[slurm-users] Slurm powersave

Thu Oct 5 13:28:59 UTC 2023

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!
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.schedmd.com/pipermail/slurm-users/attachments/20231005/badc60af/attachment.htm>