out of here that I’m missing in the interim, and hoping to get a better overall understanding in the process. I scrape the DBD Agent queue size from sdiag every 30s and the largest value I saw was 115, which is much higher than normal, but should be well below MaxDBDMsgs, where the minimum value is 10000. I would really hope that I didn’t potentially hit a 30s MessageTimeout value, but I guess thats on the table all well, not knowing if that would potentially trigger the sbatch submission failure like that. Just moving the max_rpc_cnt value up seems like an easy button, but also seems like it could have some adverse effects for backfill scheduling, and may be diminishing returns for actually keeping RPCs flowing? > Setting max_rpc_cnt to more than 256 will be only useful to let backfill continue scheduling work after locks have been yielded (i.e. each 2 seconds) if there are a maximum of MAX(max_rpc_cnt/10, 20) RPCs in the queue. i.e. max_rpc_cnt=1000, the scheduler will be allowed to continue after yielding locks only when there are less than or equal to 100 pending RPCs. Obviously, fix the storage is the real solution, but hoping that there may be more goodness to unlock, even if it is as simple as “upgrade to 23.02”. Appreciate any insight, Reed

1 0

Two jobs each with a different partition running on same node?
by Loris Bennett 29 Jan '24

29 Jan '24

Hi, I seem to remember that in the past, if a node was configured to be in two partitions, the actual partition of the node was determined by the partition associated with the jobs running on it. Moreover, at any instance where the node was running one or more jobs, the node could only actually be in a single partition. Was this indeed the case and is it still the case with version Slurm 23.02.7? Cheers, Loris -- Dr. Loris Bennett (Herr/Mr) FUB-IT (ex-ZEDAT), Freie Universität Berlin

2 1

srun only runs one job on a node
by Rike-Benjamin Schuppner 27 Jan '24

27 Jan '24

Hi, Not sure if I do not understand the scheduling algorithm correctly of if this is a bug. The bug seems similar to the one described and fixed here https://github.com/SchedMD/slurm/commit/6a2c99edbf96e50463cc2f16d8e5eb955c8… Given the node definition: NodeName=cluster01 CPUs=8 Boards=1 SocketsPerBoard=1 CoresPerSocket=4 ThreadsPerCore=2 RealMemory=64283 NodeName=cluster02 CPUs=8 Boards=1 SocketsPerBoard=1 CoresPerSocket=4 ThreadsPerCore=2 RealMemory=128814 PartitionName=DEFAULT MaxTime=2-0 MaxNodes=2 PartitionName=mini Nodes=cluster01,cluster02 State=UP and the batch script: for i in {1..6} do srun --exclusive --mem=100 --threads-per-core=1 -N1 -c1 -n1 bash -c 'echo $(hostname) $(date); sleep 10 ; echo done $(hostname) $(date)' & done wait When run with: time sbatch --wait --partition=mini --ntasks=10 --cpus-per-task=1 --mem=1000 mini.slurm I get the following log after 2m 18s: cluster01 Sat Jan 27 02:05:12 PM CET 2024 cluster01 Sat Jan 27 02:05:12 PM CET 2024 cluster01 Sat Jan 27 02:05:13 PM CET 2024 cluster01 Sat Jan 27 02:05:13 PM CET 2024 cluster02 Sat Jan 27 02:05:13 PM CET 2024 done cluster01 Sat Jan 27 02:05:22 PM CET 2024 done cluster01 Sat Jan 27 02:05:22 PM CET 2024 done cluster01 Sat Jan 27 02:05:23 PM CET 2024 done cluster01 Sat Jan 27 02:05:23 PM CET 2024 done cluster02 Sat Jan 27 02:05:23 PM CET 2024 srun: Job 650 step creation temporarily disabled, retrying (Requested nodes are busy) srun: Step created for StepId=650.6 cluster01 Sat Jan 27 02:07:14 PM CET 2024 done cluster01 Sat Jan 27 02:07:24 PM CET 2024 So apparently I am getting 4 sruns on cluster01 (why not 6 when I have 8 CPUs?), but only 1 is scheduled on cluster02. Then it is exhausted and waits endlessly long (almost 2 minutes?) before scheduling again. The step log tells me: slurmctld: debug: laying out the 1 tasks on 1 hosts cluster01 dist 1 slurmctld: STEPS: _step_alloc_lps: JobId=650 StepId=4 node 0 (cluster01) gres_cpus_alloc=0 tasks=1 cpus_per_task=1 slurmctld: STEPS: _pick_step_cores: step JobId=650 StepId=4 requires 1 cores on node 0 with cpus_per_core=1, available cpus from job: 8 slurmctld: STEPS: _pick_step_core: alloc Node:0 Socket:0 Core:3 slurmctld: STEPS: step alloc on job node 0 (cluster01) used 4 of 8 CPUs slurmctld: STEPS: _slurm_rpc_job_step_create: JobId=650 StepId=4 cluster01 usec=5502 slurmctld: STEPS: _pick_step_nodes: JobId=650 Currently running steps use 4 of allocated 8 CPUs on node cluster01 slurmctld: STEPS: _pick_step_nodes: JobId=650 Currently running steps use 1 of allocated 2 CPUs on node cluster02 slurmctld: STEPS: _mark_busy_nodes: JobId=650 StepId=0 has nodes cluster01 slurmctld: STEPS: _mark_busy_nodes: JobId=650 StepId=1 has nodes cluster02 slurmctld: STEPS: _mark_busy_nodes: JobId=650 StepId=2 has nodes cluster01 slurmctld: STEPS: _mark_busy_nodes: JobId=650 StepId=3 has nodes cluster01 slurmctld: STEPS: _mark_busy_nodes: JobId=650 StepId=4 has nodes cluster01 slurmctld: STEPS: _pick_step_nodes: step pick 1-1 nodes, avail:cluster[01-02] idle: picked:NONE slurmctld: STEPS: _pick_step_nodes: step picked 0 of 1 nodes slurmctld: STEPS: Picked nodes cluster01 when accumulating from cluster01 slurmctld: debug: laying out the 1 tasks on 1 hosts cluster01 dist 1 slurmctld: STEPS: _step_alloc_lps: JobId=650 StepId=5 node 0 (cluster01) gres_cpus_alloc=0 tasks=1 cpus_per_task=1 slurmctld: STEPS: _pick_step_cores: step JobId=650 StepId=5 requires 1 cores on node 0 with cpus_per_core=1, available cpus from job: 8 slurmctld: STEPS: unable to pick step cores for job node 0 (cluster01): Requested nodes are busy slurmctld: STEPS: Deallocating 100MB of memory on node 0 (cluster01) now used: 400 of 1000 slurmctld: STEPS: step dealloc on job node 0 (cluster01) used: 4 of 8 CPUs slurmctld: STEPS: _slurm_rpc_job_step_create for JobId=650: Requested nodes are busy So why is it unable to allocate on cluster02? And why does its need to when it is only using 4 of 8 CPUs on cluster01 anyway? Now if I change to --ntasks=8, the algorithm doesn’t even schedule cluster02 at all: cluster01 Sat Jan 27 02:15:30 PM CET 2024 cluster01 Sat Jan 27 02:15:30 PM CET 2024 cluster01 Sat Jan 27 02:15:30 PM CET 2024 cluster01 Sat Jan 27 02:15:30 PM CET 2024 done cluster01 Sat Jan 27 02:15:40 PM CET 2024 done cluster01 Sat Jan 27 02:15:40 PM CET 2024 done cluster01 Sat Jan 27 02:15:40 PM CET 2024 done cluster01 Sat Jan 27 02:15:40 PM CET 2024 srun: Job 651 step creation temporarily disabled, retrying (Requested nodes are busy) srun: Step created for StepId=651.6 cluster01 Sat Jan 27 02:17:35 PM CET 2024 srun: Job 651 step creation temporarily disabled, retrying (Requested nodes are busy) srun: Step created for StepId=651.7 cluster01 Sat Jan 27 02:17:36 PM CET 2024 done cluster01 Sat Jan 27 02:17:45 PM CET 2024 done cluster01 Sat Jan 27 02:17:46 PM CET 2024 Thanks for your help. Best /rike

1 0

LuaSQLite3 with slurm-lua-spank
by Jackson, Gary L. 26 Jan '24

26 Jan '24

What do I need to do to use the LuaSQLite3 library with Lua SPANK? For what it’s worth, I don’t know anything about Lua. -- Gary

1 0

sinfo: error: resolve_ctls_from_dns_srv: res_nsearch error: Unknown host
by Michael Lewis 26 Jan '24

26 Jan '24

Hi All, I’m trying to get slurm-23.11.3 running on Ubuntu 20.04 and running on a stand alone system. I’m running into an issue I can not find the answer to. After compiling and installing when I fire up slurmctld and slurmd I get an error from sinfo: sinfo: error: resolve_ctls_from_dns_srv: res_nsearch error: Unknown host sinfo: error: fetch_config: DNS SRV lookup failed sinfo: error: _establish_config_source: failed to fetch config sinfo: fatal: Could not establish a configuration source I looks like a DNS issue but the system has no issue resolving to its hostname or localhost. The slurm.conf file is also being read properly as I have the logs directed to a place convenient to me. I see lots have had these same issues but cannot find a clear resolution. I have slurm running on a stand alone system in another lab with and identical setup without issue. Any advice would be greatly appreciated. Thanks, Mike

2 2

Database cluster
by Daniel L'Hommedieu 26 Jan '24

26 Jan '24

Community: What do you do to ensure database reliability in your SLURM environment? We can have multiple controllers and multiple slurmdbds, but my understanding is that slurmdbd can be configured with a single MySQL server, so what do you do? Do you have that “single MySQL server” be a cluster, such as Percona XtraDB? Do you use MySQL replication, then manually switch to slurmdbd to a replication slave if the master goes down? Do you do something else? Thanks. Daniel

6 7

Problem using Podman with scrun on SLURM 23.11.3
by Marcus Lauer 25 Jan '24

25 Jan '24

I am getting an unusual error when trying to run Podman containers using scrun on SLURM 23.11.3 (and 23.11.1 previously). In short, Podman works when not configured to use scrun, but when configured to use scrun it fails. Podman gives this error: scrun: fatal: Unable to request job allocation: Job cannot be submitted without the current working directory specified. The slurmctld logs show this error: _slurm_rpc_allocate_resources: Job cannot be submitted without the current working directory specified. The scrun command does not appear to have a --chdir option so the working directory must be detected automatically, and this appears to be failing. Perhaps someone has encountered this or a similar error before? If anyone has seen this before I would love to hear about your experiences. A bit more detail in case anyone is interested. This is a two-node test system with one head node and one compute node, both running Rocky Linux 8.9. I have created /etc/containers/storage.conf and containers.conf files which are identical to the ones in the SLURM Containers Guide except that I set the rootless_storage_path to the same value as the graphroot in storage.conf. When I get rid of the /etc/containers/containers.conf file I can run containers on the head node. The backing store is ext4, though I have tried NFS (which had other problems). I have tried the vfs and overlay storage drivers but get the same results. On the SLURM side I created an oci.conf which is the same as the "oci.conf example for crun using run (suggested)" from the SLURM Container Guide. I'm using crun because I have forced cgroup v2 on both systems ( systemd.unified_cgroup_hierarchy=1 on the kernel cmdline) and crun seems to support cgroup v2 better. I am using an scrun.lua which is similar to the one in the scrun manual page with some paths modified for my setup. I should also mention that sbatch jobs run just fine and srun works as well. the test cluster seems to be working fine in general. -- Marcus Lauer Systems Administrator CETS Group, Research Support

1 0

Question about CPUs and cores
by Gestió Servidors 25 Jan '24

25 Jan '24

Hi, I want to run a simple test that uses one node and four cores. Also, in my script, I execute a binary that reports me in what core is running one of the four tasks. These are my files: * submit script: #!/bin/bash #SBATCH --job-name=test_jobs # Job name #SBATCH --output=test_job_%j.out # Output file (using array index) #SBATCH --error=test_job_%j.err # Error file (using array index) #SBATCH --ntasks=4 #SBATCH --ntasks-per-core=1 #SBATCH --cpus-per-task=1 # Number of CPU cores per task echo "START : $(date)" srun -n1 -c1 check_cores >> output_small.txt & srun -n1 -c1 check_cores >> output_small.txt & srun -n1 -c1 check_cores >> output_small.txt & srun -n1 -c1 check_cores >> output_small.txt & # Wait for all jobs to finish before exiting the job submission script wait echo "END : $(date)" * check_cores.c file: #include <stdio.h> #include<unistd.h> struct pstat { int pid; //process id char tcomm[256];//filename of the executable char state[2]; //state (R is running, S is sleeping, D is sleeping in an //uninterruptible wait, Z is zombie, T is traced or stopped) int ppid;// process id of the parent process int pgrp;// pgrp of the process int sid;// session id int tty_nr;// tty the process uses int tty_pgrp;// pgrp of the tty int flags;// task flags int min_flt;// number of minor faults int cmin_flt;// number of minor faults with child's int maj_flt;// number of major faults int cmaj_flt;// number of major faults with child's int utime;// user mode jiffies int stime;// kernel mode jiffies int cutime;// user mode jiffies with child's int cstime;// kernel mode jiffies with child's int priority;// priority level int nice;// nice level int num_threads;// number of threads int it_real_value;// (obsolete, always 0) int start_time;// time the process started after system boot int vsize;// virtual memory size int rss;// resident set memory size int rsslim;// current limit in bytes on the rss int start_code;// address above which program text can run int end_code;// address below which program text can run int start_stack;// address of the start of the stack int esp;// current value of ESP int eip;// current value of EIP int pending;// bitmap of pending signals int blocked;// bitmap of blocked signals int sigign;// bitmap of ignored signals int sigcatch;// bitmap of catched signals int wchan;// address where process went to sleep int i0;// (place holder) int i1;// (place holder) int exit_signal;// signal to send to parent thread on exit int task_cpu;// which CPU the task is scheduled on int rt_priority;// realtime priority int policy;// scheduling policy (man sched_setscheduler) int blkio_ticks;// time spent waiting for block IO int gtime;// guest time of the task in jiffies int cgtime;// guest time of the task children in jiffies } p ; int main() { char name[256]; char state[8]; FILE* f = fopen("/proc/self/stat", "r"); char hostname[1024]; gethostname(hostname, 1024); fscanf(f, "%d%s%s%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d%d", &p.pid, &p.tcomm, &p.state, &p.ppid, &p.pgrp, &p.sid, &p.tty_nr, &p.tty_pgrp, &p.flags, &p.min_flt, &p.cmin_flt, &p.maj_flt, &p.cmaj_flt, &p.utime, &p.stime, &p.cutime, &p.cstime, &p.priority, &p.nice, &p.num_threads, &p.it_real_value, &p.start_time, &p.vsize, &p.rss, &p.rsslim, &p.start_code, &p.end_code, &p.start_stack, &p.esp, &p.eip, &p.pending, &p.blocked, &p.sigign, &p.sigcatch, &p.wchan, &p.i0, &p.i1, &p.exit_signal, &p.task_cpu, &p.rt_priority, &p.policy, &p.blkio_ticks, &p.gtime, &p.cgtime); sleep (15); printf("HOST %s\n", hostname); printf("CPU %d\n", p.task_cpu); return 0; } When I run this script, I should receive an output like this: HOST clus04.hpc.local CPU 0 HOST clus04.hpc.local CPU 7 HOST clus04.hpc.local CPU 1 HOST clus04.hpc.local CPU 6 because system has scheduled my job in node "clus04.hpc.local" (a node with two 6-cores processors) and has reserved cores 0-1 of each processor. However, the real output I receive is HOST clus04.hpc.local CPU 1 HOST clus04.hpc.local CPU 1 HOST clus04.hpc.local CPU 1 HOST clus04.hpc.local CPU 6 (and others outputs where there are one CPU repeated), so it seems SLURMs is reserving correctly four cores but, then, process is running two or more times in the same core. Slurmctld.log file in server shows this information: [...] [2024-01-25T09:38:42.867] DEBUG: JobId=206129 node clus04 vpus 1 cpus 4 [2024-01-25T09:38:42.867] ==================== [2024-01-25T09:38:42.867] JobId=206129 nhosts:1 ncpus:4 node_req:1 nodes=clus04 [2024-01-25T09:38:42.867] Node[0]: [2024-01-25T09:38:42.867] Mem(MB):0:0 Sockets:2 Cores:6 CPUs:4:0 [2024-01-25T09:38:42.867] Socket[0] Core[0] is allocated [2024-01-25T09:38:42.867] Socket[0] Core[1] is allocated [2024-01-25T09:38:42.867] Socket[1] Core[0] is allocated [2024-01-25T09:38:42.867] Socket[1] Core[1] is allocated [2024-01-25T09:38:42.867] -------------------- [2024-01-25T09:38:42.867] cpu_array_value[0]:4 reps:1 [2024-01-25T09:38:42.867] ==================== [2024-01-25T09:38:42.867] DEBUG: Dump job_resources: nhosts 1 cb 0-1,6-7 [...] [2024-01-25T09:38:58.213] _job_complete: JobId=206129 WEXITSTATUS 0 [2024-01-25T09:38:58.213] deallocate_nodes: JobId=206129 [2024-01-25T09:38:58.213] DEBUG: Dump job_resources: nhosts 1 cb 0-1,6-7 [2024-01-25T09:38:58.213] _job_complete: JobId=206129 done And slurmd.log file in node clus04 shows this information: [...] [2024-01-25T09:38:30.685] [206129.batch] debug: task/cgroup: job abstract cores are '0-1,6-7' [2024-01-25T09:38:30.685] [206129.batch] debug: task/cgroup: step abstract cores are '0-1,6-7' [2024-01-25T09:38:30.685] [206129.batch] debug: task/cgroup: job physical cores are '0-1,6-7' [2024-01-25T09:38:30.685] [206129.batch] debug: task/cgroup: step physical cores are '0-1,6-7' [...] [2024-01-25T09:38:30.883] [206129.1] debug: task/cgroup: job abstract cores are '0-1,6-7' [2024-01-25T09:38:30.883] [206129.1] debug: task/cgroup: step abstract cores are '0-1,6-7' [2024-01-25T09:38:30.883] [206129.1] debug: task/cgroup: job physical cores are '0-1,6-7' [2024-01-25T09:38:30.883] [206129.1] debug: task/cgroup: step physical cores are '0-1,6-7' [...] [2024-01-25T09:38:30.911] [206129.0] debug: task/cgroup: job abstract cores are '0-1,6-7' [2024-01-25T09:38:30.911] [206129.0] debug: task/cgroup: step abstract cores are '0-1,6-7' [2024-01-25T09:38:30.911] [206129.0] debug: task/cgroup: job physical cores are '0-1,6-7' [2024-01-25T09:38:30.911] [206129.0] debug: task/cgroup: step physical cores are '0-1,6-7' [...] [2024-01-25T09:38:30.912] [206129.2] debug: task/cgroup: job abstract cores are '0-1,6-7' [2024-01-25T09:38:30.912] [206129.2] debug: task/cgroup: step abstract cores are '0-1,6-7' [2024-01-25T09:38:30.912] [206129.2] debug: task/cgroup: job physical cores are '0-1,6-7' [2024-01-25T09:38:30.912] [206129.2] debug: task/cgroup: step physical cores are '0-1,6-7' [...] [2024-01-25T09:38:30.922] [206129.3] debug: task/cgroup: job abstract cores are '0-1,6-7' [2024-01-25T09:38:30.922] [206129.3] debug: task/cgroup: step abstract cores are '0-1,6-7' [2024-01-25T09:38:30.922] [206129.3] debug: task/cgroup: job physical cores are '0-1,6-7' [2024-01-25T09:38:30.922] [206129.3] debug: task/cgroup: step physical cores are '0-1,6-7' It seems that SLURM is reserving me correctly four cores for my job but, when inside my script I run four "srun", then each srun can freely take one of the four reserved cores, not exactly one to each task. What am I doing wrong? Thanks a lot!

1 0

Slurm version 23.11.3 is now available
by Tim McMullan 24 Jan '24

24 Jan '24

We are pleased to announce the availability of Slurm version 23.11.3. The 23.11.3 fixes a single issue related to the new Debian package support. For 23.11.2, the debian changelog was not updated correctly and would generate packages as 23.11.1. This release will correctly generate 23.11.3 packages, with no other changes. Slurm can be downloaded from https://www.schedmd.com/downloads.php . -Tim > * Changes in Slurm 23.11.3 > ========================== > -- Fix debian/changelog file to reflect the correct version so the packages > are generated correctl. -- Tim McMullan Release Management, Support, and Development SchedMD LLC - Commercial Slurm Development and Support

1 0

Re: [slurm-users] slurmstepd: error: load_ebpf_prog: BPF load error (No space left on device). Please check your system limits (MEMLOCK).
by Tim Schneider 24 Jan '24

24 Jan '24

Hi, I just tested with 23.02.7-1 and the issue is gone. So it seems like the patch got released. Best, Tim On 1/24/24 16:55, Stefan Fleischmann wrote: > On Wed, 24 Jan 2024 12:37:04 -0300 Cristóbal Navarro > <cristobal.navarro.g(a)gmail.com> wrote: >> Many thanks >> One question? Do we have to apply this patch (and recompile slurm i >> guess) only on the compute-node with problems? >> Also, I noticed the patch now appears as "obsolete", is that ok? > We have Slurm installed on a NFS share, so what I did was to recompile > it and then I only replaced the library lib/slurm/cgroup_v2.so. Good > enough for now, I've been planning to update to 23.11 anyway soon. > > I suppose it's marked as obsolete because the patch went into a > release. According to the info in the bug report it should have been > included in 23.02.4. > > Cheers, > Stefan > >> On Wed, Jan 24, 2024 at 9:52 AM Stefan Fleischmann <sfle(a)kth.se> >> wrote: >> >>> Turns out I was wrong, this is not a problem in the kernel at all. >>> It's a known bug that is triggered by long bpf logs, see here >>> https://bugs.schedmd.com/show_bug.cgi?id=17210 >>> >>> There is a patch included there. >>> >>> Cheers, >>> Stefan >>> >>> On Tue, 23 Jan 2024 15:28:59 +0100 Stefan Fleischmann <sfle(a)kth.se> >>> wrote: >>>> I don't think there is much for SchedMD to do. As I said since it >>>> is working fine with newer kernels there doesn't seem to be any >>>> breaking change in cgroup2 in general, but only a regression >>>> introduced in one of the latest updates in 5.15. >>>> >>>> If Slurm was doing something wrong with cgroup2, and it >>>> accidentally worked until this recent change, then other kernel >>>> versions should show the same behavior. But as far as I can tell >>>> it still works just fine with newer kernels. >>>> >>>> Cheers, >>>> Stefan >>>> >>>> On Tue, 23 Jan 2024 15:20:56 +0100 >>>> Tim Schneider <tim.schneider1(a)tu-darmstadt.de> wrote: >>>> >>>>> Hi, >>>>> >>>>> I have filed a bug report with SchedMD >>>>> (https://bugs.schedmd.com/show_bug.cgi?id=18623), but the >>>>> support told me they cannot invest time in this issue since I >>>>> don't have a support contract. Maybe they will look into it >>>>> once it affects more people or someone important enough. >>>>> >>>>> So far, I have resorted to using 5.15.0-89-generic, but I am >>>>> also a bit concerned about the security aspect of this choice. >>>>> >>>>> Best, >>>>> >>>>> Tim >>>>> >>>>> On 23.01.24 14:59, Stefan Fleischmann wrote: >>>>>> Hi! >>>>>> >>>>>> I'm seeing the same in our environment. My conclusion is that >>>>>> it is a regression in the Ubuntu 5.15 kernel, introduced with >>>>>> 5.15.0-90-generic. Last working kernel version is >>>>>> 5.15.0-89-generic. I have filed a bug report here: >>>>>> https://bugs.launchpad.net/bugs/2050098 >>>>>> >>>>>> Please add yourself to the affected users in the bug report >>>>>> so it hopefully gets more attention. >>>>>> >>>>>> I've tested with newer kernels (6.5, 6.6 and 6.7) and the >>>>>> problem does not exist there. 6.5 is the latest hwe kernel >>>>>> for 22.04 and would be an option for now. Reverting back to >>>>>> 5.15.0-89 would work as well, but I haven't looked into the >>>>>> security aspects of that. >>>>>> >>>>>> Cheers, >>>>>> Stefan >>>>>> >>>>>> On Mon, 22 Jan 2024 13:31:15 -0300 >>>>>> cristobal.navarro.g at gmail.com wrote: >>>>>> >>>>>>> Hi Tim and community, >>>>>>> We are currently having the same issue (cgroups not working >>>>>>> it seems, showing all GPUs on jobs) on a GPU-compute node >>>>>>> (DGX A100) a couple of days ago after a full update (apt >>>>>>> upgrade). Now whenever we launch a job for that partition, >>>>>>> we get the error message mentioned by Tim. As a note, we >>>>>>> have another custom GPU-compute node with L40s, on a >>>>>>> different partition, and that one works fine. Before this >>>>>>> error, we always had small differences in kernel version >>>>>>> between nodes, so I am not sure if this can be the problem. >>>>>>> Nevertheless, here is the info of our nodes as well. >>>>>>> >>>>>>> *[Problem node]* The DGX A100 node has this kernel >>>>>>> cnavarro at nodeGPU01:~$ uname -a >>>>>>> Linux nodeGPU01 5.15.0-1042-nvidia #42-Ubuntu SMP Wed Nov 15 >>>>>>> 20:28:30 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux >>>>>>> >>>>>>> *[Functioning node]* The Custom GPU node (L40s) has this >>>>>>> kernel cnavarro at nodeGPU02:~$ uname -a >>>>>>> Linux nodeGPU02 5.15.0-91-generic #101-Ubuntu SMP Tue Nov 14 >>>>>>> 13:30:08 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux >>>>>>> >>>>>>> *And the login node *(slurmctld) >>>>>>> ? ~ uname -a >>>>>>> Linux patagon-master 5.15.0-91-generic #101-Ubuntu SMP Tue >>>>>>> Nov 14 13:30:08 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux >>>>>>> >>>>>>> Any ideas what we should check? >>>>>>> >>>>>>> On Thu, Jan 4, 2024 at 3:03?PM Tim Schneider <tim.schneider1 >>>>>>> at tu-darmstadt.de> wrote: >>>>>>> >>>>>>>> Hi, >>>>>>>> >>>>>>>> I am using SLURM 22.05.9 on a small compute cluster. Since I >>>>>>>> reinstalled two of our nodes, I get the following error when >>>>>>>> launching a job: >>>>>>>> >>>>>>>> slurmstepd: error: load_ebpf_prog: BPF load error (No space >>>>>>>> left on device). Please check your system limits (MEMLOCK). >>>>>>>> >>>>>>>> Also the cgroups do not seem to work properly anymore, as I >>>>>>>> am able to see all GPUs even if I do not request them, >>>>>>>> which is not the case on the other nodes. >>>>>>>> >>>>>>>> One difference I found between the updated nodes and the >>>>>>>> original nodes (both are Ubuntu 22.04) is the kernel >>>>>>>> version, which is "5.15.0-89-generic #99-Ubuntu SMP" on the >>>>>>>> functioning nodes and "5.15.0-91-generic #101-Ubuntu SMP" >>>>>>>> on the updated nodes. I could not figure out how to install >>>>>>>> the exact first kernel version on the updated nodes, but I >>>>>>>> noticed that when I reinstall 5.15.0 with this tool: >>>>>>>> https://github.com/pimlie/ubuntu-mainline-kernel.sh, the >>>>>>>> error message disappears. However, once I do that, the >>>>>>>> network driver does not function properly anymore, so this >>>>>>>> does not seem to be a good solution. >>>>>>>> >>>>>>>> Has anyone seen this issue before or is there maybe >>>>>>>> something else I should take a look at? I am also happy to >>>>>>>> just find a workaround such that I can take these nodes >>>>>>>> back online. >>>>>>>> >>>>>>>> I appreciate any help! >>>>>>>> >>>>>>>> Thanks a lot in advance and best wishes, >>>>>>>> >>>>>>>> Tim >>>>>>>> >>>>>>>> >>>>>>>> >>>> >>> >>

2 1

Re: [slurm-users] slurmstepd: error: load_ebpf_prog: BPF load error (No space left on device). Please check your system limits (MEMLOCK).
by Cristóbal Navarro 24 Jan '24

24 Jan '24

Many thanks One question? Do we have to apply this patch (and recompile slurm i guess) only on the compute-node with problems? Also, I noticed the patch now appears as "obsolete", is that ok? On Wed, Jan 24, 2024 at 9:52 AM Stefan Fleischmann <sfle(a)kth.se> wrote: > Turns out I was wrong, this is not a problem in the kernel at all. It's > a known bug that is triggered by long bpf logs, see here > https://bugs.schedmd.com/show_bug.cgi?id=17210 > > There is a patch included there. > > Cheers, > Stefan > > On Tue, 23 Jan 2024 15:28:59 +0100 Stefan Fleischmann <sfle(a)kth.se> > wrote: > > I don't think there is much for SchedMD to do. As I said since it is > > working fine with newer kernels there doesn't seem to be any breaking > > change in cgroup2 in general, but only a regression introduced in one > > of the latest updates in 5.15. > > > > If Slurm was doing something wrong with cgroup2, and it accidentally > > worked until this recent change, then other kernel versions should > > show the same behavior. But as far as I can tell it still works just > > fine with newer kernels. > > > > Cheers, > > Stefan > > > > On Tue, 23 Jan 2024 15:20:56 +0100 > > Tim Schneider <tim.schneider1(a)tu-darmstadt.de> wrote: > > > > > Hi, > > > > > > I have filed a bug report with SchedMD > > > (https://bugs.schedmd.com/show_bug.cgi?id=18623), but the support > > > told me they cannot invest time in this issue since I don't have a > > > support contract. Maybe they will look into it once it affects more > > > people or someone important enough. > > > > > > So far, I have resorted to using 5.15.0-89-generic, but I am also a > > > bit concerned about the security aspect of this choice. > > > > > > Best, > > > > > > Tim > > > > > > On 23.01.24 14:59, Stefan Fleischmann wrote: > > > > Hi! > > > > > > > > I'm seeing the same in our environment. My conclusion is that it > > > > is a regression in the Ubuntu 5.15 kernel, introduced with > > > > 5.15.0-90-generic. Last working kernel version is > > > > 5.15.0-89-generic. I have filed a bug report here: > > > > https://bugs.launchpad.net/bugs/2050098 > > > > > > > > Please add yourself to the affected users in the bug report so it > > > > hopefully gets more attention. > > > > > > > > I've tested with newer kernels (6.5, 6.6 and 6.7) and the problem > > > > does not exist there. 6.5 is the latest hwe kernel for 22.04 and > > > > would be an option for now. Reverting back to 5.15.0-89 would work > > > > as well, but I haven't looked into the security aspects of that. > > > > > > > > Cheers, > > > > Stefan > > > > > > > > On Mon, 22 Jan 2024 13:31:15 -0300 > > > > cristobal.navarro.g at gmail.com wrote: > > > > > > > >> Hi Tim and community, > > > >> We are currently having the same issue (cgroups not working it > > > >> seems, showing all GPUs on jobs) on a GPU-compute node (DGX A100) > > > >> a couple of days ago after a full update (apt upgrade). Now > > > >> whenever we launch a job for that partition, we get the error > > > >> message mentioned by Tim. As a note, we have another custom > > > >> GPU-compute node with L40s, on a different partition, and that > > > >> one works fine. Before this error, we always had small > > > >> differences in kernel version between nodes, so I am not sure if > > > >> this can be the problem. Nevertheless, here is the info of our > > > >> nodes as well. > > > >> > > > >> *[Problem node]* The DGX A100 node has this kernel > > > >> cnavarro at nodeGPU01:~$ uname -a > > > >> Linux nodeGPU01 5.15.0-1042-nvidia #42-Ubuntu SMP Wed Nov 15 > > > >> 20:28:30 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux > > > >> > > > >> *[Functioning node]* The Custom GPU node (L40s) has this kernel > > > >> cnavarro at nodeGPU02:~$ uname -a > > > >> Linux nodeGPU02 5.15.0-91-generic #101-Ubuntu SMP Tue Nov 14 > > > >> 13:30:08 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux > > > >> > > > >> *And the login node *(slurmctld) > > > >> ? ~ uname -a > > > >> Linux patagon-master 5.15.0-91-generic #101-Ubuntu SMP Tue Nov 14 > > > >> 13:30:08 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux > > > >> > > > >> Any ideas what we should check? > > > >> > > > >> On Thu, Jan 4, 2024 at 3:03?PM Tim Schneider <tim.schneider1 at > > > >> tu-darmstadt.de> wrote: > > > >> > > > >>> Hi, > > > >>> > > > >>> I am using SLURM 22.05.9 on a small compute cluster. Since I > > > >>> reinstalled two of our nodes, I get the following error when > > > >>> launching a job: > > > >>> > > > >>> slurmstepd: error: load_ebpf_prog: BPF load error (No space left > > > >>> on device). Please check your system limits (MEMLOCK). > > > >>> > > > >>> Also the cgroups do not seem to work properly anymore, as I am > > > >>> able to see all GPUs even if I do not request them, which is not > > > >>> the case on the other nodes. > > > >>> > > > >>> One difference I found between the updated nodes and the > > > >>> original nodes (both are Ubuntu 22.04) is the kernel version, > > > >>> which is "5.15.0-89-generic #99-Ubuntu SMP" on the functioning > > > >>> nodes and "5.15.0-91-generic #101-Ubuntu SMP" on the updated > > > >>> nodes. I could not figure out how to install the exact first > > > >>> kernel version on the updated nodes, but I noticed that when I > > > >>> reinstall 5.15.0 with this tool: > > > >>> https://github.com/pimlie/ubuntu-mainline-kernel.sh, the error > > > >>> message disappears. However, once I do that, the network driver > > > >>> does not function properly anymore, so this does not seem to be > > > >>> a good solution. > > > >>> > > > >>> Has anyone seen this issue before or is there maybe something > > > >>> else I should take a look at? I am also happy to just find a > > > >>> workaround such that I can take these nodes back online. > > > >>> > > > >>> I appreciate any help! > > > >>> > > > >>> Thanks a lot in advance and best wishes, > > > >>> > > > >>> Tim > > > >>> > > > >>> > > > >>> > > > -- Cristóbal A. Navarro

2 1

slurm-config on NFS-volume
by Werf, C.G. van der (Carel) 24 Jan '24

24 Jan '24

Hi, Among other clusters, I have a simple cluster with 2 nodes, running slurm. 1 node runs : mysqld, slurmdbd, slurmctld and slurmd. The other node, only runs slurmd. Slurm config is in node1: /etc/slurm. A copy of the config is in node2:/etc/slurm. This slurm configuration runs ok. But, as I am normally used to configure on a shared location, where there is only 1 copy of the config, I tried to change the current setup by moving the slurm-config to a shared (NFS) disk. So on both nodes : /etc/slurm -> nfs-server:/shared/etc/slurm But, I have tried several ways... The slurmctld and slurmd daemons will not start when I refer to this "nfs-shared" location. Any ideas what is missing here ? Authorizations seem ok. With Regards, | Carel van der Werf | | Developer/Administrator Linux | ICT-Bèta | Department of Science | Utrecht University |

3 3

error: Couldn't find the specified plugin name for cred/munge looking at all files
by Jesse Aiton 24 Jan '24

24 Jan '24

Hello Slurm Folks, I have a weird issue where on the same server, which acts as both a controller and a node, slurmctld can’t find cred_munge.so slurmctld: debug3: Trying to load plugin /app/slurm-24.0.8/lib/slurm/cred_munge.so slurmctld: debug4: /app/slurm-24.0.8/lib/slurm/cred_munge.so: Does not exist or not a regular file. slurmctld: error: Couldn't find the specified plugin name for cred/munge looking at all files slurmctld: error: cannot open plugin directory /app/slurm-24.0.8/lib/slurm slurmctld: error: cannot find cred plugin for cred/munge slurmctld: error: cannot create cred context for cred/munge slurmctld: fatal: failed to initialize cred plugin But slurmd can: slurmd: debug3: Trying to load plugin /app/slurm-24.0.8/lib/slurm/cred_munge.so slurmd: debug3: plugin_load_from_file->_verify_syms: found Slurm plugin name:Munge credential signature plugin type:cred/munge version:0x180800 slurmd: cred/munge: init: Munge credential signature plugin loaded slurmd: debug3: Success. This is on Ubuntu 20.04 and happens both with Slurm 20.11.09 and 24.0.8 Thank you, Jesse # slurm.conf file generated by configurator easy.html. # Put this file on all nodes of your cluster. # See the slurm.conf man page for more information. # ClusterName=prod-cluster SlurmctldHost=controller # #MailProg=/bin/mail #MpiDefault= #MpiParams=ports=#-# ProctrackType=proctrack/cgroup ReturnToService=1 SlurmctldPidFile=/var/run/slurmctld.pid #SlurmctldPort=6817 SlurmdPidFile=/var/run/slurmd.pid #SlurmdPort=6818 SlurmdSpoolDir=/var/spool/slurmd SlurmUser=slurm #SlurmdUser=root StateSaveLocation=/var/spool/slurmctld #SwitchType= TaskPlugin=task/affinity,task/cgroup # # # TIMERS #KillWait=30 #MinJobAge=300 #SlurmctldTimeout=120 #SlurmdTimeout=300 # # # SCHEDULING SchedulerType=sched/backfill SelectType=select/cons_tres # # # LOGGING AND ACCOUNTING #AccountingStorageType= #JobAcctGatherFrequency=30 #JobAcctGatherType= #SlurmctldDebug=info SlurmctldLogFile=/var/log/slurmctld.log #SlurmdDebug=info SlurmdLogFile=/var/log/slurmd.log # # # COMPUTE NODES NodeName=controller CPUs=1 State=UNKNOWN NodeName=node CPUs=1 State=UNKNOWN PartitionName=prod-part Nodes=ALL Default=YES MaxTime=INFINITE State=UP

3 5

Slurm version 23.11.2 is now available
by Tim McMullan 23 Jan '24

23 Jan '24

We are pleased to announce the availability of Slurm version 23.11.2. The 23.11.2 release includes a number of fixes to stability and various bug fixes. Some notable changes include several fixes to the new scontrol reconfigure method, including one that could result in jobs getting cancelled prematurely, a couple errors that resulted in the backup slurmctld stopping on fail-back, and an issue during upgrades with older MySQL versions with a small max_allowed_packet value for sites with a large number of associations. Slurm can be downloaded from https://www.schedmd.com/downloads.php . -Tim > * Changes in Slurm 23.11.2 > ========================== > -- slurmrestd - Reject single http query with multiple path requests. > -- Fix launching Singularity v4.x containers with srun --container by setting > .process.terminal to true in generated config.json when step has > pseudoterminal (--pty) requested. > -- Fix loading in dyanmic/cloud node jobs after net_cred expired. > -- Fix cgroup null path error on slurmd/slurmstepd tear down. > -- data_parser/v0.0.40 - Prevent failure if accounting is disabled, instead > issue a warning if needed data from the database can not be retrieved. > -- openapi/slurmctld - Prevent failure if accounting is disabled. > -- Prevent slurmscriptd processing delays from blocking other threads in > slurmctld while trying to launch various scripts. This is additional work > for a fix in 23.02.6. > -- Fix memory leak when receiving alias addrs from controller. > -- scontrol - Accept `scontrol token lifespan=infinite` to create tokens that > effectively do not expire. > -- Avoid errors when Slurmdb accounting disabled when '--json' or '--yaml' is > invoked with CLI commands and slurmrestd. Add warnings when query would > have populated data from Slurmdb instead of errors. > -- Fix slurmctld memory leak when running job with --tres-per-task=gres:shard:# > -- Fix backfill trying to start jobs outside of backfill window. > -- Fix oversubscription on partitions with PreemptMode=OFF. > -- Preserve node reason on power up if the node is downed or drained. > -- data_parser/v0.0.40 - Avoid aborting when invoking a not implemented > parser. > -- data_parser/v0.0.40 - Fix how nice values are parsed for job submissions. > -- data_parser/v0.0.40 - Fix regression where parsing error did not result in > invalid request being rejected. > -- Fix segfault in front-end node registration. > -- Prevent jobs using none typed gpus from being killed by the controller after > a reconfig or restart. > -- Fix deadlock situation in the dbd when adding associations. > -- Update default values of text/blob columns when updating from old mysql > versions in more situations. This improves a previous fix to handle an > uncommon case when upgrading mysql/mariadb. > -- Fix rpmbuild in openSUSE/SLES due to incorrect mariadb dependency. > -- Fix compilation on RHEL 7. > -- When upgrading the slurmdbd to 23.11, avoid generating a query to update > the association table that is larger than max_allowed_packet which would > result in an upgrade failure. > -- Fix rare deadlock when a dynamic node registers at the same time that a > once per minute background task occurs. > -- Fix build issue on 32-bit systems. > -- data_parser/v0.0.40 - Fix enumerated strings in OpenAPI specification not > have type field specified. > -- Improve scontrol show job -d information of used shared gres (shard/mps) > topology. > -- Allow Slurm to compile without MUNGE if --without-munge is used as an > argument to configure. > -- accounting_storage/mysql - Fix usage query to use new lineage column > instead of lft/rgt. > -- slurmrestd - Improve handling of missing parsers when content plugins > expect parsers not loaded. > -- slurmrestd - Correct parsing of StepIds when querying jobs. > -- slurmrestd - Improve error from parsing failures of lists. > -- slurmrestd - Improve parsing of singular values for lists. > -- accounting_storage/mysql - Fix PrivateData=User when listing associations. > -- Disable sorting of dynamic nodes to avoid issues when restarting with > heterogenous jobs that cause jobs to abort on restart. > -- Don't allow deletion of non-dynamic nodes. > -- accounting_storage/mysql - Fix issue adding partition based associations. > -- Respect non-"slurm" settings for I_MPI_HYDRA_BOOTSTRAP and HYDRA_BOOTSTRAP > and avoid injecting the --external-launcher option which will cause > mpirun/mpiexec to fail with an unexpected argument error. > -- Fix bug where scontrol hold would change node count for jobs with > implicitly defined node counts. > -- data_parser/v0.0.40 - Fix regression of support for "hold" in > job description. > -- Avoid sending KILL RPCs to unresolvable POWERING_UP and POWERED_DOWN nodes. > -- data_parser/v0.0.38 - Fix several potential NULL dereferences that could > cause slurmrestd to crash. > -- Add --gres-flags=one-task-per-sharing. Do not allow different tasks in to be > allocated shared gres from the same sharing gres. > -- Add SelectTypeParameters=ENFORCE_BINDING_GRES and ONE_TASK_PER_SHARING_GRES. > This gives default behavior for a job's --gres-flags. > -- Alter the networking code to try connecting to the backup controllers > if the DNS lookup for the primary SlurmctldHost fails. > -- Alter the name resolution to only log at verbose() in client commands > on failures. This allows for HA setups where the DNS entries are withdrawn > for some SlurmctldHost entries without flooding the user with errors. > -- Prevent slurmscriptd PID leaks when running slurmctld in foreground mode. > -- Open all slurmctld listening ports at startup, and persist throughout. > This also changes the backup slurmctld process to open the SlurmctldPort > range, instead of only the first. > -- Fix backup slurmctld shutting down instead of resuming standby duty if it > took control. > -- Fix race condition that delayed the primary slurmctld resuming when taking > control from a backup controller. > -- srun - Ensure processed messages are meant for this job in case of a > rapidly-reused TCP port. > -- srun - Prevent step launch failure while waiting for step allocation if > a stray message is received. > -- Fix backup slurmctld to be able to respond to configless config file > requests correctly. > -- Fix slurmctld crashing when recovering from a failed reconfigure. > -- Fix slurmscriptd operation after recovering from a failed reconfigure. -- Tim McMullan Release Management, Support, and Development SchedMD LLC - Commercial Slurm Development and Support

1 0

GPU devices mapping with job's cgroup in cgroups v2 using eBPF
by Mahendra Paipuri 23 Jan '24

23 Jan '24

Hello all, Happy new year! We have recently upgraded the cgroups on our SLURM cluster to v2. In cgroups v1, the interface `/devices.list` used to have the information of which device has been attached to that particular cgroup. From my understanding, cgroups v2 use eBPF to manage devices and so as SLURM to manage the GPUs. I was looking for a way to be able to programatically determine the job cgroups to device mapping and I came across this thread ( https://bugzilla.redhat.com/show_bug.cgi?id=1717396) which has a similar discussion in the context of VMs. So, I have used `bpftool` to inspect the job cgroups. An example output: ``` # /tmp/bpftool cgroup list /sys/fs/cgroup/system.slice/slurmstepd.scope/job_1956132 ID AttachType AttachFlags Name ``` When I add `effective` flag, I see the attached eBPF program ``` # /tmp/bpftool cgroup list /sys/fs/cgroup/system.slice/slurmstepd.scope/job_1956132 effective ID AttachType Name 4197 cgroup_device Slurm_Cgroup_v2 ``` >From my understand, `effective` flag shows the inherited eBPF programs as well. So, my question is at which level of cgroups the eBPF program is attached? I tried to inspect at various levels but all of them returned none. Then looking into translated byte code of eBPF program, I get the following ``` # /tmp/bpftool prog dump xlated id 4197 0: (61) r2 = *(u32 *)(r1 +0) 1: (54) w2 &= 65535 2: (61) r3 = *(u32 *)(r1 +0) 3: (74) w3 >>= 16 4: (61) r4 = *(u32 *)(r1 +4) 5: (61) r5 = *(u32 *)(r1 +8) 6: (55) if r2 != 0x2 goto pc+4 7: (55) if r4 != 0xc3 goto pc+3 8: (55) if r5 != 0x0 goto pc+2 9: (b7) r0 = 0 10: (95) exit 11: (55) if r2 != 0x2 goto pc+4 12: (55) if r4 != 0xc3 goto pc+3 13: (55) if r5 != 0x1 goto pc+2 14: (b7) r0 = 0 15: (95) exit 16: (55) if r2 != 0x2 goto pc+4 17: (55) if r4 != 0xc3 goto pc+3 18: (55) if r5 != 0x2 goto pc+2 19: (b7) r0 = 0 20: (95) exit 21: (55) if r2 != 0x2 goto pc+4 22: (55) if r4 != 0xc3 goto pc+3 23: (55) if r5 != 0x3 goto pc+2 24: (b7) r0 = 1 25: (95) exit 26: (b7) r0 = 1 27: (95) exit ``` >From the output, it is clear that GPU:3 (among 0,1,2,3) is the one that is attached to that job's cgroup. However, I was looking for a way to dump eBPF maps that can directly provide the major, minor numbers and permissions of device as discussed in the comment (https://bugzilla.redhat.com/show_bug.cgi?id=1717396#c5). When I inspect eBPF program, I dont see any maps associated. ``` # /tmp/bpftool prog list id 4197 4197: cgroup_device name Slurm_Cgroup_v2 tag 1a261c8a913ff67c gpl loaded_at 2024-01-02T08:19:56+0100 uid 0 xlated 224B jited 142B memlock 4096B ``` So, my second question is how can I get a similar information as `map dump` that can give us the device's major minor numbers directly instead of parsing the byte code from `prog dump`? I am still discovering the eBPF ecosystem so if I am missing something very obvious, please let me know. I would really appreciate that. Cheers! Regards Mahendra

2 1

Re: [slurm-users] slurmstepd: error: load_ebpf_prog: BPF load error (No space left on device). Please check your system limits (MEMLOCK).
by Tim Schneider 23 Jan '24

23 Jan '24

Hi, I have filed a bug report with SchedMD (https://bugs.schedmd.com/show_bug.cgi?id=18623), but the support told me they cannot invest time in this issue since I don't have a support contract. Maybe they will look into it once it affects more people or someone important enough. So far, I have resorted to using 5.15.0-89-generic, but I am also a bit concerned about the security aspect of this choice. Best, Tim On 23.01.24 14:59, Stefan Fleischmann wrote: > Hi! > > I'm seeing the same in our environment. My conclusion is that it is a > regression in the Ubuntu 5.15 kernel, introduced with 5.15.0-90-generic. > Last working kernel version is 5.15.0-89-generic. I have filed a bug > report here: https://bugs.launchpad.net/bugs/2050098 > > Please add yourself to the affected users in the bug report so it > hopefully gets more attention. > > I've tested with newer kernels (6.5, 6.6 and 6.7) and the problem does > not exist there. 6.5 is the latest hwe kernel for 22.04 and would be an > option for now. Reverting back to 5.15.0-89 would work as well, but I > haven't looked into the security aspects of that. > > Cheers, > Stefan > > On Mon, 22 Jan 2024 13:31:15 -0300 > cristobal.navarro.g at gmail.com wrote: > >> Hi Tim and community, >> We are currently having the same issue (cgroups not working it seems, >> showing all GPUs on jobs) on a GPU-compute node (DGX A100) a couple >> of days ago after a full update (apt upgrade). Now whenever we launch >> a job for that partition, we get the error message mentioned by Tim. >> As a note, we have another custom GPU-compute node with L40s, on a >> different partition, and that one works fine. >> Before this error, we always had small differences in kernel version >> between nodes, so I am not sure if this can be the problem. >> Nevertheless, here is the info of our nodes as well. >> >> *[Problem node]* The DGX A100 node has this kernel >> cnavarro at nodeGPU01:~$ uname -a >> Linux nodeGPU01 5.15.0-1042-nvidia #42-Ubuntu SMP Wed Nov 15 20:28:30 >> UTC 2023 x86_64 x86_64 x86_64 GNU/Linux >> >> *[Functioning node]* The Custom GPU node (L40s) has this kernel >> cnavarro at nodeGPU02:~$ uname -a >> Linux nodeGPU02 5.15.0-91-generic #101-Ubuntu SMP Tue Nov 14 13:30:08 >> UTC 2023 x86_64 x86_64 x86_64 GNU/Linux >> >> *And the login node *(slurmctld) >> ? ~ uname -a >> Linux patagon-master 5.15.0-91-generic #101-Ubuntu SMP Tue Nov 14 >> 13:30:08 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux >> >> Any ideas what we should check? >> >> On Thu, Jan 4, 2024 at 3:03?PM Tim Schneider <tim.schneider1 at >> tu-darmstadt.de> wrote: >> >>> Hi, >>> >>> I am using SLURM 22.05.9 on a small compute cluster. Since I >>> reinstalled two of our nodes, I get the following error when >>> launching a job: >>> >>> slurmstepd: error: load_ebpf_prog: BPF load error (No space left on >>> device). Please check your system limits (MEMLOCK). >>> >>> Also the cgroups do not seem to work properly anymore, as I am able >>> to see all GPUs even if I do not request them, which is not the >>> case on the other nodes. >>> >>> One difference I found between the updated nodes and the original >>> nodes (both are Ubuntu 22.04) is the kernel version, which is >>> "5.15.0-89-generic #99-Ubuntu SMP" on the functioning nodes and >>> "5.15.0-91-generic #101-Ubuntu SMP" on the updated nodes. I could >>> not figure out how to install the exact first kernel version on the >>> updated nodes, but I noticed that when I reinstall 5.15.0 with this >>> tool: https://github.com/pimlie/ubuntu-mainline-kernel.sh, the >>> error message disappears. However, once I do that, the network >>> driver does not function properly anymore, so this does not seem to >>> be a good solution. >>> >>> Has anyone seen this issue before or is there maybe something else I >>> should take a look at? I am also happy to just find a workaround >>> such that I can take these nodes back online. >>> >>> I appreciate any help! >>> >>> Thanks a lot in advance and best wishes, >>> >>> Tim >>> >>> >>>

2 1

Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres)
by Kherfani, Hafedh (Professional Services, TC) 23 Jan '24

23 Jan '24

Hello Experts, I'm a new Slurm user (so please bare with me :) ...). Recently we've deployed Slurm version 23.11 on a very simple cluster, which consists of a Master node (acting as a Login & Slurmdbd node as well), a Compute Node which has a NVIDIA HGX A100-SXM4-40GB GPU, detected as 4 x GPU's: GPU [0-4], and a Storage Array presenting/sharing the NFS disk (where users' home directories will be created as well). The problem is that I've never been able to run a simple/dummy batch script in a parallel way using the 4 GPU's. In fact, running the same command "sbatch gpu-job.sh" multiple times shows that only one single job is running, while the other jobs are in a pending state: [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh Submitted batch job 214 [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh Submitted batch job 215 [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh Submitted batch job 216 [slurmtest@c-a100-master test-batch-scripts]$ squeue JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 216 gpu gpu-job slurmtest PD 0:00 1 (None) 215 gpu gpu-job slurmtest PD 0:00 1 (Priority) 214 gpu gpu-job slurmtest PD 0:00 1 (Priority) 213 gpu gpu-job slurmtest PD 0:00 1 (Priority) 212 gpu gpu-job slurmtest PD 0:00 1 (Resources) 211 gpu gpu-job slurmtest R 0:14 1 c-a100-cn01 PS: CPU jobs (i.e. using the default debug partition, without call the GPU Gres) can be run in parallel. The issue with running parallel jobs is only seen when using the GPU's as Gres. I've tried many combinations of settings in gres.conf and slurm.conf, many (if not most) of these combinations would result in error messages in slurmctld and slurmd logs. The current gres.conf and slurm.conf contents is shown below. Even though it doesn't give errors when restarting slurmctld and slurmd services (on master and compute nodes, resp.), but as I said, it doesn't allow jobs to be executed in parallel. Batch script contents shared below as well, in order to give more clarity on what I'm trying to do: [root@c-a100-master slurm]# cat gres.conf | grep -v "^#" NodeName=c-a100-cn01 AutoDetect=nvml Name=gpu Type=A100 File=/dev/nvidia[0-3] [root@c-a100-master slurm]# cat slurm.conf | grep -v "^#" | egrep -i "AccountingStorageTRES|GresTypes|NodeName|partition" GresTypes=gpu AccountingStorageTRES=gres/gpu NodeName=c-a100-cn01 Gres=gpu:A100:4 CPUs=64 Boards=1 SocketsPerBoard=1 CoresPerSocket=32 ThreadsPerCore=2 RealMemory=515181 State=UNKNOWN PartitionName=debug Nodes=ALL Default=YES MaxTime=INFINITE State=UP PartitionName=gpu Nodes=ALL MaxTime=10:0:0 [slurmtest@c-a100-master test-batch-scripts]$ cat gpu-job.sh #!/bin/bash #SBATCH --job-name=gpu-job #SBATCH --partition=gpu #SBATCH --nodes=1 #SBATCH --gpus-per-node=4 #SBATCH --gres=gpu:4 #SBATCH --tasks-per-node=1 #SBATCH --output=gpu_job_output.%j # Output file name (replaces %j with job ID) #SBATCH --error=gpu_job_error.%j # Error file name (replaces %j with job ID) hostname date sleep 40 pwd Any help on which changes need to be made to the config files (mainly slurm.conf and gres.cong) and/or the batch script, so that multiple jobs can be in a "Running" state at the same time (in parallel) ? Thanks in advance for your help ! Best regards, Hafedh Kherfani

8 10

propose environment variables SLURM_STDOUT, SLURM_STDERR, SLURM_STDIN
by urbanjost 22 Jan '24

22 Jan '24

RE: Placing the full pathname of the job stdout in an environment variable Would others find it useful if new variables were added that contained the full pathnames of the standard input, error and input files of batch jobs? ## SYNOPSIS Proposed new environment variables SLURM_STDOUT,SLURM_STDERR, and SLURM_STDIN pointing to the full pathnames of the output, error output, and input files of a task would be a useful feature for batch jobs in particular. ## PROBLEM There are cases where it is desirable to have a job process the standard files (stdin, stdout, stderr) from a batch job from within the job instead of later via an epilog or via post-processing. Just a few examples where the job might want to reference the full pathname of stdout: * copy the file to a remote machine (via scp(1), for example) * move the file from local or scratch space to permanent globally accessible storage * remove the file depending on completion status * mail the output * archive it * post-process it Slurm commands like scontrol(1) and squeue(1) can help you locate the file names. But the names returned do not always expand all the macros allowed when specifying the names, and require calling a Slurm command (which might overload Slurm if millions of jobs are submitted or cause problems if Slurm is not responding for any reason(?)). For example: #!/bin/bash # Minimalist submit script for sbatch(1) #SBATCH --nodes 1-1 --ntasks=1 --time 0-0:1:00 --chdir=/tmp #SBATCH --output=stdout.%j #SBATCH --error=stderr.%A:%a:%J:%j:%N:%n:%s:%t:%u.out # query filenames via squeue(1) export SLURM_STDOUT=$(squeue --noheader --Format=STDOUT: --job=$SLURM_JOBID) export SLURM_STDERR=$(squeue --noheader --Format=STDERR: --job=$SLURM_JOBID) # PS: Current documentation for squeue(1) does not explicitly tell the # user a null size works. # query filenames via scontrol(1) declare -x $(scontrol show job=$SLURM_JOBID|grep StdOut) declare -x $(scontrol show job=$SLURM_JOBID|grep StdErr) cat <<EOF SLURM_STDOUT=$SLURM_STDOUT SLURM_STDERR=$SLURM_STDERR StdOut=$StdOut StdErr=$StdErr EOF ls stdout.* stderr.* The resulting output shows either none or some macros are expanded by the commands (and I am not sure all commands will always return a full pathname(?)) SLURM_STDOUT=/home/urbanjs/venus/V600/stdout.%j SLURM_STDERR=/home/urbanjs/venus/V600/stderr.%A:%a:%J:%j:%N:%n:%s:%t:%u.out StdOut=/home/urbanjs/venus/V600/stdout.96 StdErr=/home/urbanjs/venus/V600/stderr.96:4294967294:%J:96:%N:%n:%s:%t:urbanjs.out stderr.96:4294967294:96:96:mercury:0:4294967294:0:urbanjs.out stdout.96 One currently available work-around would be that the user just avoid the filename macros and always specify the filenames using a standard convention (which would obviously have to be specific to a specific platform). This is error-prone as it requires the user to strictly follow a protocol where the pathnames are specified that can be overridden by command-line options on sbatch(1) and so-on. Alternatively instead of the new environment variables Slurm could have a command option that returns the fully expanded names always expanding the macros. But if you query Slurm you have to worry about scaleability. Is it OK for a 100 000, jobs to query Slurm simulataneously via squeue(1) or scontrol(1)? What happens if for external reasons the Slurm daemons are down or not responding? As a current work-around using commands like realpath(1), stat(1), ls(1),find(1),and getting pids with fuser(1), pidof(1), and ps(1) become attractive. I find using the realpath(1) command at the top of the job and $SLURM_TASK_ID works the best. realpath(1) may not be available on all platforms(?) as well as the /proc/$PID/fd/1 file, in which case the other commands might be used to the same effect but it would be nice to have a simple standard way that does not require calling a command. The proposed environment variables seem the most obvious solution. # do this early in the job in case the user changes or closes # file descriptor 1 ... export SLURM_STDOUT=$(realpath /proc/$SLURM_TASK_PID/fd/1) # create softlink OUTPUT to stdout of job ln -s -f $SLURM_STDOUT OUTPUT # now user can do things like mailx(1) or scp(1) the OUTPUT file. Other reasons I might want to know the stdout pathname are to minimize network traffic and filespace on central servers. I might specify the output of a job goes to scratch or local storage or a memory-resident file system and move it all to a location on another system or a long-term archive instead of a central area that might be filled by other jobs or that I have a quota on, etc. So since the stdout might not be easily accessible from my other sessions also having something like "scontrol write batch_stdout=$SLURM_JOBID" that shows the stdout of the job would also be useful. Note the LSF job scheduler bpeek(1) command allows for a tail(1)-like interface to the job stdout to support the stdout not being accessible from all system access points, just as described. The current situation where in typical usage the output file is often in a globally mounted area and I can just access it from other nodes with grep(1), tail(1), and so on covers many users needs in simple cluster configurations; but there are several scenarios as described above primarily encountered by people running many thousands of single-node jobs where assigning output to a local scratch device and then selectively processing at job termination is preferable. As mentioned, those users can specify such a local name with "#SBATCH --output" and then, knowing it, process it as desired but having something like $SLURM_STDOUT being the full pathname to the file is much more generic. ## SUMMARY So adding the variables creates an ability for users to reliably access the standard files of a job with a low-load scalable method; and being a SLURM_* module it could be much more reliably accessed via prolog and epilog scripts and module(1) scripts as well. ## PS: Related to this would be instead of just --chdir a --mkdir and --scratch #SBATCH option would be useful that expanded file macros that would be similar to --chdir but would make sure the directory existed (using only user privelages) and a --scratch that would do the same as --mkdir but remove the directory at job termination. urbanjs@mercury:~/SLURM/proposals$ Sent with [Proton Mail](https://proton.me/) secure email.

3 2

slurmstepd: error: load_ebpf_prog: BPF load error (No space left on device). Please check your system limits (MEMLOCK).
by Tim Schneider 22 Jan '24

22 Jan '24

Hi, I am using SLURM 22.05.9 on a small compute cluster. Since I reinstalled two of our nodes, I get the following error when launching a job: slurmstepd: error: load_ebpf_prog: BPF load error (No space left on device). Please check your system limits (MEMLOCK). Also the cgroups do not seem to work properly anymore, as I am able to see all GPUs even if I do not request them, which is not the case on the other nodes. One difference I found between the updated nodes and the original nodes (both are Ubuntu 22.04) is the kernel version, which is "5.15.0-89-generic #99-Ubuntu SMP" on the functioning nodes and "5.15.0-91-generic #101-Ubuntu SMP" on the updated nodes. I could not figure out how to install the exact first kernel version on the updated nodes, but I noticed that when I reinstall 5.15.0 with this tool: https://github.com/pimlie/ubuntu-mainline-kernel.sh, the error message disappears. However, once I do that, the network driver does not function properly anymore, so this does not seem to be a good solution. Has anyone seen this issue before or is there maybe something else I should take a look at? I am also happy to just find a workaround such that I can take these nodes back online. I appreciate any help! Thanks a lot in advance and best wishes, Tim

2 1

Tried setting up GANG scheduling for timeslicing, but jobs are not alternating
by Francisco José Letterio 22 Jan '24

22 Jan '24

I'm trying to set up GANG scheduling with Slurm on my single-node server so the people at the lab can run experiments without blocking each other (so if say someone has to run some code that takes days to finish, other jobs that take less have the chance to run alternated with it and so they don't have to wait days until they can run) I followed the GANG scheduling slurm.conf setup tutorial on the slurm website, and to check if it's working properly I launched a bunch of jobs that print the current time and then sleep for a while. But when I check squeue, the jobs never alternate, they run sequentially one after the other. Why is this happening? Here's my slurm.conf file: # See the slurm.conf man page for more information. ClusterName=localcluster SlurmctldHost=localhost ProctrackType=proctrack/linuxproc ReturnToService=1 SlurmctldPidFile=/var/run/slurmctld.pid SlurmctldPort=6817 SlurmdPidFile=/var/run/slurmd.pid SlurmdPort=6818 SlurmdSpoolDir=/var/lib/slurm/slurmd SlurmUser=slurm StateSaveLocation=/var/lib/slurm/slurmctld TaskPlugin=task/none # TIMERS InactiveLimit=0 KillWait=30 MinJobAge=300 SlurmctldTimeout=120 SlurmdTimeout=300 Waittime=0 # SCHEDULING #Seteado para que cada laburo alterne cada 15 segundos SchedulerTimeSlice=15 SchedulerType=sched/builtin SelectType=select/linear SelectTypeParameters=CR_Memory PreemptMode=GANG # LOGGING AND ACCOUNTING JobCompType=jobcomp/none JobAcctGatherFrequency=30 SlurmctldDebug=info SlurmctldLogFile=/var/log/slurm/slurmctld.log SlurmdDebug=info SlurmdLogFile=/var/log/slurm/slurmd.log # COMPUTE NODES NodeName=lab04 CPUs=48 CoresPerSocket=12 ThreadsPerCore=2 State=UNKNOWN RealMemory=257249 PartitionName=LocalQ Nodes=ALL Default=YES MaxTime=INFINITE State=UP OverSubscribe=FORCE:6 DefMemPerNode=257249 MaxMemPerNode=257249 >From what I understand, SchedulerTimeSlice=15 means that jobs should alternate running every 15 seconds. This is the job I'm launching with sbatch (launching many copies of this job one after the other): #!/bin/bash #SBATCH -J test # Job name #SBATCH -o job.%j.out # Name of stdout output file (%j expands to %jobId) #SBATCH -N 1 # Total number of nodes requested echo "Test output from Slurm Testjob" date sleep 10 date sleep 10 date sleep 10 date sleep 10 I would expect jobs to print one or two dates, then the Slurm scheduler comes and lets another job run in the meantime, and then the final print(s) come with a delay way greater than 10secs However after launching many copies of this job this is what I see on squeue : Before the first job launched is done: JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 18 LocalQ test eiarussi PD 0:00 1 (Resources) 19 LocalQ test eiarussi PD 0:00 1 (Priority) 20 LocalQ test eiarussi PD 0:00 1 (Priority) 21 LocalQ test eiarussi PD 0:00 1 (Priority) 17 LocalQ test eiarussi R 0:31 1 lab04 After that job ends: JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 19 LocalQ test eiarussi PD 0:00 1 (Resources) 20 LocalQ test eiarussi PD 0:00 1 (Priority) 21 LocalQ test eiarussi PD 0:00 1 (Priority) 18 LocalQ test eiarussi R 0:02 1 lab04 Why did the first job run for 30+ seconds uninterrupted, instead of another job being allowed to run after 15 seconds? Am I misunderstanding how GANG scheduling works, or is the a problem in my conf file?

1 0

MIG-Slice: Unavailable GRES
by Dražen Jalšovec 21 Jan '24

21 Jan '24

Hi, We are testing the MIG deployment on our new slurm compute node with 4 x H100 GPUs. It looks like everything is configured correctly but we have a problem accessing mig devices. When I submit jobs requesting a mig gpu device #SBATCH --gres=gpu:H100_1g.10gb:1, the jobs get submitted to the node, but only 4 jobs get executed and all other jobs fail. I was able to "solve it" by adding code to submit to the mig device UUID. Am I missing something in the slurm/gres/cgroups configuration that would automatically assign correct CUDA_VISIBLE_DEVICES MIG UUID to the job? If I submit 10 jobs requesting --gres=gpu:H100_1g.10gb:1 8 jobs start running: #1 - CUDA_VISIBLE_DEVICES=0 - jobs 1-4 run on one of the mig devices configured on GPU-device ... #8 - CUDA_VISIBLE_DEVICES=7 - jobs 5-8 fail with the message that there are no available CUDA_VISIBLE_DEVICES ... #10 - CUDA_VISIBLE_DEVICES=? - after waiting for free mig resources jobs 9-10 runs or fails depending if GPU-device is free (migs are free) $ nvidia-smi --query-gpu=uuid --id=0 --format=csv,noheader GPU-6c7f34eb-7b0b-9b09-4a2a-a009fb1ba6d3 $ nvidia-smi --query-gpu=uuid --id=7 --format=csv,noheader No devices were found The slurm assigns CUDA_VISIBLE_DEVICES numbers based on the requested resources: Request: gpu:H100_1g.10gb:1 - CUDA_VISIBLE_DEVICES=0 gpu:H100_1g.10gb:2 - CUDA_VISIBLE_DEVICES=0,1 ... Any job with H100_1g.10gb request is going to get device number between 0 and 7 gpu:H100_1g.10gb:8 - CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 gpu:H100_2g.20gb:1 - CUDA_VISIBLE_DEVICES=8 Any job with H100_2g.20gb request is going to get device number between 8 and 10 ... gpu:H100_2g.10gb:3 - CUDA_VISIBLE_DEVICES=8,9,10 gpu:H100_3g.40gb:1 - CUDA_VISIBLE_DEVICES=11 gpu:H100_3g.40gb:2 - CUDA_VISIBLE_DEVICES=12 gpu:H100_7g.80gb:1 - CUDA_VISIBLE_DEVICES=13 I was able to "solve it" by adding this code to my slurm script: # Get the CUDA index using srun and awk CUDA_INDEX=$(srun env | grep CUDA_VISIBLE_DEVICES | awk -F '=' '{print $2}') ## GPU0 1g.10gb ## MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a ## MIG-d1679b4c-c8e7-5d78-bd63-07334ecad664 ## MIG-d3ab0675-d318-5e53-b487-b50695cf2e00 ## MIG-c1384014-7937-5942-ba2a-900bd8a4c4b0 ## MIG-85dd76da-c994-5830-adf0-467c66ae1b95 ## MIG-29a6d43b-882e-5b79-868a-15bb2c770b82 ## MIG-558debd2-dc13-5406-9256-73ef4f279737 ## GPU1 2g.20gb ## MIG-a78cd536-25b7-53cb-941a-a2db5eb0375f ## MIG-0a804c0c-aa27-5993-97cb-cedb854735ce ## MIG-8af9af6d-8720-5763-81e2-83afc43eb42b ## GPU1 1g.10gb ## MIG-68d47441-f5fb-5a3b-ab5b-a0f3449c5b2a ## GPU2 3g.40gb ## MIG-2ed972aa-6640-5116-9271-3d61c0dc9f3b ## MIG-3bdad788-60a5-5bef-85b9-fbdb36dc71fc ## GPU3 7g.80gb ## MIG-ec043869-9176-577e-bac0-46c8411e4e37 # Define the list of UUIDs UUIDS=( "MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a" "MIG-d1679b4c-c8e7-5d78-bd63-07334ecad664" "MIG-d3ab0675-d318-5e53-b487-b50695cf2e00" "MIG-c1384014-7937-5942-ba2a-900bd8a4c4b0" "MIG-85dd76da-c994-5830-adf0-467c66ae1b95" "MIG-29a6d43b-882e-5b79-868a-15bb2c770b82" "MIG-558debd2-dc13-5406-9256-73ef4f279737" "MIG-68d47441-f5fb-5a3b-ab5b-a0f3449c5b2a" "MIG-a78cd536-25b7-53cb-941a-a2db5eb0375f" "MIG-0a804c0c-aa27-5993-97cb-cedb854735ce" "MIG-8af9af6d-8720-5763-81e2-83afc43eb42b" "MIG-2ed972aa-6640-5116-9271-3d61c0dc9f3b" "MIG-3bdad788-60a5-5bef-85b9-fbdb36dc71fc" "MIG-ec043869-9176-577e-bac0-46c8411e4e37" ) # Assign the UUID based on the CUDA index SELECTED_UUID=${UUIDS[$CUDA_INDEX]} # Print the selected UUID echo $SELECTED_UUID export CUDA_VISIBLE_DEVICES=$SELECTED_UUID echo "CUDA_VISIBLE_DEVICES set to: "$SELECTED_UUID echo "Test: "$CUDA_VISIBLE_DEVICES CUDA_VISIBLE_DEVICES=$SELECTED_UUID python3.8 gpu_script.py Result: Script start MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a CUDA_VISIBLE_DEVICES set to: MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a Matrix calculation on CUDA device completed successfully. Script end +---------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=======================================================================================| | 0 7 0 28686 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 8 0 28744 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 9 0 28300 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 10 0 28512 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 11 0 28506 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 12 0 28516 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 13 0 28511 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 1 9 0 28552 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | +---------------------------------------------------------------------------------------+ This is the current setup: RockyLinux 9.2 slurm-22.05.9-1.el9.x86_64 CUDA Version: 12.1 Driver Version: 530.30.02 # nvidia-smi -L GPU 0: NVIDIA H100 80GB HBM3 (UUID: GPU-6c7f34eb-7b0b-9b09-4a2a-a009fb1ba6d3) MIG 1g.10gb Device 0: (UUID: MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a) MIG 1g.10gb Device 1: (UUID: MIG-d1679b4c-c8e7-5d78-bd63-07334ecad664) MIG 1g.10gb Device 2: (UUID: MIG-d3ab0675-d318-5e53-b487-b50695cf2e00) MIG 1g.10gb Device 3: (UUID: MIG-c1384014-7937-5942-ba2a-900bd8a4c4b0) MIG 1g.10gb Device 4: (UUID: MIG-85dd76da-c994-5830-adf0-467c66ae1b95) MIG 1g.10gb Device 5: (UUID: MIG-29a6d43b-882e-5b79-868a-15bb2c770b82) MIG 1g.10gb Device 6: (UUID: MIG-558debd2-dc13-5406-9256-73ef4f279737) GPU 1: NVIDIA H100 80GB HBM3 (UUID: GPU-1439a39c-948b-0657-98e9-aff8595a8729) MIG 2g.20gb Device 0: (UUID: MIG-a78cd536-25b7-53cb-941a-a2db5eb0375f) MIG 2g.20gb Device 1: (UUID: MIG-0a804c0c-aa27-5993-97cb-cedb854735ce) MIG 2g.20gb Device 2: (UUID: MIG-8af9af6d-8720-5763-81e2-83afc43eb42b) MIG 1g.10gb Device 3: (UUID: MIG-68d47441-f5fb-5a3b-ab5b-a0f3449c5b2a) GPU 2: NVIDIA H100 80GB HBM3 (UUID: GPU-3b182fed-bfdf-bc11-d1bd-c41fd1468d2a) MIG 3g.40gb Device 0: (UUID: MIG-2ed972aa-6640-5116-9271-3d61c0dc9f3b) MIG 3g.40gb Device 1: (UUID: MIG-3bdad788-60a5-5bef-85b9-fbdb36dc71fc) GPU 3: NVIDIA H100 80GB HBM3 (UUID: GPU-7eba9d78-e908-3db8-2633-a269aeec395e) MIG 7g.80gb Device 0: (UUID: MIG-ec043869-9176-577e-bac0-46c8411e4e37) $ scontrol show node node16 NodeName=node16 Arch=x86_64 CoresPerSocket=1 CPUAlloc=1 CPUEfctv=112 CPUTot=112 CPULoad=0.00 AvailableFeatures=(null) ActiveFeatures=(null) Gres=gpu:H100_1g.10gb:8,gpu:H100_2g.20gb:3,gpu:H100_3g.40gb:2,gpu:H100_7g.80gb:1 NodeAddr=node16 NodeHostName=node16 Version=22.05.9 OS=Linux 5.14.0-284.30.1.el9_2.x86_64 #1 SMP PREEMPT_DYNAMIC Sat Sep 16 09:55:41 UTC 2023 RealMemory=1030000 AllocMem=0 FreeMem=1011801 Sockets=112 Boards=1 State=MIXED ThreadsPerCore=1 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A Partitions=gpu_H100 BootTime=2024-01-18T13:52:27 SlurmdStartTime=2024-01-19T11:25:44 LastBusyTime=2024-01-19T10:59:13 CfgTRES=cpu=112,mem=1030000M,billing=112,gres/gpu=14 AllocTRES=cpu=1 CapWatts=n/a CurrentWatts=0 AveWatts=0 ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s $ scontrol show partition gpu_H100 PartitionName=gpu_H100 AllowGroups=ALL AllowAccounts=ALL AllowQos=ALL AllocNodes=ALL Default=NO QoS=N/A DefaultTime=NONE DisableRootJobs=NO ExclusiveUser=NO GraceTime=0 Hidden=NO MaxNodes=UNLIMITED MaxTime=UNLIMITED MinNodes=0 LLN=NO MaxCPUsPerNode=UNLIMITED Nodes=node16 PriorityJobFactor=1 PriorityTier=1 RootOnly=NO ReqResv=NO OverSubscribe=NO OverTimeLimit=NONE PreemptMode=OFF State=UP TotalCPUs=112 TotalNodes=1 SelectTypeParameters=NONE JobDefaults=(null) DefMemPerNode=UNLIMITED MaxMemPerNode=UNLIMITED TRES=cpu=112,mem=1030000M,node=1,billing=112,gres/gpu=14 slurm.conf: AccountingStorageTRES=gres/gpu,gres/gpu:A100,gres/gpu:H100,gres/gpu:H100_1g.10gb,gres/gpu:H100_2g.20gb,gres/gpu:H100_3g.40gb,gres/gpu:H100_7g.80gb GresTypes=gpu ... NodeName=node16 CPUs=112 RealMemory=1030000 Gres=gpu:H100_1g.10gb:8,gpu:H100_2g.20gb:3,gpu:H100_3g.40gb:2,gpu:H100_7g.80gb:1 PartitionName=gpu_H100 Nodes=node16 Default=NO MaxTime=INFINITE State=UP gres.conf #AutoDetect=nvml # GPU 0 MIG 0 /proc/driver/nvidia/capabilities/gpu0/mig/gi7/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap66,/dev/nvidia-caps/nvidia-cap67 CPUs=0-55 # GPU 0 MIG 1 /proc/driver/nvidia/capabilities/gpu0/mig/gi8/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap75,/dev/nvidia-caps/nvidia-cap76 CPUs=0-55 # GPU 0 MIG 2 /proc/driver/nvidia/capabilities/gpu0/mig/gi9/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap84,/dev/nvidia-caps/nvidia-cap85 CPUs=0-55 # GPU 0 MIG 3 /proc/driver/nvidia/capabilities/gpu0/mig/gi10/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap93,/dev/nvidia-caps/nvidia-cap94 CPUs=0-55 # GPU 0 MIG 4 /proc/driver/nvidia/capabilities/gpu0/mig/gi11/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap102,/dev/nvidia-caps/nvidia-cap103 CPUs=0-55 # GPU 0 MIG 5 /proc/driver/nvidia/capabilities/gpu0/mig/gi12/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap111,/dev/nvidia-caps/nvidia-cap112 CPUs=0-55 # GPU 0 MIG 6 /proc/driver/nvidia/capabilities/gpu0/mig/gi13/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap120,/dev/nvidia-caps/nvidia-cap121 CPUs=0-55 # GPU 1 MIG 0 /proc/driver/nvidia/capabilities/gpu1/mig/gi3/access NodeName=node16 Name=gpu Type=H100_2g.20gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap165,/dev/nvidia-caps/nvidia-cap166 CPUs=0-55 # GPU 1 MIG 1 /proc/driver/nvidia/capabilities/gpu1/mig/gi5/access NodeName=node16 Name=gpu Type=H100_2g.20gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap183,/dev/nvidia-caps/nvidia-cap184 CPUs=0-55 # GPU 1 MIG 2 /proc/driver/nvidia/capabilities/gpu1/mig/gi6/access NodeName=node16 Name=gpu Type=H100_2g.20gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap192,/dev/nvidia-caps/nvidia-cap193 CPUs=0-55 # GPU 1 MIG 3 /proc/driver/nvidia/capabilities/gpu1/mig/gi9/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap219,/dev/nvidia-caps/nvidia-cap220 CPUs=0-55 # GPU 2 MIG 0 /proc/driver/nvidia/capabilities/gpu2/mig/gi1/access NodeName=node16 Name=gpu Type=H100_3g.40gb MultipleFiles=/dev/nvidia2,/dev/nvidia-caps/nvidia-cap282,/dev/nvidia-caps/nvidia-cap283 CPUs=0-55 # GPU 2 MIG 1 /proc/driver/nvidia/capabilities/gpu2/mig/gi2/access NodeName=node16 Name=gpu Type=H100_3g.40gb MultipleFiles=/dev/nvidia2,/dev/nvidia-caps/nvidia-cap291,/dev/nvidia-caps/nvidia-cap292 CPUs=0-55 # GPU 3 MIG 0 /proc/driver/nvidia/capabilities/gpu3/mig/gi0/access NodeName=node16 Name=gpu Type=H100_7g.80gb MultipleFiles=/dev/nvidia3,/dev/nvidia-caps/nvidia-cap408,/dev/nvidia-caps/nvidia-cap409 CPUs=0-55 [2024-01-19T14:31:13.701] debug: gres/gpu: init: loaded [2024-01-19T14:31:13.701] debug: gpu/generic: init: init: GPU Generic plugin loaded [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_2g.20gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_2g.20gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_2g.20gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_3g.40gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_3g.40gb Count=1 [2024-01-19T14:31:13.702] Gres Name=gpu Type=H100_7g.80gb Count=1 Best, Drazen Jalsovec

1 0

Is there a way to map Azure AD users to NIS?
by Ben Wellborn 20 Jan '24

20 Jan '24

In the past, we would connect AD to NIS users using Centrify and adnisd, but I don't think adnisd can connect to Azure AD (HCI stack). Thanks, Ben

1 0

Re: [slurm-users] Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres)
by Jason Macklin 19 Jan '24

19 Jan '24

If you run "scontrol show jobid <jobid>" of your pending job with the "(Resources)" tag you may see more about what is unavailable to your job. Slurm default configs can cause an entire compute node of resources to be "allocated" to a running job regardless of whether it needs all of them or not so you may need to alter one or both of the following settings to allow more than one job to run on a single node at once. You'll find these in your slurm.conf. Don't forget to "scontrol reconf" and even potentially restart both "slurmctld" & "slurmd" on your nodes if you do end up making changes. SelectType=select/cons_tres SelectTypeParameters=CR_Core_Memory I hope this helps. Kind regards, Jason ---- Jason Macklin Manager Cyberinfrastructure, Research Cyberinfrastructure 860.837.2142 t | 860.202.7779 m jason.macklin(a)jax.org The Jackson Laboratory Maine | Connecticut | California | Shanghai www.jax.org<http://www.jax.org> The Jackson Laboratory: Leading the search for tomorrow's cures ________________________________ From: slurm-users <slurm-users-bounces(a)lists.schedmd.com> on behalf of slurm-users-request(a)lists.schedmd.com <slurm-users-request(a)lists.schedmd.com> Sent: Friday, January 19, 2024 9:24 AM To: slurm-users(a)lists.schedmd.com <slurm-users(a)lists.schedmd.com> Subject: [EXTERNAL]slurm-users Digest, Vol 75, Issue 31 Send slurm-users mailing list submissions to slurm-users(a)lists.schedmd.com To subscribe or unsubscribe via the World Wide Web, visit https://lists.schedmd.com/cgi-bin/mailman/listinfo/slurm-users or, via email, send a message with subject or body 'help' to slurm-users-request(a)lists.schedmd.com You can reach the person managing the list at slurm-users-owner(a)lists.schedmd.com When replying, please edit your Subject line so it is more specific than "Re: Contents of slurm-users digest..." Today's Topics: 1. Re: Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres) (Marko Markoc) 2. Re: Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres) (?mit Seren) ---------------------------------------------------------------------- Message: 1 Date: Fri, 19 Jan 2024 06:12:24 -0800 From: Marko Markoc <mmarkoc(a)pdx.edu> To: Slurm User Community List <slurm-users(a)lists.schedmd.com> Subject: Re: [slurm-users] Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres) Message-ID: <CABnuMe4JTA0e6=VbO8D+To=8FGO+3Byv1dK_MC+OuRitzN5dXg(a)mail.gmail.com> Content-Type: text/plain; charset="utf-8" +1 on checking the memory allocation. Or add/check if you have any DefMemPerX set in your slurm.conf On Fri, Jan 19, 2024 at 12:33?AM mohammed shambakey <shambakey1(a)gmail.com> wrote: > Hi > > I'm not an expert, but is it possible that the currently running jobs is > consuming the whole node because it is allocated the whole memory of the > node (so the other 2 jobs had to wait until it finishes)? > Maybe if you try to restrict the required memory for each job? > > Regards > > On Thu, Jan 18, 2024 at 4:46?PM ?mit Seren <uemit.seren(a)gmail.com> wrote: > >> This line also has tobe changed: >> >> >> #SBATCH --gpus-per-node=4 ? #SBATCH --gpus-per-node=1 >> >> --gpus-per-node seems to be the new parameter that is replacing the --gres= >> one, so you can remove the ?gres line completely. >> >> >> >> Best >> >> ?mit >> >> >> >> *From: *slurm-users <slurm-users-bounces(a)lists.schedmd.com> on behalf of >> Kherfani, Hafedh (Professional Services, TC) <hafedh.kherfani(a)hpe.com> >> *Date: *Thursday, 18. January 2024 at 15:40 >> *To: *Slurm User Community List <slurm-users(a)lists.schedmd.com> >> *Subject: *Re: [slurm-users] Need help with running multiple >> instances/executions of a batch script in parallel (with NVIDIA HGX A100 >> GPU as a Gres) >> >> Hi Noam and Matthias, >> >> >> >> Thanks both for your answers. >> >> >> >> I changed the ?#SBATCH --gres=gpu:4? directive (in the batch script) with >> ?#SBATCH --gres=gpu:1? as you suggested, but it didn?t make a difference, >> as running this batch script 3 times will result in the first job to be in >> a running state, while the second and third jobs will still be in a pending >> state ? >> >> >> >> [slurmtest@c-a100-master test-batch-scripts]$ cat gpu-job.sh >> >> #!/bin/bash >> >> #SBATCH --job-name=gpu-job >> >> #SBATCH --partition=gpu >> >> #SBATCH --nodes=1 >> >> #SBATCH --gpus-per-node=4 >> >> #SBATCH --gres=gpu:1 # <<<< Changed from ?4? >> to ?1? >> >> #SBATCH --tasks-per-node=1 >> >> #SBATCH --output=gpu_job_output.%j >> >> #SBATCH --error=gpu_job_error.%j >> >> >> >> hostname >> >> date >> >> sleep 40 >> >> pwd >> >> >> >> [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh >> >> Submitted batch job *217* >> >> [slurmtest@c-a100-master test-batch-scripts]$ squeue >> >> JOBID PARTITION NAME USER ST TIME NODES >> NODELIST(REASON) >> >> 217 gpu gpu-job slurmtes R 0:02 1 >> c-a100-cn01 >> >> [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh >> >> Submitted batch job *218* >> >> [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh >> >> Submitted batch job *219* >> >> [slurmtest@c-a100-master test-batch-scripts]$ squeue >> >> JOBID PARTITION NAME USER ST TIME NODES >> NODELIST(REASON) >> >> 219 gpu gpu-job slurmtes *PD* 0:00 1 >> (Priority) >> >> 218 gpu gpu-job slurmtes *PD* 0:00 1 >> (Resources) >> >> 217 gpu gpu-job slurmtes *R* 0:07 1 >> c-a100-cn01 >> >> >> >> Basically I?m seeking for some help/hints on how to tell Slurm, from the >> batch script for example: ?I want only 1 or 2 GPUs to be used/consumed by >> the job?, and then I run the batch script/job a couple of times with sbatch >> command, and confirm that we can indeed have multiple jobs using a GPU and >> running in parallel, at the same time. >> >> >> >> Makes sense ? >> >> >> >> >> >> Best regards, >> >> >> >> *Hafedh * >> >> >> >> *From:* slurm-users <slurm-users-bounces(a)lists.schedmd.com> *On Behalf >> Of *Bernstein, Noam CIV USN NRL (6393) Washington DC (USA) >> *Sent:* jeudi 18 janvier 2024 2:30 PM >> *To:* Slurm User Community List <slurm-users(a)lists.schedmd.com> >> *Subject:* Re: [slurm-users] Need help with running multiple >> instances/executions of a batch script in parallel (with NVIDIA HGX A100 >> GPU as a Gres) >> >> >> >> On Jan 18, 2024, at 7:31 AM, Matthias Loose <m.loose(a)mindcode.de> wrote: >> >> >> >> Hi Hafedh, >> >> Im no expert in the GPU side of SLURM, but looking at you current >> configuration to me its working as intended at the moment. You have defined >> 4 GPUs and start multiple jobs each consuming 4 GPUs each. So the jobs wait >> for the ressource the be free again. >> >> I think what you need to look into is the MPS plugin, which seems to do >> what you are trying to achieve: >> https://slurm.schedmd.com/gres.html#MPS_Management >> >> >> >> I agree with the first paragraph. How many GPUs are you expecting each >> job to use? I'd have assumed, based on the original text, that each job is >> supposed to use 1 GPU, and the 4 jobs were supposed to be running >> side-by-side on the one node you have (with 4 GPUs). If so, you need to >> tell each job to request only 1 GPU, and currently each one is requesting 4. >> >> >> >> If your jobs are actually supposed to be using 4 GPUs each, I still don't >> see any advantage to MPS (at least in what is my usual GPU usage pattern): >> all the jobs will take longer to finish, because they are sharing the fixed >> resource. If they take turns, at least the first ones finish as fast as >> they can, and the last one will finish no later than it would have if they >> were all time-sharing the GPUs. I guess NVIDIA had something in mind when >> they developed MPS, so I guess our pattern may not be typical (or at least >> not universal), and in that case the MPS plugin may well be what you need. >> > > > -- > Mohammed >

1 0

slurmctld/slurmdbd (code=exited, status=217/USER)
by Miriam Olmi 19 Jan '24

19 Jan '24

Hi all, I am having some issue with the new version of slurm 23.11.0-1. I had already installed and configured slurm 23.02.3-1 on my cluster and all the services were active and running properly. After I install with the same procedure the new version of slurm I have that the slurmctld and slurmdbd daemons fail to start all with the same error: (code=exited, status=217/USER) And investigating the problem with the command journalctl -xe I find: slurmctld.service: Failed to determine user credentials: No such process slurmctld.service: Failed at step USER spawning /usr/sbin/slurmctld: No such process I had a look at the slurmctld.service file for both the slurm versions and I found the following differences in the [Service] section. >From the slurmctld.service file of slurm 23.02.3-1: [Service] Type=simple EnvironmentFile=-/etc/sysconfig/slurmctld EnvironmentFile=-/etc/default/slurmctld ExecStart=/usr/sbin/slurmctld -D -s $SLURMCTLD_OPTIONS ExecReload=/bin/kill -HUP $MAINPID LimitNOFILE=65536 TasksMax=infinity >From the slurmctld.service file of slurm 23.11.0-1: [Service] Type=notify EnvironmentFile=-/etc/sysconfig/slurmctld EnvironmentFile=-/etc/default/slurmctld User=slurm Group=slurm ExecStart=/usr/sbin/slurmctld --systemd $SLURMCTLD_OPTIONS ExecReload=/bin/kill -HUP $MAINPID LimitNOFILE=65536 TasksMax=infinity I think the presence of the new lines regarding the slurm user might be the problem but I am not sure and I have no idea how to solve it. Can anyone halp me? Thanks in advance, Miriam

2 1

Jobs exiting together
by Alexander Silva 19 Jan '24

19 Jan '24

Recently, i have built an hpc cluster with slurm as workload. The test jobs with quatum chemistry codes have worked fine. However, production jobs with lammps have shown an unexpected behavior when the first job completed, normally or not, cause the termination of the others in the same compute node. Initially, I thought that was due to mpi malfunction, but this behavior is algo observed for serial lammps code. The lammps group said to me that behavior could be generated by slurm. My question to you is about what parameter in slurm.conf could be responsible for the termination of the other jobs. I am using an epilogue script that work normally in another cluster. Thanks.

1 0

Running slurm job on requested nvidia mig device
by Dražen Jalšovec 18 Jan '24

18 Jan '24

Hi, We are testing the MIG deployment on our new slurm compute node with 4 x H100 GPUs. It looks like everything is configured correctly but we have a problem accessing mig devices. When I submit jobs requesting a mig gpu device #SBATCH --gres=gpu:H100_1g.10gb:1, the jobs get submitted to the node, but only 4 jobs get executed and all other jobs fail. I was able to "solve it" by adding code to submit to the mig device UUID. Am I missing something in the slurm/gres/cgroups configuration that would automatically assign correct CUDA_VISIBLE_DEVICES MIG UUID to the job? If I submit 10 jobs requesting --gres=gpu:H100_1g.10gb:1 8 jobs start running: #1 - CUDA_VISIBLE_DEVICES=0 - jobs 1-4 run on one of the mig devices configured on GPU-device ... #8 - CUDA_VISIBLE_DEVICES=7 - jobs 5-8 fail with the message that there are no available CUDA_VISIBLE_DEVICES ... #10 - CUDA_VISIBLE_DEVICES=? - after waiting for free mig resources jobs 9-10 runs or fails depending if GPU-device is free (migs are free) $ nvidia-smi --query-gpu=uuid --id=0 --format=csv,noheader GPU-6c7f34eb-7b0b-9b09-4a2a-a009fb1ba6d3 $ nvidia-smi --query-gpu=uuid --id=7 --format=csv,noheader No devices were found The slurm assigns CUDA_VISIBLE_DEVICES numbers based on the requested resources: Request: gpu:H100_1g.10gb:1 - CUDA_VISIBLE_DEVICES=0 gpu:H100_1g.10gb:2 - CUDA_VISIBLE_DEVICES=0,1 ... Any job with H100_1g.10gb request is going to get device number between 0 and 7 gpu:H100_1g.10gb:8 - CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 gpu:H100_2g.20gb:1 - CUDA_VISIBLE_DEVICES=8 Any job with H100_2g.20gb request is going to get device number between 8 and 10 ... gpu:H100_2g.10gb:3 - CUDA_VISIBLE_DEVICES=8,9,10 gpu:H100_3g.40gb:1 - CUDA_VISIBLE_DEVICES=11 gpu:H100_3g.40gb:2 - CUDA_VISIBLE_DEVICES=12 gpu:H100_7g.80gb:1 - CUDA_VISIBLE_DEVICES=13 I was able to "solve it" by adding this code to my slurm script: # Get the CUDA index using srun and awk CUDA_INDEX=$(srun env | grep CUDA_VISIBLE_DEVICES | awk -F '=' '{print $2}') ## GPU0 1g.10gb ## MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a ## MIG-d1679b4c-c8e7-5d78-bd63-07334ecad664 ## MIG-d3ab0675-d318-5e53-b487-b50695cf2e00 ## MIG-c1384014-7937-5942-ba2a-900bd8a4c4b0 ## MIG-85dd76da-c994-5830-adf0-467c66ae1b95 ## MIG-29a6d43b-882e-5b79-868a-15bb2c770b82 ## MIG-558debd2-dc13-5406-9256-73ef4f279737 ## GPU1 2g.20gb ## MIG-a78cd536-25b7-53cb-941a-a2db5eb0375f ## MIG-0a804c0c-aa27-5993-97cb-cedb854735ce ## MIG-8af9af6d-8720-5763-81e2-83afc43eb42b ## GPU1 1g.10gb ## MIG-68d47441-f5fb-5a3b-ab5b-a0f3449c5b2a ## GPU2 3g.40gb ## MIG-2ed972aa-6640-5116-9271-3d61c0dc9f3b ## MIG-3bdad788-60a5-5bef-85b9-fbdb36dc71fc ## GPU3 7g.80gb ## MIG-ec043869-9176-577e-bac0-46c8411e4e37 # Define the list of UUIDs UUIDS=( "MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a" "MIG-d1679b4c-c8e7-5d78-bd63-07334ecad664" "MIG-d3ab0675-d318-5e53-b487-b50695cf2e00" "MIG-c1384014-7937-5942-ba2a-900bd8a4c4b0" "MIG-85dd76da-c994-5830-adf0-467c66ae1b95" "MIG-29a6d43b-882e-5b79-868a-15bb2c770b82" "MIG-558debd2-dc13-5406-9256-73ef4f279737" "MIG-68d47441-f5fb-5a3b-ab5b-a0f3449c5b2a" "MIG-a78cd536-25b7-53cb-941a-a2db5eb0375f" "MIG-0a804c0c-aa27-5993-97cb-cedb854735ce" "MIG-8af9af6d-8720-5763-81e2-83afc43eb42b" "MIG-2ed972aa-6640-5116-9271-3d61c0dc9f3b" "MIG-3bdad788-60a5-5bef-85b9-fbdb36dc71fc" "MIG-ec043869-9176-577e-bac0-46c8411e4e37" ) # Assign the UUID based on the CUDA index SELECTED_UUID=${UUIDS[$CUDA_INDEX]} # Print the selected UUID echo $SELECTED_UUID export CUDA_VISIBLE_DEVICES=$SELECTED_UUID echo "CUDA_VISIBLE_DEVICES set to: "$SELECTED_UUID echo "Test: "$CUDA_VISIBLE_DEVICES CUDA_VISIBLE_DEVICES=$SELECTED_UUID python3.8 gpu_script.py Result: Script start MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a CUDA_VISIBLE_DEVICES set to: MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a Matrix calculation on CUDA device completed successfully. Script end +---------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=======================================================================================| | 0 7 0 28686 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 8 0 28744 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 9 0 28300 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 10 0 28512 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 11 0 28506 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 12 0 28516 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 0 13 0 28511 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | | 1 9 0 28552 C ...gpu-cuda12.1-python38/bin/python3.8 866MiB | +---------------------------------------------------------------------------------------+ This is the current setup: RockyLinux 9.2 slurm-22.05.9-1.el9.x86_64 CUDA Version: 12.1 Driver Version: 530.30.02 # nvidia-smi -L GPU 0: NVIDIA H100 80GB HBM3 (UUID: GPU-6c7f34eb-7b0b-9b09-4a2a-a009fb1ba6d3) MIG 1g.10gb Device 0: (UUID: MIG-fbb82c87-989b-5e2c-98e5-a57919c5cb1a) MIG 1g.10gb Device 1: (UUID: MIG-d1679b4c-c8e7-5d78-bd63-07334ecad664) MIG 1g.10gb Device 2: (UUID: MIG-d3ab0675-d318-5e53-b487-b50695cf2e00) MIG 1g.10gb Device 3: (UUID: MIG-c1384014-7937-5942-ba2a-900bd8a4c4b0) MIG 1g.10gb Device 4: (UUID: MIG-85dd76da-c994-5830-adf0-467c66ae1b95) MIG 1g.10gb Device 5: (UUID: MIG-29a6d43b-882e-5b79-868a-15bb2c770b82) MIG 1g.10gb Device 6: (UUID: MIG-558debd2-dc13-5406-9256-73ef4f279737) GPU 1: NVIDIA H100 80GB HBM3 (UUID: GPU-1439a39c-948b-0657-98e9-aff8595a8729) MIG 2g.20gb Device 0: (UUID: MIG-a78cd536-25b7-53cb-941a-a2db5eb0375f) MIG 2g.20gb Device 1: (UUID: MIG-0a804c0c-aa27-5993-97cb-cedb854735ce) MIG 2g.20gb Device 2: (UUID: MIG-8af9af6d-8720-5763-81e2-83afc43eb42b) MIG 1g.10gb Device 3: (UUID: MIG-68d47441-f5fb-5a3b-ab5b-a0f3449c5b2a) GPU 2: NVIDIA H100 80GB HBM3 (UUID: GPU-3b182fed-bfdf-bc11-d1bd-c41fd1468d2a) MIG 3g.40gb Device 0: (UUID: MIG-2ed972aa-6640-5116-9271-3d61c0dc9f3b) MIG 3g.40gb Device 1: (UUID: MIG-3bdad788-60a5-5bef-85b9-fbdb36dc71fc) GPU 3: NVIDIA H100 80GB HBM3 (UUID: GPU-7eba9d78-e908-3db8-2633-a269aeec395e) MIG 7g.80gb Device 0: (UUID: MIG-ec043869-9176-577e-bac0-46c8411e4e37) $ scontrol show node node16 NodeName=node16 Arch=x86_64 CoresPerSocket=1 CPUAlloc=1 CPUEfctv=112 CPUTot=112 CPULoad=0.00 AvailableFeatures=(null) ActiveFeatures=(null) Gres=gpu:H100_1g.10gb:8,gpu:H100_2g.20gb:3,gpu:H100_3g.40gb:2,gpu:H100_7g.80gb:1 NodeAddr=node16 NodeHostName=node16 Version=22.05.9 OS=Linux 5.14.0-284.30.1.el9_2.x86_64 #1 SMP PREEMPT_DYNAMIC Sat Sep 16 09:55:41 UTC 2023 RealMemory=1030000 AllocMem=0 FreeMem=1011801 Sockets=112 Boards=1 State=MIXED ThreadsPerCore=1 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A Partitions=gpu_H100 BootTime=2024-01-18T13:52:27 SlurmdStartTime=2024-01-19T11:25:44 LastBusyTime=2024-01-19T10:59:13 CfgTRES=cpu=112,mem=1030000M,billing=112,gres/gpu=14 AllocTRES=cpu=1 CapWatts=n/a CurrentWatts=0 AveWatts=0 ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s $ scontrol show partition gpu_H100 PartitionName=gpu_H100 AllowGroups=ALL AllowAccounts=ALL AllowQos=ALL AllocNodes=ALL Default=NO QoS=N/A DefaultTime=NONE DisableRootJobs=NO ExclusiveUser=NO GraceTime=0 Hidden=NO MaxNodes=UNLIMITED MaxTime=UNLIMITED MinNodes=0 LLN=NO MaxCPUsPerNode=UNLIMITED Nodes=node16 PriorityJobFactor=1 PriorityTier=1 RootOnly=NO ReqResv=NO OverSubscribe=NO OverTimeLimit=NONE PreemptMode=OFF State=UP TotalCPUs=112 TotalNodes=1 SelectTypeParameters=NONE JobDefaults=(null) DefMemPerNode=UNLIMITED MaxMemPerNode=UNLIMITED TRES=cpu=112,mem=1030000M,node=1,billing=112,gres/gpu=14 slurm.conf: AccountingStorageTRES=gres/gpu,gres/gpu:A100,gres/gpu:H100,gres/gpu:H100_1g.10gb,gres/gpu:H100_2g.20gb,gres/gpu:H100_3g.40gb,gres/gpu:H100_7g.80gb GresTypes=gpu ... NodeName=node16 CPUs=112 RealMemory=1030000 Gres=gpu:H100_1g.10gb:8,gpu:H100_2g.20gb:3,gpu:H100_3g.40gb:2,gpu:H100_7g.80gb:1 PartitionName=gpu_H100 Nodes=node16 Default=NO MaxTime=INFINITE State=UP gres.conf #AutoDetect=nvml # GPU 0 MIG 0 /proc/driver/nvidia/capabilities/gpu0/mig/gi7/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap66,/dev/nvidia-caps/nvidia-cap67 CPUs=0-55 # GPU 0 MIG 1 /proc/driver/nvidia/capabilities/gpu0/mig/gi8/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap75,/dev/nvidia-caps/nvidia-cap76 CPUs=0-55 # GPU 0 MIG 2 /proc/driver/nvidia/capabilities/gpu0/mig/gi9/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap84,/dev/nvidia-caps/nvidia-cap85 CPUs=0-55 # GPU 0 MIG 3 /proc/driver/nvidia/capabilities/gpu0/mig/gi10/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap93,/dev/nvidia-caps/nvidia-cap94 CPUs=0-55 # GPU 0 MIG 4 /proc/driver/nvidia/capabilities/gpu0/mig/gi11/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap102,/dev/nvidia-caps/nvidia-cap103 CPUs=0-55 # GPU 0 MIG 5 /proc/driver/nvidia/capabilities/gpu0/mig/gi12/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap111,/dev/nvidia-caps/nvidia-cap112 CPUs=0-55 # GPU 0 MIG 6 /proc/driver/nvidia/capabilities/gpu0/mig/gi13/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia0,/dev/nvidia-caps/nvidia-cap120,/dev/nvidia-caps/nvidia-cap121 CPUs=0-55 # GPU 1 MIG 0 /proc/driver/nvidia/capabilities/gpu1/mig/gi3/access NodeName=node16 Name=gpu Type=H100_2g.20gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap165,/dev/nvidia-caps/nvidia-cap166 CPUs=0-55 # GPU 1 MIG 1 /proc/driver/nvidia/capabilities/gpu1/mig/gi5/access NodeName=node16 Name=gpu Type=H100_2g.20gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap183,/dev/nvidia-caps/nvidia-cap184 CPUs=0-55 # GPU 1 MIG 2 /proc/driver/nvidia/capabilities/gpu1/mig/gi6/access NodeName=node16 Name=gpu Type=H100_2g.20gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap192,/dev/nvidia-caps/nvidia-cap193 CPUs=0-55 # GPU 1 MIG 3 /proc/driver/nvidia/capabilities/gpu1/mig/gi9/access NodeName=node16 Name=gpu Type=H100_1g.10gb MultipleFiles=/dev/nvidia1,/dev/nvidia-caps/nvidia-cap219,/dev/nvidia-caps/nvidia-cap220 CPUs=0-55 # GPU 2 MIG 0 /proc/driver/nvidia/capabilities/gpu2/mig/gi1/access NodeName=node16 Name=gpu Type=H100_3g.40gb MultipleFiles=/dev/nvidia2,/dev/nvidia-caps/nvidia-cap282,/dev/nvidia-caps/nvidia-cap283 CPUs=0-55 # GPU 2 MIG 1 /proc/driver/nvidia/capabilities/gpu2/mig/gi2/access NodeName=node16 Name=gpu Type=H100_3g.40gb MultipleFiles=/dev/nvidia2,/dev/nvidia-caps/nvidia-cap291,/dev/nvidia-caps/nvidia-cap292 CPUs=0-55 # GPU 3 MIG 0 /proc/driver/nvidia/capabilities/gpu3/mig/gi0/access NodeName=node16 Name=gpu Type=H100_7g.80gb MultipleFiles=/dev/nvidia3,/dev/nvidia-caps/nvidia-cap408,/dev/nvidia-caps/nvidia-cap409 CPUs=0-55 [2024-01-19T14:31:13.701] debug: gres/gpu: init: loaded [2024-01-19T14:31:13.701] debug: gpu/generic: init: init: GPU Generic plugin loaded [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_1g.10gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_2g.20gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_2g.20gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_2g.20gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_3g.40gb Count=1 [2024-01-19T14:31:13.701] Gres Name=gpu Type=H100_3g.40gb Count=1 [2024-01-19T14:31:13.702] Gres Name=gpu Type=H100_7g.80gb Count=1 Best, Drazen Jalsovec

1 0

Potential Side Effects of larger MessageTimeout value
by Herc Silverstein 18 Jan '24

18 Jan '24

Hi, What are potential bad side effects of using a large/larger MessageTimeout? And is there a value at which this setting is too large (long)? Thanks, Herc

1 0

error
by Felix 18 Jan '24

18 Jan '24

Hello I started a new AMD node, and the error is as follows: "CPU frequency setting not configured for this node" extended looks like this: [2024-01-18T18:28:06.682] CPU frequency setting not configured for this node [2024-01-18T18:28:06.691] slurmd started on Thu, 18 Jan 2024 18:28:06 +0200 [2024-01-18T18:28:06.691] CPUs=128 Boards=1 Sockets=1 Cores=64 Threads=2 Memory=256786 TmpDisk=875797 Uptime=4569 CPUSpecList=(null) FeaturesAvail=(null) FeaturesActive=(null) In the configuration file I have the following: NodeName=awn-1[04] NodeAddr=192.168.4.[111] CPUs=128 RealMemory=256000 Sockets=1 CoresPerSocket=64 ThreadsPerCore=2 Feature=HyperThread Could you please help me? Thank you Felix -- Dr. Eng. Farcas Felix National Institute of Research and Development of Isotopic and Molecular Technology, IT - Department - Cluj-Napoca, Romania Mobile: +40742195323

2 1

Re: [slurm-users] [BULK] slurm-users Digest, Vol 75, Issue 26
by Jason Macklin 18 Jan '24

18 Jan '24

If you run "scontrol show jobid <jobid>" of your pending job with the "(Resources)" tag you may see more about what is unavailable to your job. Slurm default configs can cause an entire compute node of resources to be "allocated" to a running job regardless of whether it needs all of them or not so you may need to alter one or both of the following settings to allow more than one job to run on a single node at once. You'll find these in your slurm.conf. Don't forget to "scontrol reconf" and even potentially restart both "slurmctld" & "slurmd" on your nodes if you do end up making changes. SelectType=select/cons_tres SelectTypeParameters=CR_Core_Memory I hope this helps. Kind regards, Jason ---- Jason Macklin Manager Cyberinfrastructure, Research Cyberinfrastructure 860.837.2142 t | 860.202.7779 m jason.macklin(a)jax.org The Jackson Laboratory Maine | Connecticut | California | Shanghai www.jax.org<http://www.jax.org> The Jackson Laboratory: Leading the search for tomorrow's cures ________________________________ From: slurm-users <slurm-users-bounces(a)lists.schedmd.com> on behalf of slurm-users-request(a)lists.schedmd.com <slurm-users-request(a)lists.schedmd.com> Sent: Thursday, January 18, 2024 9:46 AM To: slurm-users(a)lists.schedmd.com <slurm-users(a)lists.schedmd.com> Subject: [BULK] slurm-users Digest, Vol 75, Issue 26 Send slurm-users mailing list submissions to slurm-users(a)lists.schedmd.com To subscribe or unsubscribe via the World Wide Web, visit https://lists.schedmd.com/cgi-bin/mailman/listinfo/slurm-users or, via email, send a message with subject or body 'help' to slurm-users-request(a)lists.schedmd.com You can reach the person managing the list at slurm-users-owner(a)lists.schedmd.com When replying, please edit your Subject line so it is more specific than "Re: Contents of slurm-users digest..." Today's Topics: 1. Re: Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres) (Baer, Troy) ---------------------------------------------------------------------- Message: 1 Date: Thu, 18 Jan 2024 14:46:48 +0000 From: "Baer, Troy" <troy(a)osc.edu> To: Slurm User Community List <slurm-users(a)lists.schedmd.com> Subject: Re: [slurm-users] Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres) Message-ID: <CH0PR01MB6924127AF471DED69151805BCF712(a)CH0PR01MB6924.prod.exchangelabs.com> Content-Type: text/plain; charset="utf-8" Hi Hafedh, Your job script has the sbatch directive ??gpus-per-node=4? set. I suspect that if you look at what?s allocated to the running job by doing ?scontrol show job <jobid>? and looking at the TRES field, it?s been allocated 4 GPUs instead of one. Regards, --Troy From: slurm-users <slurm-users-bounces(a)lists.schedmd.com> On Behalf Of Kherfani, Hafedh (Professional Services, TC) Sent: Thursday, January 18, 2024 9:38 AM To: Slurm User Community List <slurm-users(a)lists.schedmd.com> Subject: Re: [slurm-users] Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres) Hi Noam and Matthias, Thanks both for your answers. I changed the ?#SBATCH --gres=gpu:?4? directive (in the batch script) with ?#SBATCH --gres=gpu:?1? as you suggested, but it didn?t make a difference, as running Hi Noam and Matthias, Thanks both for your answers. I changed the ?#SBATCH --gres=gpu:4? directive (in the batch script) with ?#SBATCH --gres=gpu:1? as you suggested, but it didn?t make a difference, as running this batch script 3 times will result in the first job to be in a running state, while the second and third jobs will still be in a pending state ? [slurmtest@c-a100-master test-batch-scripts]$ cat gpu-job.sh #!/bin/bash #SBATCH --job-name=gpu-job #SBATCH --partition=gpu #SBATCH --nodes=1 #SBATCH --gpus-per-node=4 #SBATCH --gres=gpu:1 # <<<< Changed from ?4? to ?1? #SBATCH --tasks-per-node=1 #SBATCH --output=gpu_job_output.%j #SBATCH --error=gpu_job_error.%j hostname date sleep 40 pwd [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh Submitted batch job 217 [slurmtest@c-a100-master test-batch-scripts]$ squeue JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 217 gpu gpu-job slurmtes R 0:02 1 c-a100-cn01 [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh Submitted batch job 218 [slurmtest@c-a100-master test-batch-scripts]$ sbatch gpu-job.sh Submitted batch job 219 [slurmtest@c-a100-master test-batch-scripts]$ squeue JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 219 gpu gpu-job slurmtes PD 0:00 1 (Priority) 218 gpu gpu-job slurmtes PD 0:00 1 (Resources) 217 gpu gpu-job slurmtes R 0:07 1 c-a100-cn01 Basically I?m seeking for some help/hints on how to tell Slurm, from the batch script for example: ?I want only 1 or 2 GPUs to be used/consumed by the job?, and then I run the batch script/job a couple of times with sbatch command, and confirm that we can indeed have multiple jobs using a GPU and running in parallel, at the same time. Makes sense ? Best regards, Hafedh From: slurm-users <slurm-users-bounces(a)lists.schedmd.com<mailto:slurm-users-bounces@lists.schedmd.com>> On Behalf Of Bernstein, Noam CIV USN NRL (6393) Washington DC (USA) Sent: jeudi 18 janvier 2024 2:30 PM To: Slurm User Community List <slurm-users(a)lists.schedmd.com<mailto:slurm-users@lists.schedmd.com>> Subject: Re: [slurm-users] Need help with running multiple instances/executions of a batch script in parallel (with NVIDIA HGX A100 GPU as a Gres) On Jan 18, 2024, at 7:31 AM, Matthias Loose <m.loose(a)mindcode.de<mailto:m.loose@mindcode.de>> wrote: Hi Hafedh, Im no expert in the GPU side of SLURM, but looking at you current configuration to me its working as intended at the moment. You have defined 4 GPUs and start multiple jobs each consuming 4 GPUs each. So the jobs wait for the ressource the be free again. I think what you need to look into is the MPS plugin, which seems to do what you are trying to achieve: https://slurm.schedmd.com/gres.html#MPS_Management<https://urldefense.com/v3/__https:/slurm.schedmd.com/gres.html*MPS_Manageme…> I agree with the first paragraph. How many GPUs are you expecting each job to use? I'd have assumed, based on the original text, that each job is supposed to use 1 GPU, and the 4 jobs were supposed to be running side-by-side on the one node you have (with 4 GPUs). If so, you need to tell each job to request only 1 GPU, and currently each one is requesting 4. If your jobs are actually supposed to be using 4 GPUs each, I still don't see any advantage to MPS (at least in what is my usual GPU usage pattern): all the jobs will take longer to finish, because they are sharing the fixed resource. If they take turns, at least the first ones finish as fast as they can, and the last one will finish no later than it would have if they were all time-sharing the GPUs. I guess NVIDIA had something in mind when they developed MPS, so I guess our pattern may not be typical (or at least not universal), and in that case the MPS plugin may well be what you need.

1 0

2024

slurm-users ----- 2024 ----- December 2024 November 2024 October 2024 September 2024 August 2024 July 2024 June 2024 May 2024 April 2024 March 2024 February 2024 January 2024

slurm-users