[slurm-users] Implementing a "soft" wall clock limit

[Davide DelVento via slurm-users]

In the institution where I work, so far we have managed to live without mandatory wallclock limits (a policy decided well before I joined the organization), and that has been possible because the cluster was not very much utilized.

Now that is changing, with more jobs being submitted and those being larger ones. As such I would like to introduce wallclock limits to allow slurm to be more efficient in scheduling jobs, including with backfill.

My concern is that this user base is not used to it and therefore I want to make it easier for them, and avoid common complaints. I anticipate one of them would be "my job was cancelled even though there were enough nodes idle and no other job in line after mine" (since the cluster utilization is increasing, but not yet always full like it has been at most other places I know).

So my question is: is it possible to implement "soft" wallclock limits in slurm, namely ones which would not be enforced unless necessary to run more jobs? In other words, is it possible to change the pre-emptability of a job only after some time has passed? I can think of some ways to hack this functionality myself with some cron or at jobs, and that might be easy enough to do, but I am not sure I can make it robust enough to cover all situations, so I'm looking for something either slurm-native or (if external solution) field-tested by someone else already, so that at least the worst kinks have been already ironed out.

Thanks in advance for any suggestions you may provide!

[Loris Bennett via slurm-users]

Hi Davide,

Davide DelVento via slurm-users

We just have a default wallclock limit of 14 days, but we also have QOS
with shorter wallclock limits but with higher priorities, albeit with
for fewer jobs and resources:

$ sqos
Name Priority MaxWall MaxJobs MaxSubmit MaxTRESPU
---------- ---------- ----------- ------- --------- --------------------
hiprio 100000 03:00:00 50 100 cpu=128,gres/gpu=4
prio 1000 3-00:00:00 500 1000 cpu=256,gres/gpu=8
standard 0 14-00:00:00 2000 10000 cpu=768,gres/gpu=16

We also have a page of documentation which explains how users can profit
from backfill. Thus users have a certain incentive to specify a shorter
wallclock limit, if they can.

'sqos' is just an alias for

sacctmgr show qos format=name,priority,maxwall,maxjobs,maxsubmitjobs,maxtrespu%20

Cheers,

Loris

--
Dr. Loris Bennett (Herr/Mr)
FUB-IT, Freie Universität Berlin

--
slurm-users mailing list -- slurm...@lists.schedmd.com
To unsubscribe send an email to slurm-us...@lists.schedmd.com

[Davide DelVento via slurm-users]

Thanks Loris,

Am I correct if reading in between the lines you're saying: rather than going on with my "soft" limit idea, just use the regular hard limits, being generous with the default and providing user education instead? In fact that is an alternative approach that I am considering too.

[Loris Bennett via slurm-users]

Hi Davide,

Davide DelVento <davide....@gmail.com> writes:

> Thanks Loris,
>
> Am I correct if reading in between the lines you're saying: rather than going on with my "soft" limit idea, just use the regular hard limits, being generous with
> the default and providing user education instead? In fact that is an alternative approach that I am considering too.

Yes. We in fact never get anyone complaining about jobs being cancelled
due to having reached their time-limit, even though other resources were
idle. Having said that, we do occasionally extend the time-limit for
individual jobs, when requested. We also don't pre-empt any jobs.

Apart from that, I imaging implementing your 'soft' limits robustly
might be quite challenging and/or time-consuming, as I am not aware that
Slurm has anything like that built in.

Cheers,

Loris

[Ansgar Esztermann-Kirchner via slurm-users]

Hi Davide,

I think it should be possible to emulate this via preemption: if you
set PreemptMode to CANCEL, a preempted job will behave just as if it
reached the end of its wall time. Then, you can use PreemptExemptTime
as your soft wall time limit -- the job will not be preempted before
PreemptExemptTime has passed.

See https://slurm.schedmd.com/preempt.html


Best,

A.

--
Ansgar Esztermann
Sysadmin Dep. Theoretical and Computational Biophysics
https://www.mpinat.mpg.de/person/11315/3883774

[Davide DelVento via slurm-users]

Hi Ansgar,

This is indeed what I was looking for: I was not aware of PreemptExemptTime.

From my cursory glance at the documentation, it seems that PreemptExemptTime is QOS-based and not job based though. Is that correct? Or could it be set per-job, perhaps on a prolog/submit lua script? I'm thinking that the user could use the regular wallclock limit setting in slurm and the script could remove that and use it to set the PreemptExemptTime.

Thanks,
Davide

[Ansgar Esztermann-Kirchner via slurm-users]

On Thu, Jun 12, 2025 at 04:52:24AM -0600, Davide DelVento wrote:
> Hi Ansgar,
>
> This is indeed what I was looking for: I was not aware of PreemptExemptTime.
>
> From my cursory glance at the documentation, it seems
> that PreemptExemptTime is QOS-based and not job based though. Is that
> correct? Or could it be set per-job, perhaps on a prolog/submit lua script?

Yes, that's correct.
I guess you could create a bunch of QOS with different
PremptExemptTimes and then let the user select one (or indeed select
it from lua) but as far as I know, there is no way to set arbitrary
per-job values.

[Davide DelVento via slurm-users]

Sounds good, thanks for confirming it. 

Let me sleep on it wrt the "too many" QOS, or think if I should ditch this idea.

If I'll implement it, I'll post in this conversation details on how I did it.

Cheers

[Prentice Bisbal via slurm-users]

I think the idea of having a generous default timelimit is the wrong way to go. In fact, I think any defaults for jobs are a bad way to go.  The majority of your users will just use that default time limit, and backfill scheduling will remain useless to you.

Instead, I recommend you use your job_submit.lua to reject all jobs that don't have a wallclock time and print out a helpful error message to inform users they now need to specify a wallclock time, and provide a link to documentation on how to do that.

Requiring users to specify a time limit themselves does two things:

1. It reminds them that it's important to be conscious of timelimits when submitting jobs

2. If a job is killed before it's done and all the progress is lost because the job wasn't checkpointing, they can't blame you as the admin.

If you do this, it's easy to get the users on board by first providing useful and usable documentation on why timelimits are needed and how to set them. Be sure to hammer home the point that effective timelimits can lead to their jobs running sooner, and that effective timelimits can increase cluster efficiency/utilization, helping them get a better return on their investment (if they contribute to the clusters cost) or they'll get more science done. I like to frame it that accurate wallclock times will give them a competitive edge in getting their jobs running before other cluster users. Everyone likes to think what they're doing will give them an advantage!

My 4 cents (adjusted for inflation).

Prentice

-- 
Prentice Bisbal
HPC Systems Engineer III
Computational & Information Systems Laboratory (CISL)
NSF National Center for Atmospheric Research (NSF NCAR)
https://www.cisl.ucar.edu
https://ncar.ucar.edu

[Loris Bennett via slurm-users]

Hi Prentice,

Prentice Bisbal via slurm-users

<slurm...@lists.schedmd.com> writes:

> I think the idea of having a generous default timelimit is the wrong way to go. In fact, I think any defaults for jobs are a bad way to go. The majority of your
> users will just use that default time limit, and backfill scheduling will remain useless to you.

Horses for courses, I would say. We have a default time of 14 days, but
because we also have QoS with increased priority, but shorter time
limits, there is still an incentive for users to set the time limit
themselves. So currently we have around 900 jobs running, only 100 of
which are using the default time limit. Many of these will be
long-running Gaussian jobs and will indeed need the time.

> Instead, I recommend you use your job_submit.lua to reject all jobs that don't have a wallclock time and print out a helpful error message to inform users they
> now need to specify a wallclock time, and provide a link to documentation on how to do that.
>
> Requiring users to specify a time limit themselves does two things:
>
> 1. It reminds them that it's important to be conscious of timelimits when submitting jobs

This is a good point. We use 'jobstats', which provides information
after a job has completed, about run time relative to time limit,
amongst other things, although unfortunately many people don't seem to
read this. However, even if you do force people to set a time limit,
they can still choose not to think about it and just set the maximum.

> 2. If a job is killed before it's done and all the progress is lost because the job wasn't checkpointing, they can't blame you as the admin.

I don't really understand this point. The limit is just the way it is,
just as we have caps on the total number of cores or GPUs the jobs given
user can use at any one time. Up to now no-one has blamed us for this.

> If you do this, it's easy to get the users on board by first providing useful and usable documentation on why timelimits are needed and how to set them. Be
> sure to hammer home the point that effective timelimits can lead to their jobs running sooner, and that effective timelimits can increase cluster
> efficiency/utilization, helping them get a better return on their investment (if they contribute to the clusters cost) or they'll get more science done. I like to
> frame it that accurate wallclock times will give them a competitive edge in getting their jobs running before other cluster users. Everyone likes to think what
> they're doing will give them an advantage!

I agree with all this and this is also what we also try to do. The only
thing I don't concur with is your last sentence. In my experience, as
long as things work, users will in general not give a fig about whether
they are using resources efficiently. Only when people notice a delay
in jobs starting do they become more aware about it and are prepared to
take action. It is particularly a problem with new users, because
fairshare means that their jobs will start pretty quickly, no matter how
inefficiently they have configured them. Maybe we should just give new
users fewer share initially and only later bump them up to some standard
value.

Cheers,

Loris

--
Dr. Loris Bennett (Herr/Mr)
FUB-IT, Freie Universität Berlin

[Davide DelVento via slurm-users]

This conversation is drifting a bit away from my initial questions and covering various other related topics. In fact I do agree with almost everything written in the last few messages. However, that is somewhat orthogonal to my initial request, which I now understand has the answer "not possible with slurm configuration, possible with ugly hacks which are probably error prone and not worth the hassle". Just for the sake of the discussion (since I'm enjoying hearing the various perspectives) I'll restate my request and why I think slurm does not support this need.

Most clusters have very high utilization all the time. This is good for ROI etc but annoying to users. Forcing users to specify a firm wallclock limit helps slurm make good scheduling decisions, which keep utilization (ROI, etc) high and minimizes wait time for everybody.

At the place where I work there is a quite different situation: there are moments of high pressure and long wait, and there are moments in which its utilization drops under 50% and sometimes even under 25% (e.g. during long weekends). We can have a discussion about it, but the bottom line is that management (ROI, etc) is fine with it, so that's the way it is. This circumstance, I agree, is quite peculiar and not shared by any other place I worked before or where I ever had an account and saw how things were, but that is what it is. In this circumstance it feels at least silly and perhaps even extremely wasteful and annoying to let slurm cancel jobs at their wallclock limit without considering other context. I mean, imagine a user with a weeklong job who estimated a 7 day wallclock limit and "for good measure" requested 8 days, but then the job would actually take 9 days. Imagine that the 8th day happened in the middle of on a long weekend when utilization was 25% and there was not a single other job pending. Maybe this job is a one-off experiment quickly cobbled together to test one thing, so it's not a well-designed piece of code and does not have checkpoint-restart capabilities. Why enforce the wallclock limit in that situation?

The way around this problem in the past was to simply not make the wallclock limit mandatory (which was decided by my predecessor, who has now left). That worked, only because the cluster was not in a very good usability status so most people avoided it anyway and there seldom was a long line of jobs pending in the queue, so slurm did not need to work very hard to schedule things. Now that I've improved the usability situation, this has become a problem, because utilization has become much higher. Perhaps in a short time people will learn to plan ahead and submit more jobs and fill the machine up during the weekends too (I'm working on user education towards that), and if that happens, it will make the above dilemma go away. But for now I have it.

I'm still mulling on how to best proceed. Maybe just force the users to set a wallclock limit and live with it.

Here is another idea that just came to me. Does slurm have a "global" switch to turn on/off cancelling jobs hitting their wallclock limit? If so, I could have a cron job checking if there are pending jobs in the queue and if not shut it off, and if so turn it on. Granted, that may be sloppy (e.g. one job pending for one resources causing the cancelling of jobs using other resources) but it's something and it'd be easy to implement compared to the turn on/off pre-emption as discussed in a previous message.

Great conversation folks, enjoying reading the various perspectives at different sites!

[Michael DiDomenico via slurm-users]

there is one reason to have a hard time limit. garbage collection.
if your institution is like mine, visiting academics often come and go
and many full-timers are "forgetful" of what they startup. at some
point someone has to clean all that up

[Smith, Sebastian via slurm-users]

Another 4 cents:

I think automatically increasing job time limits, or otherwise disabling job termination due to time, will cause you headaches down the road.  Coupling Slurm’s behavior to the effort of the cluster, or other state, will be difficult to communicate to users because the behavior of their jobs becomes non-deterministic.  You’ll answer a lot of questions that start, “this job completed the last time I ran it…”.  And, you’ll have to evaluate the state of the system near the job time limit to understand what happened (write good logs!!).  I’d avoid playing the detective at that, recurring, crime scene…

 

I forced my users to specify time limits and they quickly adapted:

 

`JobSubmitPlugins=require_timelimit`

 

Good luck!

 

Sebastian

 

-- 

 

 

>> >  https://urldefense.com/v3/__https://www.mpinat.mpg.de/person/11315/3883774__;!!NuzbfyPwt6ZyPHQ!pzCg16AzlVsRop9e83uUcbZt-GVO2XTyrusyMX9GljKv3d9nJtQDwTCAyGKAkgH4Ov1hF5BZIMqK0idZxv4Cu-WFy5yFuD5KsH9Nrhw$[mpinat[.]mpg[.]de]

>> >

>> >

>> > --

>> > Prentice Bisbal

>> > HPC Systems Engineer III

>> > Computational & Information Systems Laboratory (CISL)

>> > NSF National Center for Atmospheric Research (NSF NCAR)

>> > https://urldefense.com/v3/__https://www.cisl.ucar.edu__;!!NuzbfyPwt6ZyPHQ!pzCg16AzlVsRop9e83uUcbZt-GVO2XTyrusyMX9GljKv3d9nJtQDwTCAyGKAkgH4Ov1hF5BZIMqK0idZxv4Cu-WFy5yFuD5K_CSMT48$[cisl[.]ucar[.]edu]

>> > https://urldefense.com/v3/__https://ncar.ucar.edu__;!!NuzbfyPwt6ZyPHQ!pzCg16AzlVsRop9e83uUcbZt-GVO2XTyrusyMX9GljKv3d9nJtQDwTCAyGKAkgH4Ov1hF5BZIMqK0idZxv4Cu-WFy5yFuD5KArhlxNM$[ncar[.]ucar[.]edu]

>> --

>> Dr. Loris Bennett (Herr/Mr)

>> FUB-IT, Freie Universität Berlin

>> 

>> 

>> --

>> slurm-users mailing list -- slurm...@lists.schedmd.com

>> To unsubscribe send an email to slurm-us...@lists.schedmd.com

> 

> 

> --

> slurm-users mailing list -- slurm...@lists.schedmd.com

> To unsubscribe send an email to slurm-us...@lists.schedmd.com

-- 

slurm-users mailing list -- slurm...@lists.schedmd.com

To unsubscribe send an email to slurm-us...@lists.schedmd.com

CONFIDENTIALITY NOTICE: This e-mail, including any attachments, is for the sole use of the intended recipient(s) and may contain confidential and privileged information protected by law. Any unauthorized review, use, disclosure or distribution is prohibited. If you are not the intended recipient, please contact the sender by reply e-mail and destroy all copies of the original message.

[Christopher Samuel via slurm-users]

On 6/20/25 2:36 pm, Smith, Sebastian via slurm-users wrote:

> I forced my users to specify time limits and they quickly adapted:

You can also set a default (and maximum) time limit per partition, our
default time limits are set to 10 minutes, QOS's have limits in between
the partition default and the partition maximum.

All the best,
Chris
--
Chris Samuel : http://www.csamuel.org/ : Berkeley, CA, USA

[Prentice Bisbal via slurm-users]

I'm  the one to blame for taking this conversation off target. Sorry about that!

Unfortunately, people are hard, and I don't think any amount of technology will ever fix that. 😉

Thanks for explaining your situation, it's certainly different than what most of us see. I would say you need to plan for growth (i.e., a busy cluster). It sounds like you're already heading that way now that you've already fixed the usability issue(s). Every time you make a policy change, it takes effort to get the word out and retrain/recondition your users to adapt to the change, so my  advice is whatever policy you go with now, try to choose one that would survive at least a couple of increased cluster usage at the rate your seeing it increase now.

I take the "if you build it, they will come" attitude - if you design your cluster to handle a lot of traffic, it will!

imagine a user with a weeklong job who estimated a 7 day wallclock limit and "for good measure" requested 8 days, but then the job would actually take 9 days.

As much as I advocate for accurate timelimits, I always tell my users to specify a bit more, with a goal of 10-15% over.  If they're not sure how long the job will run, or they have low confidence in predicting run time then they shouldn't be afraid to be be more generous with their estimates, and as they run more jobs, they'll get a better feel for predicting run time. Having a job get killed 5 minutes before it finishes after running for 72 hours is also wasteful of computer time. It's a balancing act. 

Yes there's checkpointing, but that's way outside the scope of this conversation.

Prentice