- slurm-users - lists.schedmd.com

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