[slurm-users] Slurm powersave

[Davide DelVento]

I'm experimenting with slurm powersave and I have several questions. I'm following the guidance from https://slurm.schedmd.com/power_save.html and the great presentation from our own https://slurm.schedmd.com/SLUG23/DTU-SLUG23.pdf

I am running slurm 23.02.3

1) I'm not sure I fully understand ReconfigFlags=KeepPowerSaveSettings

The documentations ways that if set, an "scontrol reconfig" command will preserve the current state of SuspendExcNodes, SuspendExcParts and SuspendExcStates. Why would one *NOT* want to preserve that? What would happen if one does not (or does) have this setting? For now I'm using it, assuming that it means "if I run scontrol reconfig" don't shut off nodes that are up because I said so that they should be up in slurm.conf with those three options" --- but I am not clear if that is really what it says.

2) the PDF above says that the problem with nodes in down and drained state is solved in 23.02 but that does not appear to be the case. Before running my experiment, I had 

$ sinfo -R
REASON               USER      TIMESTAMP           NODELIST
Not responding       root      2023-09-13T13:14:50 node31
ECC memory errors    root      2023-08-26T07:21:04 node27

and after it became

$ sinfo -R
REASON               USER      TIMESTAMP           NODELIST
Not responding       root      2023-09-13T13:14:50 node31
none                 Unknown   Unknown             node27

And that despite having excluded drain'ed nodes as below:

--- a/slurm/slurm.conf
+++ b/slurm/slurm.conf
@@ -140,12 +140,15 @@ SlurmdLogFile=/var/log/slurm/slurmd.log
 #
 #
 # POWER SAVE SUPPORT FOR IDLE NODES (optional)
+SuspendProgram=/opt/slurm/poweroff
+ResumeProgram=/opt/slurm/poweron
+SuspendTimeout=120
+ResumeTimeout=240
 #ResumeRate=
+SuspendExcNodes=node[13-32]:2
+SuspendExcStates=down,drain,fail,maint,not_responding,reserved
+BatchStartTimeout=60
+ReconfigFlags=KeepPowerSaveSettings # not sure if needed: preserve current status when running "scontrol reconfig"

-PartitionName=compute512 Default=False Nodes=node[13-32] State=UP DefMemPerCPU=9196

+PartitionName=compute512 Default=False Nodes=node[13-32] State=UP DefMemPerCPU=9196 SuspendTime=600

so probably that's not solved? Anyway, that's a nuisance, not a deal breaker

3) The whole thing does not appear to be working as I intended. My understanding of the "exclude node" above should have meant that slurm should never attempt to shut off more than all idle nodes in that partition minus 2. Instead it shut them off all of them, and then tried to turn them back on:

$ sinfo | grep 512
compute512     up   infinite      1 alloc# node15
compute512     up   infinite      2  idle# node[14,32]
compute512     up   infinite      3  down~ node[16-17,31]
compute512     up   infinite      1 drain~ node27
compute512     up   infinite     12  idle~ node[18-26,28-30]
compute512     up   infinite      1  alloc node13

But again this is a minor nuisance which I can live with (especially if it happens only when I "flip the switch"), and I'm mentioning only in case it's a symptom of something else I'm doing wrong. I did try to use both the SuspendExcNodes=node[13-32]:2 syntax as it seem more reasonable to me (compared to the rest of the file, e.g. partitions definition) and the SuspendExcNodes=node[13\-32]:2 as suggested in the slurm powersave documentation. Behavior, exactly identical

4) Most importantly from the output above you may have noticed two nodes (actually three by the time I ran the command below) that slurm deemed down

$ sinfo -R
REASON               USER      TIMESTAMP           NODELIST
Not responding       root      2023-09-13T13:14:50 node31
reboot timed out     slurm     2023-10-04T14:51:28 node14
reboot timed out     slurm     2023-10-04T14:52:28 node15
reboot timed out     slurm     2023-10-04T14:49:58 node32
none                 Unknown   Unknown             node27

This can't be the case, the nodes are fine, and cannot have timed out while "rebooting", because for now my poweroff and poweron script are identical and literally a simple one-liner bash script doing almost nothing and the log file is populated correctly as I would expect 

echo "Pretending to $0 the following node(s): $1"  >> $log_file 2>&1

So I can confirm slurm invoked the script, but then waited for something (what? starting slurmd?) which failed to occur and marked the node down.  When I removed the suspend time from the partition to end the experiment, the other nodes went "magically" in production , without slurm calling my poweron script. Of course the nodes were never powered off, but slurm thought they were, so why it did not have the problem it id with the node which instead intentionally tried to power on?

Thanks for any light you can shed on these issues, particularly the last one!

[Ole Holm Nielsen]

Hi Davide,

On 10/4/23 23:03, Davide DelVento wrote:
> I'm experimenting with slurm powersave and I have several questions. I'm
> following the guidance from https://slurm.schedmd.com/power_save.html
> <https://slurm.schedmd.com/power_save.html> and the great presentation
> from our own https://slurm.schedmd.com/SLUG23/DTU-SLUG23.pdf

> <https://slurm.schedmd.com/SLUG23/DTU-SLUG23.pdf>

I presented that talk at SLUG'23 :-)

> I am running slurm 23.02.3
>
> 1) I'm not sure I fully understand ReconfigFlags=KeepPowerSaveSettings
> The documentations ways that if set, an "scontrol reconfig" command will
> preserve the current state of SuspendExcNodes, SuspendExcParts and
> SuspendExcStates. Why would one *NOT* want to preserve that? What would
> happen if one does not (or does) have this setting? For now I'm using it,
> assuming that it means "if I run scontrol reconfig" don't shut off nodes
> that are up because I said so that they should be up in slurm.conf with
> those three options" --- but I am not clear if that is really what it says.

As I understand it, the ReconfigFlags means that if you updated some
settings using scontrol, they will be lost when slurmctld is reconfigured,
and the settings from slurm.conf will be used in stead.

> 2) the PDF above says that the problem with nodes in down and drained
> state is solved in 23.02 but that does not appear to be the case. Before
> running my experiment, I had
>
> $ sinfo -R
> REASON               USER      TIMESTAMP           NODELIST
> Not responding       root      2023-09-13T13:14:50 node31
> ECC memory errors    root      2023-08-26T07:21:04 node27
>
> and after it became
>
> $ sinfo -R
> REASON               USER      TIMESTAMP           NODELIST
> Not responding       root      2023-09-13T13:14:50 node31
> none                 Unknown   Unknown             node27

Please use "sinfo -lR" so that we can see the node STATE.

> And that despite having excluded drain'ed nodes as below:
>
> --- a/slurm/slurm.conf
> +++ b/slurm/slurm.conf
> @@ -140,12 +140,15 @@ SlurmdLogFile=/var/log/slurm/slurmd.log
>  #
>  #
>  # POWER SAVE SUPPORT FOR IDLE NODES (optional)
> +SuspendProgram=/opt/slurm/poweroff
> +ResumeProgram=/opt/slurm/poweron
> +SuspendTimeout=120
> +ResumeTimeout=240
>  #ResumeRate=
> +SuspendExcNodes=node[13-32]:2
> +SuspendExcStates=down,drain,fail,maint,not_responding,reserved
> +BatchStartTimeout=60
> +ReconfigFlags=KeepPowerSaveSettings # not sure if needed: preserve
> current status when running "scontrol reconfig"
> -PartitionName=compute512 Default=False Nodes=node[13-32] State=UP
> DefMemPerCPU=9196
> +PartitionName=compute512 Default=False Nodes=node[13-32] State=UP
> DefMemPerCPU=9196 SuspendTime=600
>
> so probably that's not solved? Anyway, that's a nuisance, not a deal breaker

With my 23.02.5 the SuspendExcStates is working as documented :-)

> 3) The whole thing does not appear to be working as I intended. My
> understanding of the "exclude node" above should have meant that slurm
> should never attempt to shut off more than all idle nodes in that
> partition minus 2. Instead it shut them off all of them, and then tried to
> turn them back on:
>
> $ sinfo | grep 512
> compute512     up   infinite      1 alloc# node15
> compute512     up   infinite      2  idle# node[14,32]
> compute512     up   infinite      3  down~ node[16-17,31]
> compute512     up   infinite      1 drain~ node27
> compute512     up   infinite     12  idle~ node[18-26,28-30]
> compute512     up   infinite      1  alloc node13

I agree that 2 nodes from node[13-32] shouldn't be suspended, according to
SuspendExcNodes in the slurm.conf manual. I haven't tested this feature.

IMHO, "pretending" to power down nodes defies the logic of the Slurm
power_save plugin. Slurmctld expects suspended nodes to *really* power
down (slurmd is stopped). When slurmctld resumes a suspended node, it
expects slurmd to start up when the node is powered on. There is a
ResumeTimeout parameter which I've set to about 15-30 minutes in case of
delays due to BIOS updates and the like - the default of 60 seconds is WAY
too small!

Have you tried to experiment with the IPMI based power down/up method
explained in the above presentation? I'd appreciate independent testing
of my scripts in
https://github.com/OleHolmNielsen/Slurm_tools/tree/master/power_save :-)

Best regards,
Ole

[Davide DelVento]

Hi Ole,

Thanks for getting back to me.

> the great presentation
> from our own 

I presented that talk at SLUG'23 :-)

Yes! That's why I wrote "from our own", but perhaps these are local slangs where I live (and English is my second language)

 

> 1) I'm not sure I fully understand ReconfigFlags=KeepPowerSaveSettings 

As I understand it, the ReconfigFlags means that if you updated some

settings using scontrol, they will be lost when slurmctld is reconfigured,
and the settings from slurm.conf will be used in stead.

I see, so that applies to the case in which I change the (power) state of the nodes by scontrol.

 

> 2) the PDF above says that the problem with nodes in down and drained
> state is solved in 23.02 but that does not appear to be the case. Before
> running my experiment, I had
>
> $ sinfo -R
> REASON               USER      TIMESTAMP           NODELIST
> Not responding       root      2023-09-13T13:14:50 node31
> ECC memory errors    root      2023-08-26T07:21:04 node27
>
> and after it became
>
> $ sinfo -R
> REASON               USER      TIMESTAMP           NODELIST
> Not responding       root      2023-09-13T13:14:50 node31
> none                 Unknown   Unknown             node27

Please use "sinfo -lR" so that we can see the node STATE.

$ sinfo -lR
Thu Oct 05 07:08:18 2023
REASON               USER         TIMESTAMP           STATE  NODELIST
Not responding       root(0)      2023-09-13T13:14:50 down~  node31
none                 root(0)      Unknown             drain  node27

Somewhat it has now remembered that the user was root (it now shows that even with plain sinfo -R)

> so probably that's not solved? Anyway, that's a nuisance, not a deal breaker

With my 23.02.5 the SuspendExcStates is working as documented :-)

Okay so perhaps something happened between 23.02.3 and 23.02.5. I might need to sleuth in the ticketing system.

 

> 3) The whole thing does not appear to be working as I intended. My
> understanding of the "exclude node" above should have meant that slurm
> should never attempt to shut off more than all idle nodes in that
> partition minus 2. Instead it shut them off all of them, and then tried to
> turn them back on:
>
> $ sinfo | grep 512
> compute512     up   infinite      1 alloc# node15
> compute512     up   infinite      2  idle# node[14,32]
> compute512     up   infinite      3  down~ node[16-17,31]
> compute512     up   infinite      1 drain~ node27
> compute512     up   infinite     12  idle~ node[18-26,28-30]
> compute512     up   infinite      1  alloc node13

I agree that 2 nodes from node[13-32] shouldn't be suspended, according to
SuspendExcNodes in the slurm.conf manual.  I haven't tested this feature.

Good to know that an independent read of the manual is understood the same way as mine. If you don't use this feature, what do you do? Shutting off all idle nodes and leaving newly submitted jobs waiting for a boot? Or something else?

 

> But again this is a minor nuisance which I can live with

> 4) Most importantly from the output above you may have noticed two nodes
> (actually three by the time I ran the command below) that slurm deemed down
>

> So I can confirm slurm invoked the script, but then waited for something
> (what? starting slurmd?) which failed to occur and marked the node down. 
> When I removed the suspend time from the partition to end the experiment,
> the other nodes went "magically" in production , without slurm calling my
> poweron script. Of course the nodes were never powered off, but slurm
> thought they were, so why it did not have the problem it id with the node
> which instead intentionally tried to power on?

IMHO, "pretending" to power down nodes defies the logic of the Slurm
power_save plugin. 

And it is sure useless ;)

But I was using the suggestion from https://slurm.schedmd.com/power_save.html which says

You can also configure Slurm with programs that perform no action as SuspendProgram and ResumeProgram to assess the potential impact of power saving mode before enabling it.

 

Slurmctld expects suspended nodes to *really* power
down (slurmd is stopped).  When slurmctld resumes a suspended node, it
expects slurmd to start up when the node is powered on.  There is a
ResumeTimeout parameter which I've set to about 15-30 minutes in case of
delays due to BIOS updates and the like - the default of 60 seconds is WAY
too small!

Sure in fact I upped that to 4 minutes. Typically our nodes reboot in 3 minutes and will not update BIOS or OS automatically. Sometimes they become "hosed" and slower (firmware bug throttling CPU speed for no reason) but in that case better Slurm recognizes it and deems the node down. But in any case this is a moot point since the node is not going down....

 

Have you tried to experiment with the IPMI based power down/up method
explained in the above presentation?  I'd appreciate independent testing
of my scripts in
https://github.com/OleHolmNielsen/Slurm_tools/tree/master/power_save :-)

I have not, because I already have my time-tested script from the vendor of the cluster. Those accept slurm's syntax as the node(s) to shut off/on or query so it's extremely convenient, and I use them all the time (interactively) 

Thanks!

[Ole Holm Nielsen]

Hi Davide,

On 10/5/23 15:28, Davide DelVento wrote:
> IMHO, "pretending" to power down nodes defies the logic of the Slurm
> power_save plugin.
>
> And it is sure useless ;)
> But I was using the suggestion from
> https://slurm.schedmd.com/power_save.html
> <https://slurm.schedmd.com/power_save.html> which says
>
> You can also configure Slurm with programs that perform no action as

> *SuspendProgram* and *ResumeProgram* to assess the potential impact of

> power saving mode before enabling it.

I had not noticed the above sentence in the power_save manual before! So
I decided to test a "no action" power saving script, similar to what you
have done, applying it to a test partition. I conclude that "no action"
power saving DOES NOT WORK, at least in Slurm 23.02.5. So I opened a bug
report https://bugs.schedmd.com/show_bug.cgi?id=17848 to find out if the
documentation is obsolete, or if there may be a bug. Please follow that
bug to find out the answer from SchedMD.

What I *believe* (but not with 100% certainty) really happens with power
saving in the current Slurm versions is what I wrote yesterday:

> Slurmctld expects suspended nodes to *really* power
> down (slurmd is stopped).  When slurmctld resumes a suspended node, it
> expects slurmd to start up when the node is powered on.  There is a
> ResumeTimeout parameter which I've set to about 15-30 minutes in case of
> delays due to BIOS updates and the like - the default of 60 seconds is
> WAY too small!

I hope this helps,
Ole

[Davide DelVento]

In case it's useful to others: I've been able to get this working by having the "no action" script stop the slurmd daemon and start it *with the -b option*.