I built 23.02.7 and tried that and had the same problems.
BTW, I am using the slurm.spec rpm build method (built on Rocky 8 boxes
with NVIDIA 535.54.03 proprietary drives installed).
The behavior I was seeing was one would start a GPU job. It was fine at
first but at some point the slurmstepd processes for it would crash/die.
Sometimes the user process for it would die too, sometimes not. In an
interactive job you would sometimes see some final line about "malloc" and
"invalid value" and the terminal would be hung until the job was
'scancel'ed. 'ps' would show a 'slurmd <defunct>' process that was
unkillable (killing the main slurmd process would get rid of it)
How the slurm controller saw the job seemed random. Sometimes it saw it
as a crashed job and it would be reported like that in the system.
Sometimes it was stuck as a permanently CG (completing) job. Sometimes it
did not notice anything wrong and just stayed as a seemingly perfect
running job according to slurmctld (I never waited for the TimeLimit to
hit to see what happened then but did scancel).
I scancelled a "failed" job in the CG or R state, it would not actually
kill the user processes on the node but it would clear the job from
squeue.
Jobs on my non-GPU Rocky 8 nodes or on my Ubuntu GPU nodes (slurm 23.11.3
install built sepearately on these) have all been working fine so far.
The install on the Ubuntu GPU boxes was built separately on them and also
another difference is they are still using NVIDIA 470 drivers.
I tried downgrading a Rocky 8 GPU box to NVIDIA 470 and rebuilding
slurm 23.11.3 there and installing it to see if that worked
to fix things. It did not.
I then tried installing old 22.05.6 RPMs I had built on my Rocky 8
box back in Nov 2022 on all Rocky 8 GPU boxes. This seems to
have fixed the problem and jobs are no longer showing the issue.
Not and ideal solution but good enough for now.
Both those 22.05.6 RPMs and the 23.11.3 RPMs are built on the
same Rocky 8 GPU box. So the differences are:
1) slurm version obviously
2) built using different gcc/lib versions mostly likely
due to OS updates between Nov 2022 and now
3) built with a different NVIDIA driver/cuda installed
between then and now but I am not sure what I had
in Nov 2022
I highly suspect #2 or #3 as the underlying issue here and I wonder if the
NVML library at the time of build is the key (though like I said I tried
rebuiling with NVIDIA 470 and that still had the issue)
-- Paul Raines (http://help.nmr.mgh.harvard.edu)
On Tue, 30 Jan 2024 3:36am, Fokke Dijkstra wrote:
External Email - Use Caution
We had similar issues with Slurm 23.11.1 (and 23.11.2). Jobs get stuck in
a completing state and slurmd daemons can't be killed because they are left
in a CLOSE-WAIT state. See my previous mail to the mailing list for the
details. And also
https://secure-web.cisco.com/16zDED-XPBGKgr-e6-N-CrY1J1qxgnDyBeTc5vVc0ghVjnVNEtDKVr66UkuO1IT_ALDy93Yc0AmTeUXzIZviB36h_FFVEoHsmy36XMo6j8usUs-kk5AjTgDUP4KUbfOzFSwWiOoSZGqRWubpW1s47fI9tY-S6zRRk3tl2C_RZ7VpbkyCXnPlNaSsYQvxM_MkVKpbsUrwVDWQ0aqr4jIDrKr-ddHpVhwaN7YmukIpwG4dXoXlh5yqq1VznG-xDJ2Xabhr08FF6AdRHwHQXYPWTR1hLchTEGLxENISVpgGs_PwosuU-VzgMEGp9KSRIjCM9y3MV5gXoJhG44TKX9jkYQg/https%3A%2F%2Fbugs.schedmd.com%2Fshow_bug.cgi%3Fid%3D18561
for
another site having issues.
We've now downgraded the clients (slurmd and login nodes) to 23.02.7 which
gets rid of most issues. If possible, I would try to also downgrade
slurmctld to an earlier release, but this requires getting rid of all
running and queued jobs.
Kind regards,
Fokke
Op ma 29 jan 2024 om 01:00 schreef Paul Raines <rai...@nmr.mgh.harvard.edu>:
Some more info on what I am seeing after the 23.11.3 upgrade.
Here is a case where a job is cancelled but seems permanently
stuck in 'CG' state in squeue
[2024-01-28T17:34:11.002] debug3: sched: JobId=3679903 initiated
[2024-01-28T17:34:11.002] sched: Allocate JobId=3679903 NodeList=rtx-06
#CPUs=4 Partition=rtx8000
[2024-01-28T17:34:11.002] debug3: create_mmap_buf: loaded file
`/var/slurm/spool/ctld/hash.3/job.3679903/script` as buf_t
[2024-01-28T17:42:27.724] _slurm_rpc_kill_job: REQUEST_KILL_JOB
JobId=3679903 uid 5875902
[2024-01-28T17:42:27.725] debug: email msg to sg1526: Slurm
Job_id=3679903 Name=sjob_1246 Ended, Run time 00:08:17, CANCELLED,
ExitCode 0
[2024-01-28T17:42:27.725] debug3: select/cons_tres: job_res_rm_job:
JobId=3679903 action:normal
[2024-01-28T17:42:27.725] debug3: select/cons_tres: job_res_rm_job:
removed JobId=3679903 from part rtx8000 row 0
[2024-01-28T17:42:27.726] job_signal: 9 of running JobId=3679903
successful 0x8004
[2024-01-28T17:43:19.000] Resending TERMINATE_JOB request JobId=3679903
Nodelist=rtx-06
[2024-01-28T17:44:20.000] Resending TERMINATE_JOB request JobId=3679903
Nodelist=rtx-06
[2024-01-28T17:45:20.000] Resending TERMINATE_JOB request JobId=3679903
Nodelist=rtx-06
[2024-01-28T17:46:20.000] Resending TERMINATE_JOB request JobId=3679903
Nodelist=rtx-06
[2024-01-28T17:47:20.000] Resending TERMINATE_JOB request JobId=3679903
Nodelist=rtx-06
So at 17:42 the user must of done an scancel. In the slurmd log on the
node I see:
[2024-01-28T17:42:27.727] debug: _rpc_terminate_job: uid = 1150
JobId=3679903
[2024-01-28T17:42:27.728] debug: credential for job 3679903 revoked
[2024-01-28T17:42:27.728] debug: _rpc_terminate_job: sent SUCCESS for
3679903, waiting for prolog to finish
[2024-01-28T17:42:27.728] debug: Waiting for job 3679903's prolog to
complete
[2024-01-28T17:43:19.002] debug: _rpc_terminate_job: uid = 1150
_JobId=3679903
[2024-01-28T17:44:20.001] debug: _rpc_terminate_job: uid = 1150
JobId=3679903
[2024-01-28T17:45:20.002] debug: _rpc_terminate_job: uid = 1150
JobId=3679903
[2024-01-28T17:46:20.001] debug: _rpc_terminate_job: uid = 1150
JobId=3679903
[2024-01-28T17:47:20.002] debug: _rpc_terminate_job: uid = 1150
JobId=3679903
Strange that a prolog is being called on job cancel
slurmd seems to be getting the repeated calls to terminate the job
from slurmctld but it is not happening. Also the process table has
[root@rtx-06 ~]# ps auxw | grep slurmd
root 161784 0.0 0.0 436748 21720 ? Ssl 13:44 0:00
/usr/sbin/slurmd --systemd
root 190494 0.0 0.0 0 0 ? Zs 17:34 0:00
[slurmd] <defunct>
where there is now a zombie slurmd process I cannot kill even with kill -9
If I do a 'systemctl stop slurmd' it takes a long time but eventually stop
slurmd and gets rid of the zombie process but kills the "good" running
jobs too with NODE_FAIL.
Another case is where a job will be cancelled and SLURM acts like it
is cancelled with it not showing up in squeue but the process keep running
on the box.
# pstree -u sg1526 -p | grep ^slurm
slurm_script(185763)---python(185796)-+-{python}(185797)
# strings /proc/185763/environ | grep JOB_ID
SLURM_JOB_ID=3679888
# squeue -j 3679888
slurm_load_jobs error: Invalid job id specified
sacct shows that job being cancelled. In the slurmd log we see
[2024-01-28T17:33:58.757] debug: _rpc_terminate_job: uid = 1150
JobId=3679888
[2024-01-28T17:33:58.757] debug: credential for job 3679888 revoked
[2024-01-28T17:33:58.757] debug: _step_connect: connect() failed for
/var/slurm/spool/d/rtx-06_3679888.4294967292: Connection refused
[2024-01-28T17:33:58.757] debug: Cleaned up stray socket
/var/slurm/spool/d/rtx-06_3679888.4294967292
[2024-01-28T17:33:58.757] debug: signal for nonexistent
StepId=3679888.extern stepd_connect failed: Connection refused
[2024-01-28T17:33:58.757] debug: _step_connect: connect() failed for
/var/slurm/spool/d/rtx-06_3679888.4294967291: Connection refused
[2024-01-28T17:33:58.757] debug: Cleaned up stray socket
/var/slurm/spool/d/rtx-06_3679888.4294967291
[2024-01-28T17:33:58.757] _handle_stray_script: Purging vestigial job
script /var/slurm/spool/d/job3679888/slurm_script
[2024-01-28T17:33:58.757] debug: signal for nonexistent
StepId=3679888.batch stepd_connect failed: Connection refused
[2024-01-28T17:33:58.757] debug2: No steps in jobid 3679888 were able to
be signaled with 18
[2024-01-28T17:33:58.757] debug2: No steps in jobid 3679888 to send signal
15
[2024-01-28T17:33:58.757] debug2: set revoke expiration for jobid 3679888
to 1706481358 UTS
[2024-01-28T17:33:58.757] debug: Waiting for job 3679888's prolog to
complete
[2024-01-28T17:33:58.757] debug: Finished wait for job 3679888's prolog
to complete
[2024-01-28T17:33:58.771] debug: completed epilog for jobid 3679888
[2024-01-28T17:33:58.774] debug: JobId=3679888: sent epilog complete msg:
rc = 0
-- Paul Raines
(http://secure-web.cisco.com/19HZZ6loaWDjqm3fm3-mLdIlOO486WzCozFRPfjXcoODk0os6A1hBP3WXDCQ4u_tmUyP3Bf-6x2dhzz0q8D1mA84xHUvIyvr2xldvTcSkypQnYUyPES7j9mmdxcCaxjp7TP6LMC_KgCVSvIMODRvyZ5H-YNzmogGLC0uiZeyvm7U_3jb1TSfOVFL4vQRJ45wu2nK3upi8FNEvNVCu43hEKoEPJq37vkmZKq1lPK8drgpcN1dEqGGOXp-iuaEdLoEqJwAzdocv_Ozep4OchdGzU7KjRMe2J-pKuXJV3_GICfbHOXHqtAV6KtJy2d_LTozW4EqyODnrwPH8GHd01bFLpw/http%3A%2F%2Fhelp.nmr.mgh.harvard.edu)
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.
--
Fokke Dijkstra <f.dijks...@rug.nl> <f.dijks...@rug.nl>
Team High Performance Computing
Center for Information Technology, University of Groningen
Postbus 11044, 9700 CA Groningen, The Netherlands
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.
