Ticket 12826

In addition, if I remove the cgroup_allowed_devices_file.conf (which I realise isn't necessary since about slurm 18) the particular message  about Invalid arguments goes away, however I still can't see the device.

Dr Chris Hines
Senior Research DevOps Engineer

Created attachment 22161 [details]
Slurm configurations

Hi,
Some more information. I added additional printf statements and realised that all of my devices are being added to devices.deny

My node looks like
ubuntu@mlerp-login0:~$ scontrol show node mlerp-node05
NodeName=mlerp-node05 Arch=x86_64 CoresPerSocket=1 
   CPUAlloc=64 CPUTot=64 CPULoad=0.00
   AvailableFeatures=(null)
   ActiveFeatures=(null)
   Gres=gpu:2g.10gb:3(S:0-63)
   NodeAddr=mlerp-node05 NodeHostName=mlerp-node05 Version=21.08.3
   OS=Linux 5.4.0-89-generic #100-Ubuntu SMP Fri Sep 24 14:50:10 UTC 2021 
   RealMemory=604160 AllocMem=0 FreeMem=603434 Sockets=64 Boards=1
   MemSpecLimit=8192
   State=ALLOCATED ThreadsPerCore=1 TmpDisk=3145728 Weight=1 Owner=N/A MCS_label=N/A
   Partitions=batch 
   BootTime=2021-11-05T03:19:23 SlurmdStartTime=2021-11-08T02:49:45
   LastBusyTime=2021-11-08T02:50:00
   CfgTRES=cpu=64,mem=590G,billing=64
   AllocTRES=cpu=64,mem=590G,billing=64
   CapWatts=n/a
   CurrentWatts=0 AveWatts=0
   ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s

and my jobscript looks like

#!/bin/bash
#SBATCH --mem=1G
#SBATCH --time=10:00:00
#SBATCH --gres=gpu:2g.10gb:1
#SBATCH -w mlerp-node05

hostname
nvidia-smi

but all the nvidia-caps are being added to devices.deny

Regards,
--
Dr Chris Hines

(In reply to Damien from comment #0)
> I'm having some difficulty with configuring MIG on my new cluster.
> Using Ubuntu 20.04 Slurm 21.08.3 MIG devices seem to be picked up by
> autodetect but do not appear to be added to the cgroup allow list for any
> steps. 
> 
> In particular my slurmd.log looks like
Damien, could you set SlurmdDebug=debug2 and restart the slurmd? That should give us more information.

Thanks,
-Michael

(In reply to Damien from comment #0)
> I hacked in an extra error statement to the cgroup plugin to find out that
> the string being written is "0 rwm" ... which is a little odd since I though
> it was supposed to be strings like "c NNN:MMM rwm"
I think that could occur if the dev_path passed into gres_device_major() in gres.c doesn't point to a device file... But I'm not sure how this can happen.

Created attachment 22178 [details]
slurmd log with debug2

Hi Michael,
Thanks @Damien for submitting this bug for me.

Attached is a slurmd log from 'systemctl start slurmd' to job cancel

Note that lines that include the word "wrote" are due to me hacking in 

        error("%s: wrote %s to cgroup %s: %m",
              __func__, content, file_path);

at line 401 of cgroup_common.c function common_file_write_content.

This tells me that slurmd is denying access to all my MIG devices on purpose, but its not clear why.

Is there a debug level to ask slurmd to report what gres it thinks its supposed to allocate to the job?

Hi Michael 
re the entries for "0 rwm" in devices.allow I think I understand this now. I think this occurred when I was using a quickly copied cgroup_allowed_devices_file that included entries for /dev/sd* which didn't exist on this node. It looks like it tried to identify the correct major and allow the device by default. So thats how you can force it to point to a device file that doesn't exist. Stupidity and quickly copy pasting when I should know better

Anyway I've since removed the cgroup_allowed_devices_file because it appears to be unnecessary.

Regards,
--
Chris Hines.

Hi Michael,
You can resolve this ticket now as I managed to figure out my mistake.
It seems that somewhere along the lines I got SelectType="select/linear" in my slurm.conf which was mangling the gres. I ended up with jobs like

JobId=80 JobName=test.sh
   UserId=chines(1001) GroupId=chines(1003) MCS_label=N/A
   Priority=13756 Nice=0 Account=chines QOS=normal
   JobState=RUNNING Reason=None Dependency=(null)
   Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0
   DerivedExitCode=0:0
   RunTime=00:00:05 TimeLimit=10:00:00 TimeMin=N/A
   SubmitTime=2021-11-09T04:30:13 EligibleTime=2021-11-09T04:30:13
   AccrueTime=2021-11-09T04:30:13
   StartTime=2021-11-09T04:30:13 EndTime=2021-11-09T14:30:13 Deadline=N/A
   SuspendTime=None SecsPreSuspend=0 LastSchedEval=2021-11-09T04:30:13 Scheduler=Backfill
   Partition=batch AllocNode:Sid=mlerp-login0:707018
   ReqNodeList=mlerp-node05 ExcNodeList=(null)
   NodeList=mlerp-node05
   BatchHost=mlerp-node05
   NumNodes=1 NumCPUs=64 NumTasks=1 CPUs/Task=1 ReqB:S:C:T=0:0:*:*
   TRES=cpu=64,node=1,billing=64
   Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*
   JOB_GRES=gpu:7g.40gb:0
     Nodes=mlerp-node05 CPU_IDs=0-63 Mem=0 GRES=gpu:7g.40gb:0(IDX:)
   MinCPUsNode=1 MinMemoryNode=20G MinTmpDiskNode=0
   Features=(null) DelayBoot=00:00:00
   OverSubscribe=NO Contiguous=0 Licenses=(null) Network=(null)
   Command=/home/chines/test.sh
   WorkDir=/home/chines
   StdErr=/home/chines/slurm-80.out
   StdIn=/dev/null
   StdOut=/home/chines/slurm-80.out
   Power=
   TresPerNode=gres:gpu:7g.40gb:1

Notice that TresPerNode lists 1 device, but JOB_GRES lists 0 devices.

Having switched my selecttype to cons_tres things behave much better

Thanks for your help

(In reply to chris.hines from comment #7)
> Anyway I've since removed the cgroup_allowed_devices_file because it appears
> to be unnecessary.
Yes, it is now unnecessary, as all devices are blacklisted by default, and only devices listed in Files= in gres.conf or autodetected will be allowed to be used by Slurm.

(In reply to chris.hines from comment #8)
> It seems that somewhere along the lines I got SelectType="select/linear" in
> my slurm.conf which was mangling the gres.
I saw you had SelectType="select/linear", but I didn't know that was on accident. We recommend using select/cons_tres, because that supports all the latest GPU features, and you can still get select/linear behavior with cons_tres by setting partitions to Oversubscribe=Exclusive.

Marking as resolved. Thanks!
-Michael