February 2024 - slurm-users - lists.schedmd.com

job_container/tmpfs and srun.
by Phill Harvey-Smith 17 Jan '25

17 Jan '25

Hi all, On our setup we are using job_container/tmpfs to give each job it's own temp space. Since our compute nodes have reasonably sized disks for tasks that do a lot of disk I/O on user's data we have asked users to copy their data to the local disk at the beginning of the task and (if needed) copy it back at the end. This saves lots of NFS thrashing slowing down both the task and the NFS servers. However some of our users are having problems with this, their initial sbatch script will create a temp directory in their private /tmp copy their data to it and then try to srun a program. The srun will fall over as it doesn't seem to have have access to the copied data. I suspect this is because the srun task is getting it's own private /tmp. So my question is, is there a way to have the srun task inherit the /tmp of the initial sbatch? I'll include a sample of the script our user is using below. If any further information is required please feel free to ask. Cheers. Phill. #!/usr/bin/bash #SBATCH --nodes 1 #SBATCH --ntasks-per-node=1 #SBATCH --cpus-per-task=1 #SBATCH --time=00:00:10 #SBATCH --mem-per-cpu=3999 #SBATCH --output=script_out.log #SBATCH --error=script_error.log # The above options puts the STDOUT and STDERR of sbatch in # log files prefixed with 'script_'. # Create a randomly-named directory under /tmp jobtmpdir=$(mktemp -d) # Register a function to try and cleanup in case of job failure cleanup_handler() { echo "Cleaning up ${jobtmpdir}" rm -rf ${jobtmpdir} } trap 'cleanup_handler' SIGTERM EXIT # Change working directory to this directory cd ${jobtmpdir} # Copy the executable and input files from # where the job was submitted to the temporary directory. cp ${SLURM_SUBMIT_DIR}/a.out . cp ${SLURM_SUBMIT_DIR}/input.txt . # Run the executable, handling the collection of stdout # and stderr ourselves by redirecting to file srun ./a.out 2> task_error.log > task_out.log # Copy output data back to the submit directory. cp output.txt ${SLURM_SUBMIT_DIR} cp task_out.log ${SLURM_SUBMIT_DIR} cp task_error.log ${SLURM_SUBMIT_DIR} # Cleanup cd ${SLURM_SUBMIT_DIR} cleanup_handler

2 1

Can't schedule on cloud node: State=IDLE+CLOUD+POWERED_DOWN+NOT_RESPONDING
by Xaver Stiensmeier 20 Sep '24

20 Sep '24

Dear slurm-user list, I have a cloud node that is powered up and down on demand. Rarely it can happen that slurm's resumeTimeout is reached and the node is therefore powered down. We have set ReturnToService=2 in order to avoid the node being marked down, because the instance behind that node is created on demand and therefore after a failure nothing stops the system to start the node again as it is a different instance. I thought this would be enough, but apparently the node is still marked with "NOT_RESPONDING" which leads to slurm not trying to schedule on it. After a while NOT_RESPONDING is removed, but I would like to move it directly from within my fail script if possible so that the node can return to service immediately and not be blocked by "NOT_RESPONDING". Best regards, Xaver

2 3

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

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

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

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

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