- slurm-users - lists.schedmd.com

cgroup issue on non-systemd system
by Rafał Lalik 14 Jun '24

14 Jun '24

Hello, per documentation, it is possible to run slurm on non systemd system with IgnoreSystemd=yes in cgroup.conf. However I had an error with slurmd: error: common_file_write_content: unable to open '/sys/fs/cgroup/system.slice/cgroup.subtree_control' for writing: No such file or directory error: Cannot enable cpuset in /sys/fs/cgroup/system.slice/cgroup.subtree_control: No such file or directory error: common_file_write_content: unable to open '/sys/fs/cgroup/system.slice/cgroup.subtree_control' for writing: No such file or directory error: Cannot enable memory in /sys/fs/cgroup/system.slice/cgroup.subtree_control: No such file or directory error: common_file_write_content: unable to open '/sys/fs/cgroup/system.slice/cgroup.subtree_control' for writing: No such file or directory error: Cannot enable cpu in /sys/fs/cgroup/system.slice/cgroup.subtree_control: No such file or directory error: Could not create scope directory /sys/fs/cgroup/system.slice/slurmstepd.scope: No such file or directory error: Couldn't load specified plugin name for cgroup/v2: Plugin init() callback failed error: cannot create cgroup context for cgroup/v2 error: Unable to initialize cgroup plugin error: slurmd initialization failed I could link it to the fact that '/sys/fs/cgroup/system.slice' does not exists when creating '/sys/fs/cgroup/system.slice/slurmstepd.scope' subdir. I propose small patch to the slurm which allows to create the path recursively (aka `mkdir -p` from shell): (also here: https://github.com/rlalik/slurm/tree/recurse_mkdir) diff --git a/src/plugins/cgroup/v2/cgroup_v2.c b/src/plugins/cgroup/v2/cgroup_v2.c index a18b9e62bd..6b7e2a2e4a 100644 --- a/src/plugins/cgroup/v2/cgroup_v2.c +++ b/src/plugins/cgroup/v2/cgroup_v2.c @@ -743,11 +743,40 @@ static int _init_stepd_system_scope(pid_t pid) return SLURM_SUCCESS; } +/* This create path recursively.*/ +bool recurse_mkdir(const char *dirname, int mode) +{ + const char *p = dirname; + char *temp = calloc(1, strlen(dirname)+1); + int ret = 0; + + while ((p = strchr(p, '/')) != NULL) { + /* Skip empty elements. Could be just multiple separators + * which is okay. */ + if (p != dirname && *(p-1) == '/') { + p++; + continue; + } + /* Put the path up to this point into a temporary to + * pass to the make directory function. */ + memcpy(temp, dirname, p-dirname); + temp[p-dirname] = '\0'; + p++; + if ((ret = mkdir(temp, mode)) != 0) { + if (errno != EEXIST) { + break; + } + } + } + free(temp); + return ret; +} + static int _init_new_scope(char *scope_path) { int rc; - rc = mkdir(scope_path, 0755); + rc = recurse_mkdir(scope_path, 0755); if (rc && (errno != EEXIST)) { error("Could not create scope directory %s: %m", scope_path); return SLURM_ERROR;

1 0

Debian RPM build for arm64?
by Christopher Harrop - NOAA Affiliate 14 Jun '24

14 Jun '24

Hello, Are the instructions for building Debian RPMs found at https://slurm.schedmd.com/quickstart_admin.html#debuild expected to work on ARM machines? I am having trouble with the "debuild -b -uc -us” step. #10 29.01 configure: exit 1 #10 29.01 dh_auto_configure: error: cd obj-aarch64-linux-gnu && ../configure --build=aarch64-linux-gnu --prefix=/usr --includedir=\${prefix}/include --mandir=\${prefix}/share/man --infodir=\${prefix}/share/info --sysconfdir=/etc --localstatedir=/var --disable-silent-rules --libdir=\${prefix}/lib/aarch64-linux-gnu --runstatedir=/run --disable-maintainer-mode --disable-dependency-tracking --sysconfdir=/etc/slurm --disable-debug --with-slurmrestd --with-pmix --enable-pam --with-systemdsystemunitdir=/lib/systemd/system/ SUCMD=/bin/su SLEEP_CMD=/bin/sleep returned exit code 1 #10 29.01 make[1]: *** [debian/rules:21: override_dh_auto_configure] Error 25 #10 29.01 make[1]: Leaving directory '/tmp/slurm-23.11.7' #10 29.02 make: *** [debian/rules:6: build] Error 2 #10 29.02 dpkg-buildpackage: error: debian/rules build subprocess returned exit status 2 #10 29.02 debuild: fatal error at line 1182: #10 29.02 dpkg-buildpackage -us -uc -ui -b failed #10 ERROR: process "/bin/sh -c cd /tmp && wget https://download.schedmd.com/slurm/slurm-23.11.7.tar.bz2 && tar -xaf slurm-23.11.7.tar.bz2 && cd slurm-23.11.7 && mk-build-deps -t \"apt-get -o Debug::pkgProblemResolver=yes -y\" -i debian/control && debuild -b -uc -us && cd .. && ARCH=$(dpkg --print-architecture) && dpkg --install slurm-smd_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-client_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-dev_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-doc_23.11.7-1_all.deb && dpkg --install slurm-smd-libnss-slurm_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-libpam-slurm-adopt_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-libpmi0_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-libpmi2-0_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-libslurm-perl_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-sackd_23.11.7-1_${ARCH}.deb && dpkg --install slurm-smd-sview_23.11.7-1_${ARCH}.deb" did not complete successfully: exit code: 29 Chris --------------------------------------------------------------------------------------------------- Christopher W. Harrop voice: (720) 649-0316 NOAA Global Systems Laboratory, R/GSL6 fax: (303) 497-7259 325 Broadway Boulder, CO 80303

3 3

Re: slurmstepd: error: task_g_set_affinity: Operation not permitted
by Christopher Harrop - NOAA Affiliate 14 Jun '24

14 Jun '24

There is a permission problem somewhere, but I don’t know where. If I run as root, it works: admin@slurmfrontend:~$ srun hostname srun: error: task 0 launch failed: Slurmd could not execve job slurmstepd: error: task_g_set_affinity: Operation not permitted slurmstepd: error: _exec_wait_child_wait_for_parent: failed: No error admin@slurmfrontend:~$ sudo srun hostname slurmnode1 admin@slurmfrontend:~$ sudo srun -N 3 hostname slurmnode1 slurmnode3 slurmnode2 admin@slurmfrontend:~$ Chris --------------------------------------------------------------------------------------------------- Christopher W. Harrop voice: (720) 649-0316 NOAA Global Systems Laboratory, R/GSL6 fax: (303) 497-7259 325 Broadway Boulder, CO 80303

2 1

Re: Limit GPU depending on type
by Gestió Servidors 14 Jun '24

14 Jun '24

Hi, because of my real scenario (in mi first post I explained my testing scenario), with several differents users of differents types (researchers, university students and/or teachers, etc), I have distributed my GPUs in 3 differents partitions: * PartitionName=cuda-staff.q Nodes=gpu-[1-4] OverSubscribe=No MaxTime=INFINITE State=UP AllocNodes=node[0-22],node-login,node-login-bak AllowGroups=caos,profesor * PartitionName=cuda-int.q Nodes=gpu-[2,4] OverSubscribe=No MaxTime=30:00 State=UP AllocNodes=node[0-22] * PartitionName=cuda-ext.q Nodes=gpu-[1,3] OverSubscribe=No MaxTime=30:00 State=UP AllocNodes=node[0-22],node-login,node-login-bak Explanation: * In “cuda-staff.q”, only teachers could submit and they can submit from each lab node or each login node. * In “cuda-int.q” everybody can submit, but only from lab nodes. * In “cuda-ex.q” everybody can also submit, but in this case, from lab nodes and login nodes. Since now, I have not used “QoS”... but I’m going to install a new data/user server and I want to reconfigure SLURM. If I distributed GPUs in the way that Gerhard Strangar explains (both similar RTX3080) in a partition with restricted QoS and all other GPUs in other partition, some GPUs that now are restricted for “inside lab user” will be accessible from “outside lab user”. So I think (all teachers want this way ☹ ) I must have these partitions distribution. However, if I apply a QoS limiting only one GPU in each partition. it woulb be possible that a user could user one RTX3080 from outside lab and the other RTX3080 from inside lab… and this is what I want to deny. I will reread documentation. All help will be appreciated, of course!!!! Thanks.

1 0

Slurm version 23.11.8 is now available
by Marshall Garey 13 Jun '24

13 Jun '24

We are pleased to announce the availability of Slurm version 23.11.8. The 23.11.8 release fixes some potential crashes in slurmctld, slurmrestd, and slurmd when using less common features; two issues in auth/slurm; and a few other minor bugs. Slurm can be downloaded from https://www.schedmd.com/downloads.php . -Marshall > -- Fix slurmctld crash when reconfiguring with a PrologSlurmctld is running. > -- Fix slurmctld crash after a job has been resized. > -- Fix slurmctld and slurmdbd potentially stopping instead of performing a > logrotate when recieving SIGUSR2 when using auth/slurm. > -- Fix not having a disabled value for keepalive CommunicationParameters in > slurm.conf when these parameters are not set. This can log an error when > setting a socket, for example during slurmdbd registration with ctld. > -- switch/hpe_slingshot - Fix slurmctld crash when upgrading from 23.02. > -- Fix "Could not find group" errors from validate_group() when using > AllowGroups with large /etc/group files. > -- slurmrestd - Prevent a slurmrestd segfault when parsing the crontab field, > which was never usable. Now it explicitly ignores the value and emits a > warning if it is used for the following endpoints: > 'POST /slurm/v0.0.39/job/{job_id}' > 'POST /slurm/v0.0.39/job/submit' > 'POST /slurm/v0.0.40/job/{job_id}' > 'POST /slurm/v0.0.40/job/submit' > -- Fix getting user environment when using sbatch with "--get-user-env" or > "--export=" when there is a user profile script that reads /proc. > -- Prevent slurmd from crashing if acct_gather_energy/gpu is configured but > GresTypes is not configured. > -- Do not log the following errors when AcctGatherEnergyType plugins are used > but a node does not have or cannot find sensors: > "error: _get_joules_task: can't get info from slurmd" > "error: slurm_get_node_energy: Zero Bytes were transmitted or received" > However, the following error will continue to be logged: > "error: Can't get energy data. No power sensors are available. Try later" > -- Fix cloud nodes not being able to forward to nodes that restarted with new > IP addresses. > -- sacct - Fix printing of job group for job steps. > -- Fix error in scrontab jobs when using slurm.conf:PropagatePrioProcess=1. > -- Fix slurmctld crash on a batch job submission with "--nodes 0,...". > -- Fix dynamic IP address fanout forwarding when using auth/slurm.

1 0

slurmstepd: error: task_g_set_affinity: Operation not permitted
by Christopher Harrop - NOAA Affiliate 13 Jun '24

13 Jun '24

Hi, I am building a containerized Slurm cluster with Ubuntu 20.04 and have it almost working. The daemons start, and an “sinfo” command shows compute nodes up and available: admin@slurmfrontend:~$ sinfo PARTITION AVAIL TIMELIMIT NODES STATE NODELIST slurmpar* up infinite 3 idle slurmnode[1-3] admin@slurmfrontend:~$ However, if I try to use “srun” to test a job submission it fails saying it could not execve the job: admin@slurmfrontend:~$ srun hostname srun: error: task 0 launch failed: Slurmd could not execve job slurmstepd: error: task_g_set_affinity: Operation not permitted slurmstepd: error: _exec_wait_child_wait_for_parent: failed: No error admin@slurmfrontend:~$ If I go to the slurmnode1 container where the job should run, and look at the slurmd log, all I see is this: admin@slurmnode1:/$ sudo cat /var/log/slurmd.log [2024-06-13T14:58:36.238] CPU frequency setting not configured for this node [2024-06-13T14:58:36.239] warning: Core limit is only 0 KB [2024-06-13T14:58:36.239] slurmd version 23.11.7 started [2024-06-13T14:58:36.243] slurmd started on Thu, 13 Jun 2024 14:58:36 +0000 [2024-06-13T14:58:36.243] CPUs=8 Boards=1 Sockets=8 Cores=1 Threads=1 Memory=47926 TmpDisk=59767 Uptime=71713 CPUSpecList=(null) FeaturesAvail=(null) FeaturesActive=(null) [2024-06-13T14:58:47.230] launch task StepId=1.0 request from UID:1000 GID:1000 HOST:172.20.0.2 PORT:50618 [2024-06-13T14:58:47.230] task/affinity: lllp_distribution: JobId=1 implicit auto binding: sockets,one_thread, dist 8192 [2024-06-13T14:58:47.230] task/affinity: _task_layout_lllp_cyclic: _task_layout_lllp_cyclic [2024-06-13T14:58:47.230] task/affinity: _lllp_generate_cpu_bind: _lllp_generate_cpu_bind jobid [1]: mask_cpu,one_thread, 0x01 [2024-06-13T14:58:47.243] [1.0] error: task_g_set_affinity: Operation not permitted [2024-06-13T14:58:47.243] [1.0] error: _exec_wait_child_wait_for_parent: failed: No error [2024-06-13T14:58:47.244] [1.0] error: job_manager: exiting abnormally: Slurmd could not execve job [2024-06-13T14:58:47.247] [1.0] stepd_cleanup: done with step (rc[0xfb4]:Slurmd could not execve job, cleanup_rc[0xfb4]:Slurmd could not execve job) admin@slurmnode1:/$ I’ve installed by following the instructions for building/installing the Debian RPMs and can see that all the daemons are up and running. I have this slurm.conf (on all nodes): admin@slurmfrontend:~$ grep -v '#' /etc/slurm/slurm.conf ClusterName=cluster SlurmctldHost=slurmmaster MpiDefault=none ProctrackType=proctrack/linuxproc ReturnToService=1 SlurmdParameters=config_overrides SlurmctldPidFile=/var/run/slurmctld.pid SlurmctldPort=6817 SlurmdPidFile=/var/run/slurmd.pid SlurmdPort=6818 SlurmdSpoolDir=/var/spool/slurmd SlurmUser=root StateSaveLocation=/var/spool/slurmctld SwitchType=switch/none TaskPlugin=task/affinity InactiveLimit=0 KillWait=30 MinJobAge=300 SlurmctldTimeout=120 SlurmdTimeout=300 Waittime=0 SchedulerType=sched/backfill SelectType=select/cons_tres AccountingStorageType=accounting_storage/none JobCompType=jobcomp/none JobAcctGatherFrequency=30 JobAcctGatherType=jobacct_gather/none SlurmctldDebug=info SlurmctldLogFile=/var/log/slurmctld.log SlurmdDebug=info SlurmdLogFile=/var/log/slurmd.log NodeName=slurmnode[1-3] CPUs=8 State=UNKNOWN PartitionName=slurmpar Nodes=ALL Default=YES MaxTime=INFINITE State=UP admin@slurmfrontend:~$ And I have this group.conf (on all nodes): admin@slurmfrontend:~$ grep -v '#' /etc/slurm/cgroup.conf CgroupPlugin=cgroup/v1 ConstrainCores=yes ConstrainDevices=yes ConstrainRAMSpace=yes admin@slurmfrontend:~$ Does anyone have any clues about where to look for why “srun” can’t run a job and where the "task_g_set_affinity: Operation not permitted” may be coming from? Chris --------------------------------------------------------------------------------------------------- Christopher W. Harrop voice: (720) 649-0316 NOAA Global Systems Laboratory, R/GSL6 fax: (303) 497-7259 325 Broadway Boulder, CO 80303

1 0

Limit GPU depending on type
by Gestió Servidors 13 Jun '24

13 Jun '24

Hello, I would like to know if it would be possible to limit, using "sacctmgr", use of a certain type of GPU according the name I have assigned in "gres.conf" file. For example, my small cluster has 3 GPUs nodes sharing 2 GPUs each one. Two of that GPUs are the same model but they are located in different servers. Because of my scenario, I would like to limit users to user only one of that GPU type and not allowing to use both of them. For example: * gpu-node-1: * GTX1080 * RTX3080 * gpu-node-2: * GTX750 * GTX680 * gpu-node-3: * RTX2070 * RTX3080 What I want is users could user all of them but simultaniously, a user only could use one of the RTX3080. Using QoS I have created a new "qos" with "sacctmgr add qos test-gpu-limit MaxTRESPerUser=gres/gpu=1", but with this new QoS, users are limited to use only one GPU, even they need to use different GPUs models. I have tried with "sacctmgr add qos test-gpu-limit MaxTRESPerUser=gres/gpu:RTX3080:1" but system returns this error "sacctmgr: error: slurmdb_format_tres_str: no TRES id found for gres/gpu:RTX3080:1". So, could be possible to apply limits I want to apply? Thanks. -- [cid:image003.jpg@01DABD6A.8DFD6D30]<http://www.uab.cat/> Daniel Ruiz Molina Tècnic Mitjà Informàtic Arquitectura de Computadors i Sistemes Operatius Escola d'Enginyeria Edifici Q - Despatx QC/3052 - Carrer de les Sitges Campus de la UAB · 08193 Bellaterra (Cerdanyola del Vallès) · Barcelona · Spain +34 93 581 35 44 www.uab.cat<http://www.uab.cat/> Daniel Ruiz at UAB<https://tinyurl.com/yd95zb8j> [cid:image004.jpg@01DABD6A.8DFD6D30]<www.linkedin.com/in/daniel-ruiz-molina-50a83b27> Aquest missatge s'adreça exclusivament a la persona destinatària i pot contenir informació privada o confidencial. Si l'heu rebut per error, comuniqueu-nos-ho i destruïu-lo, i tingueu present que no teniu autorització per fer-ne cap ús. Abans d'imprimir aquest missatge penseu en el medi ambient.

2 1

srun hostname - Socket timed out on send/recv operation
by Arnuld 11 Jun '24

11 Jun '24

I have two machines. When I run "srum hostname" on one machine (it's both a controller and a node) then I get the hostname fine but I get socket timed out error in these two situations: 1) "srun hostname" on 2nd machine (it's a node) 2) "srun -N 2 hostname" on controller "scontrol show node" shows both mach2 and mach4. "sinfo" shows both nodes too. Also the job gets stuck forever in CG state after the error. Here is the output: $ srun -N 2 hostname mach2 srun: error: slurm_receive_msgs: [[mach4]:6818] failed: Socket timed out on send/recv operation srun: error: Task launch for StepId=2222.0 failed on node hpc4: Socket timed out on send/recv operation srun: error: Application launch failed: Socket timed out on send/recv operation srun: Job step aborted Output form "squeue" 3 seconds apart Tue Jun 11 05:09:56 2024 JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 2222 poxo hostname arnuld R 0:19 2 mach4,mach2 Tue Jun 11 05:09:59 2024 JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) 2222 poxo hostname arnuld CG 0:20 1 mach4

1 1

sbatch: Node count specification invalid - when only specifying --ntasks
by George Leaver 11 Jun '24

11 Jun '24

Hello, Previously we were running 22.05.10 and could submit a "multinode" job using only the total number of cores to run, not the number of nodes. For example, in a cluster containing only 40-core nodes (no hyperthreading), Slurm would determine two nodes were needed with only: sbatch -p multinode -n 80 --wrap="...." Now in 23.02.1 this is no longer the case - we get: sbatch: error: Batch job submission failed: Node count specification invalid At least -N 2 is must be used (-n 80 can be added) sbatch -p multinode -N 2 -n 80 --wrap="...." The partition config was, and is, as follows (MinNodes=2 to reject small jobs submitted to this partition - we want at least two nodes requested) PartitionName=multinode State=UP Nodes=node[081-245] DefaultTime=168:00:00 MaxTime=168:00:00 PreemptMode=OFF PriorityTier=1 DefMemPerCPU=4096 MinNodes=2 QOS=multinode Oversubscribe=EXCLUSIVE Default=NO All nodes are of the form NodeName=node245 NodeAddr=node245 State=UNKNOWN Procs=40 Sockets=2 CoresPerSocket=20 ThreadsPerCore=1 RealMemory=187000 slurm.conf has EnforcePartLimits = ANY SelectType = select/cons_tres TaskPlugin = task/cgroup,task/affinity A few fields from: sacctmgr show qos multinode Name|Flags|MaxTRES multinode|DenyOnLimit|node=5 The sbatch/srun man page states: -n, --ntasks .... If -N is not specified, the default behavior is to allocate enough nodes to satisfy the requested resources as expressed by per-job specification options, e.g. -n, -c and --gpus. I've had a look through release notes back to 22.05.10 but can't see anything obvious (to me). Has this behaviour changed? Or, more likely, what have I missed ;-) ? Many thanks, George -- George Leaver Research Infrastructure, IT Services, University of Manchester http://ri.itservices.manchester.ac.uk | @UoM_eResearch

2 3

Issue about selecting cpus for optimization
by Purvesh Parmar 10 Jun '24

10 Jun '24

Hi, We have 16 nodes cluster with DGX-A100 (80 GB). We have 128 cores of each node separated in to a separate partition for cpu only jobs and 8 GPUs and 128 cores in other partitions for cpugpu jobs. We want to ensure that only selected 128 cores should be part of the cpu partition. (NUMA / Symmetry) for optimization. How to achieve it? cores parameter in gres.conf help? Regards, Purvesh

1 0

scontrol create partition fails
by Long, Daniel S. 10 Jun '24

10 Jun '24

Hi, I need to temporarily dedicated one of our compute nodes to a single account. To do this, I was going to create a new partition but I'm running into an error where scontrol create partition outputs "scontrol: error: Invalid input: partition Request aborted" regardless of what parameters I give it. As far as I can tell, this should be allowed; the man page for scontrol has a whole section titled "Partitions - Specficiations for Create, Update, and Delete Commands". What am I missing? Also, is there a better way to approach this? This is really just a one or two day thing and I'm a little surprised there's no easy way to cordon off a node for a user or a project without spinning up an entire partition. Am I missing something obvious? Thanks for any help you can provide.

2 2

Software builds using slurm
by Duane Ellis 10 Jun '24

10 Jun '24

I have been lurking here for a while hoping to see some examples that would help but have not fit several months We have a slurm system setup for xilnix FPGA builds (hdl) I want to use this for sw builds too What I seem to see is slurm talks about cpus, GPUs and memory etc I am looking for a “run my make file (or shell script) on any available node” In our case we have 3 top level jobs A B and C These can all run in parallel and are independent (ie bootloader, linux kernel, and the Linux root file system via buildroot) Job A (boot) is actually about 7 small builds that are independent I am looking for a means to fork n jobs (ie job A B and C above) across the cluster and wait/collect the std output of those n jobs and the exit status Job A would then fork and build 7 to 8 sub jobs When they are done it would assemble the result into what Xilinix calls boot.bin Job B is a Linux kernel build Job C is buildroot so there are several (n=50) smaller builds ie bash, busybody, and other tools like python for the target agian each of these can be executed in parallel Really do not (cannot) re architect my build to be a slurm only build because it also need to be able to run without slurm ie build everything on my laptop without slurm present In that case the jobs would run serially and take an hour or so the hope is by parallelizing the sw build jobs our overall cycle time will improve It would also be nice if the slurm cluster would adapt to the available nodes automatically Our hope is we can run our lab pcs as duel boot they normally boot windows but we can duel boot them into Linux and they become a compile node and auto join the cluster and the cluster sees them as going off line when somebody reboots the machine back to windows Sent from my iPhone

3 2

Re: sbatch: Node count specification invalid - when only specifying --ntasks
by George Leaver 10 Jun '24

10 Jun '24

Noam, Thanks for the suggestion but no luck: sbatch -p multinode -n 80 --ntasks-per-core=1 --wrap="..." sbatch: error: Batch job submission failed: Node count specification invalid sbatch -p multinode -n 2 -c 40 --ntasks-per-core=1 --wrap="..." sbatch: error: Batch job submission failed: Node count specification invalid sbatch -p multinode -N 2 -n 80 --ntasks-per-core=1 --wrap="..." Submitted batch job I guess that the MinNodes=2 in the partition def is now being enforced somewhat more strictly, or earlier in the submission process, before it can be determined that the request will satisfy the constraint. Regards, George -- George Leaver Research Infrastructure, IT Services, University of Manchester http://ri.itservices.manchester.ac.uk | @UoM_eResearch ________________________________________ From: Bernstein, Noam CIV USN NRL WASHINGTON DC (USA) <noam.bernstein.civ(a)us.navy.mil> Sent: 09 June 2024 19:33 To: George Leaver; slurm-users(a)lists.schedmd.com Subject: Re: sbatch: Node count specification invalid - when only specifying --ntasks It would be a shame to lose this capability. Have you tried adding `--ntasks-per-core` explicitly (but not number of nodes)? Noam

1 0

cpu distribution question
by Alan Stange 08 Jun '24

08 Jun '24

All, I have a very simple slurm cluster. It's just a single system with 2 sockets and 16 cores in each socket. I would like to be able to submit a simple task into this cluster, and to have the cpus assigned to that task allocated round robin across the two sockets. Everything I try is putting all the cpus for this single task on the same socket. I have not specified any CpuBind options in the slurm.conf file. For example, if I try $ srun -c 4 --pty bash I get a shell prompt on the system, and can run $ taskset -cp $$ pid 12345 current affinity list: 0,2,4,6 and I get this same set of cpus no matter what options I try (the cluster is idle with no tasks consuming slots). I've tried various srun command line options like: --hint=compute_bound --hint=memory_bound various --cpubind options -B 2:2 -m block:cyclic and block:fcyclic Note that if I try to allocation 17 cpus, then I do get the 17th cpu allocated on the 2nd socket. What magic incantation is needed to get an allocation where the cpus are chosen round robin across the sockets? Thank you! Alan

3 3

maxrss reported by sachet is wrong
by Feng Zhang 07 Jun '24

07 Jun '24

Hi All, I am having trouble calculating the real RSS memory usage by some kind of users' jobs. Which the sacct returned wrong numbers. Rocky Linux release 8.5, Slurm 21.08 (slurm.conf) ProctrackType=proctrack/cgroup JobAcctGatherType=jobacct_gather/linux The troubling jobs are like: 1. python spawn multithreading 96 threads; 2. Each thread uses SKlearn which again spawns 96 threads using openmp. Which is obviously over running the node, and I want to address it. The node has 300GB RAM, but the "sacct" (and seff) reports 1.2TB MaxRSS(also AveRSS). This does not look correct. I am suspecting that whether the SLurm+jobacct_gather/linux repeatedly sums up the memory used by all these threads, multiple counted the same thing many times. For the openMP part, maybe it is fine for slurm; while for python/multithreading, maybe it can not work well with Slurm for memory accounting? So, if this is the case, maybe 1.2TB/96= 12GB MaxRSS? I want to get the right MaxRSS to report to users. Thanks! Best, Feng

2 1

need to set From: address for slurm
by Vanhorn, Mike 07 Jun '24

07 Jun '24

All, When the slurm daemon is sending out emails, they are coming from “slurm(a)servername.subdomain.domain.edu<mailto:slurm@servername.subdomain.domain.edu>”. This has worked okay in the past, but due to a recent mail server change (over which I have no control whatsoever) this will no longer work. Now, the From: address is going to have to be something like “slurm-servername(a)domain.xn--edu-9o0a , or, at least something that ends in “(a)domain.xn--edu-9o0a (the subdomain being present will cause it to get rejected by the mail server. I am not seeing in the documentation how to change the “From:” address tha slurm uses. Is there a way to do this and I’m just missing it? --- Mike VanHorn Senior Computer Systems Administrator College of Engineering and Computer Science Wright State University 265 Russ Engineering Center 937-775-5157 michael.vanhorn(a)wright.edu

2 2

Re: Not being able to ssh to node with running job
by Ratnasamy, Fritz 06 Jun '24

06 Jun '24

So the squeue issue was resolved and was due to the partition being hidden. Unhiding it solves the problem. However, the ssh issue remains (looks like both were separate issues). The pam_slurm_adopt is working on all the other nodes but not on the new ones. Any idea how to solve this? Best, *Fritz Ratnasamy* Data Scientist Information Technology On Thu, Jun 6, 2024 at 2:11 PM Ratnasamy, Fritz via slurm-users < slurm-users(a)lists.schedmd.com> wrote: > As admin on the cluster, we do not observe any issue on our newly added > gpu nodes. > However, for regular users, they are not seeing their jobs running on > these gpu nodes when running squeue -u <username> ( it is > however showing as running status with sacct) and they are not able to ssh > to these newly added nodes when they have a running job on it. > I am not sure if these 2 are related (not being to ssh to mgpu node with > a running job on it and not listing a job with squeue for a user on the > same node). There are no issues reported on the other nodes. Anyone know > what is happening? > Best, > > *Fritz Ratnasamy* > > Data Scientist > > Information Technology > > > CAUTION: This email has originated outside of University email systems. > Please do not click links or open attachments unless you recognize the > sender and trust the contents as safe. > >

1 0

Re: JWKS File Update - Reread File
by Pedro Daniel Da Rocha Santos 06 Jun '24

06 Jun '24

Thanks for the reply. I was not aware of that change in 23.11. I am still using Slurm 23.02.7. Best regards, Pedro D.R. Santos Specialist, Ph.D., High Performance Computing Vestas Technology Centre Porto Vestas Wind Systems A/S Vestas will as part of your communication, interaction and business relationship with us collect and process personal data about you. You can read more about Vestas' collection and processing of your personal data and your rights as a data subject in our Privacy Policy<https://www.vestas.com/en/about/profile/privacy-policy>. Classification: Restricted

1 0

Not being able to ssh to node with running job
by Ratnasamy, Fritz 06 Jun '24

06 Jun '24

As admin on the cluster, we do not observe any issue on our newly added gpu nodes. However, for regular users, they are not seeing their jobs running on these gpu nodes when running squeue -u <username> ( it is however showing as running status with sacct) and they are not able to ssh to these newly added nodes when they have a running job on it. I am not sure if these 2 are related (not being to ssh to mgpu node with a running job on it and not listing a job with squeue for a user on the same node). There are no issues reported on the other nodes. Anyone know what is happening? Best, *Fritz Ratnasamy* Data Scientist Information Technology

1 0

JWKS File Update - Reread File
by Pedro Daniel Da Rocha Santos 06 Jun '24

06 Jun '24

Hello everyone, I am currently implementing a service to monitor for jwks key rotation. If a change is detected, the jwks file is updated. Is it enough to use "scontrol reconfigure" to reread the jwks file? Or does it require a full restart of the slurmctl daemon? Thanks in advance for the inputs. Best regards, Pedro D.R. Santos Specialist, Ph.D., High Performance Computing Vestas Technology Centre Porto Vestas Wind Systems A/S +351 924 241 907 pddro(a)vestas.com<mailto:pddro@vestas.com> http://www.vestas.com<http://www.vestas.com/> Company reg. name: Vestas Wind Systems A/S This e-mail is subject to our e-mail disclaimer statement. Please refer to www.vestas.com/legal/notice<http://www.vestas.com/legal/notice> If you have received this e-mail in error please contact the sender. Vestas will as part of your communication, interaction and business relationship with us collect and process personal data about you. You can read more about Vestas' collection and processing of your personal data and your rights as a data subject in our Privacy Policy<https://www.vestas.com/en/about/profile/privacy-policy>. Classification: Restricted

2 1

diagnosing why interactive/non-interactive job waits are so long with State=MIXED
by Robert Kudyba 05 Jun '24

05 Jun '24

At the moment we have 2 nodes that are having long wait times. Generally this is when the nodes are fully allocated. What would be the other reasons if there is still enough available memory and CPU available, that a job would take so long? Slurm version is 23.02.4 via Bright Computing. Note the compute nodes have hyperthreading enabled but that should be irrelevant. Is there a way to determine what else could be holding jobs up? srun --pty -t 0-01:00:00 --nodelist=node001 --gres=gpu:1 -A ourts -p short /bin/bash srun: job 672204 queued and waiting for resources scontrol show node node001 NodeName=m001 Arch=x86_64 CoresPerSocket=48 CPUAlloc=24 CPUEfctv=192 CPUTot=192 CPULoad=20.37 AvailableFeatures=location=local ActiveFeatures=location=local Gres=gpu:A6000:8 NodeAddr=node001 NodeHostName=node001 Version=23.02.4 OS=Linux 5.14.0-70.13.1.el9_0.x86_64 #1 SMP PREEMPT Thu Apr 14 12:42:38 EDT 2022 RealMemory=1031883 AllocMem=1028096 FreeMem=222528 Sockets=2 Boards=1 State=MIXED ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A Partitions=ours,short BootTime=2024-04-29T16:18:30 SlurmdStartTime=2024-05-18T16:48:11 LastBusyTime=2024-06-03T10:49:49 ResumeAfterTime=None CfgTRES=cpu=192,mem=1031883M,billing=192,gres/gpu=8 AllocTRES=cpu=24,mem=1004G,gres/gpu=2,gres/gpu:a6000=2 CapWatts=n/a CurrentWatts=0 AveWatts=0 ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s grep 672204 /var/log/slurmctld [2024-06-04T15:50:35.627] sched: _slurm_rpc_allocate_resources JobId=672204 NodeList=(null) usec=852

3 5

Re: Problems with gres.conf
by Gestió Servidors 05 Jun '24

05 Jun '24

Hi, my GPU testing system (named “gpu-node”) is a simple computer with one socket and a processor " Intel(R) Core(TM) i7 CPU 950 @ 3.07GHz". Executing "lscpu", I can see there are 4 cores per socket, 2 threads per core and 8 CPUs: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 On-line CPU(s) list: 0-7 Thread(s) per core: 2 Core(s) per socket: 4 Socket(s): 1 NUMA node(s): 1 Vendor ID: GenuineIntel CPU family: 6 Model: 26 Model name: Intel(R) Core(TM) i7 CPU 950 @ 3.07GHz My “gres.conf” file is: NodeName=gpu-node Name=gpu Type=GeForce-GTX-TITAN-X File=/dev/nvidia0 CPUs=0-1 NodeName=gpu-node Name=gpu Type=GeForce-GTX-TITAN-Black File=/dev/nvidia1 CPUs=2-3 Running “numactl -H” in “gpu-node” host, reports: available: 1 nodes (0) node 0 cpus: 0 1 2 3 4 5 6 7 node 0 size: 7809 MB node 0 free: 6597 MB node distances: node 0 0: 10 CPUs are assigned 0-1 for first GPU and 2-3 for second GPU. However, “lscpu” shows 8 CPUs… If I rewrite “gres.conf” in this way: NodeName=gpu-node Name=gpu Type=GeForce-GTX-TITAN-X File=/dev/nvidia0 CPUs=0-3 NodeName=gpu-node Name=gpu Type=GeForce-GTX-TITAN-Black File=/dev/nvidia1 CPUs=4-7 when I run “scontrol reconfigure”, slurmctld log reports this error message: [2024-06-05T11:42:18.558] error: _node_config_validate: gres/gpu: invalid GRES core specification (4-7) on node gpu-node So I think SLURM only can get physical cores and not threads, so my node only can serve 4 cores (in “lspcu”) but in gres.conf I need to write “CPUs”, not “Cores”… isn’t it? But if “numactl -H” shows 8 CPUs, why I can use this 8 CPUs in “gres.conf”? Sorry about this large email. Thanks.

1 0

Problems with gres.conf
by Gestió Servidors 04 Jun '24

04 Jun '24

Hello, I am trying to rewrite my gres.conf file. Before changes, this file was just like this: NodeName=node-gpu-1 AutoDetect=off Name=gpu Type=GeForceRTX2070 File=/dev/nvidia0 Cores=0-11 NodeName=node-gpu-1 AutoDetect=off Name=gpu Type=GeForceGTX1080Ti File=/dev/nvidia1 Cores=12-23 NodeName=node-gpu-2 AutoDetect=off Name=gpu Type=GeForceGTX1080Ti File=/dev/nvidia0 Cores=0-11 NodeName=node-gpu-2 AutoDetect=off Name=gpu Type=GeForceGTX1080 File=/dev/nvidia1 Cores=12-23 NodeName=node-gpu-3 AutoDetect=off Name=gpu Type=GeForceRTX3080 File=/dev/nvidia0 Cores=0-11 NodeName=node-gpu-4 AutoDetect=off Name=gpu Type=GeForceRTX3080 File=/dev/nvidia0 Cores=0-7 # you can seee that nodes node-gpu-1 and node-gpu-2 have two GPUs each one, whereas nodes node-gpu-3 and node-gpu-4 have only one GPU each one And my slurmd.conf was this: [...] AccountingStorageTRES=gres/gpu GresTypes=gpu NodeName=node-gpu-1 CPUs=24 SocketsPerBoard=2 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=96000 TmpDisk=47000 Gres=gpu:GeForceRTX2070:1,gpu:GeForceGTX1080Ti:1 NodeName=node-gpu-2 CPUs=24 SocketsPerBoard=2 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=96000 TmpDisk=47000 Gres=gpu:GeForceGTX1080Ti:1,gpu:GeForceGTX1080:1 NodeName=node-gpu-3 CPUs=12 SocketsPerBoard=1 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=23000 Gres=gpu:GeForceRTX3080:1 NodeName=node-gpu-4 CPUs=8 SocketsPerBoard=1 CoresPerSocket=4 ThreadsPerCore=2 RealMemory=7800 Gres=gpu:GeForceRTX3080:1 NodeName=node-worker-[0-22] CPUs=12 SocketsPerBoard=1 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=47000 [...] With this configuration, all seems works fine, except slurmctld.log reports: [...] error: _node_config_validate: gres/gpu: invalid GRES core specification (0-11) on node node-gpu-3 error: _node_config_validate: gres/gpu: invalid GRES core specification (12-23) on node node-gpu-1 error: _node_config_validate: gres/gpu: invalid GRES core specification (12-23) on node node-gpu-2 error: _node_config_validate: gres/gpu: invalid GRES core specification (0-7) on node node-gpu-4 [...] However, even these errors, users can submit jobs and request GPUs resources. Now, I have tried to reconfigure gres.conf and slurmd.conf in this way: gres.conf: Name=gpu Type=GeForceRTX2070 File=/dev/nvidia0 Name=gpu Type=GeForceGTX1080Ti File=/dev/nvidia1 Name=gpu Type=GeForceGTX1080Ti File=/dev/nvidia0 Name=gpu Type=GeForceGTX1080 File=/dev/nvidia1 Name=gpu Type=GeForceRTX3080 File=/dev/nvidia0 Name=gpu Type=GeForceRTX3080 File=/dev/nvidia0 # there is no NodeName attribute slurmd.conf: [...] NodeName=node-gpu-1 SocketsPerBoard=2 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=96000 TmpDisk=47000 Gres=gpu:GeForceRTX2070:1,gpu:GeForceGTX1080Ti:1 NodeName=node-gpu-2 SocketsPerBoard=2 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=96000 TmpDisk=47000 Gres=gpu:GeForceGTX1080Ti:1,gpu:GeForceGTX1080:1 NodeName=node-gpu-3 SocketsPerBoard=1 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=23000 Gres=gpu:GeForceRTX3080:1 NodeName=node-gpu-4 SocketsPerBoard=1 CoresPerSocket=4 ThreadsPerCore=2 RealMemory=7800 Gres=gpu:GeForceRTX3080:1 NodeName=node-worker-[0-22] SocketsPerBoard=1 CoresPerSocket=6 ThreadsPerCore=2 RealMemory=47000 # there is no CPUs attribute [...] With this new configuration, nodes with GPU start correctly slurmd.service daemon, but nodes without GPU (node-worker-[0-22]) can't start slurmd.service daemon and returns this error: [...] error: Waiting for gres.conf file /dev/nvidia0 fatal: can't stat gres.conf file /dev/nvidia0: No such file or directory [...] It seems SLURM is waiting that "node-workers" have also an nvidia GPU but not, theses nodes haven't GPU... So, where is my configuration error? I have read in https://slurm.schedmd.com/gres.conf.html about syntax and examples but it seems I'm doing some wrong. Thanks!!

2 1

Tracking GPU utilization with sacct
by Robertson, Jordan 03 Jun '24

03 Jun '24

Hi all, I’m trying to pull (and understand) some GPU usage metrics for historical purposes, and dug into sacct’s TRES reporting a bit. We have AccountingStorageTRES=gres/gpu set in slurm.conf so we do see gres/gpuutil and gres/gpumem numbers available, but I’m struggling to find Slurm-side documentation that describes the units of these values. In looking at the code for gpu_nvml.c it seems the “nvmlDeviceGetProcessUtilization” function is being used and returns units in percentages, but I’m lost on the rest of the calculation. Does anyone know if these units are percentages, and how they are calculated for the final job record, especially wrt multi-GPU jobs with a bunch of processes/moving parts? For context I’ve been looking at TRESUsageInTot and TRESUsageInAve so far. Also we’re currently running Slurm v23.02.6 Thanks in advance! -- Jordan Robertson Preferred pronouns: he/him/his Technology Architect | Research Technology Services DigITs, Technology Division Memorial Sloan Kettering Cancer Center 929-687-1066 robertj8(a)mskcc.org<mailto:robertj8@mskcc.org> ===================================================================== Please note that this e-mail and any files transmitted from Memorial Sloan Kettering Cancer Center may be privileged, confidential, and protected from disclosure under applicable law. If the reader of this message is not the intended recipient, or an employee or agent responsible for delivering this message to the intended recipient, you are hereby notified that any reading, dissemination, distribution, copying, or other use of this communication or any of its attachments is strictly prohibited. If you have received this communication in error, please notify the sender immediately by replying to this message and deleting this message, any attachments, and all copies and backups from your computer. Disclaimer ID:MSKCC

1 0

sbatch and --nodes
by Michael DiDomenico 31 May '24

31 May '24

its friday and i'm either doing something silly or have a misconfig somewhere, i can't figure out which when i run sbatch --nodes=1 --cpus-per-task=1 --array=1-100 --output test_%A_%a.txt --wrap 'uname -n' sbatch doesn't seem to be adhering to the --nodes param. when i look at my output files it's spreading them across more nodes. in the simple case above it's 50/50, but if i through a random sleep in, it'll be more. and if i expand the array it'll use even more nodes. i'm using con/tres and have cr_core_memory,cr_one_core_per_task set

2 1