[slurm-users] ticking time bomb? launching too many jobs in parallel
Guillaume Perrault Archambault
gperr050 at uottawa.ca
Tue Aug 27 17:16:48 UTC 2019
Thanks a lot for your recommendations.
I'll do my best to address your three points in line. I hope I've
understood you correctly, please correct me if i've misunderstood parts.
"1) If you can, either use xargs or parallel to do the forking so you can
limit the number of simultaneous submissions "
I think that's a great idea! I will look into it
"2) I have yet to see where it is a good idea to have many separate jobs
when using an array can work.
If you can prep up a proper input file for a script, a single
submission is all it takes. Then you can control how many are currently
running (MaxArrayTask) and can change that to scale up/down"
The reason I don't use job arrays is to be able limit the number of jobs
per users (as explained, not sure how well, in my reply to Paul's second
Perhaps it doesn't fall into the "good idea" because there are good methods
to limit jobs run per user at the job array level; I'm still trying to
figure out if there are.
"Here is where you may want to look into slurmdbd and sacct
Then you can create a qos that has MaxJobsPerUser to limit the total number
running on a per-user basis: https://slurm.schedmd.com/resource_limits.html"
Correct me if I'm wrong, but this an admin-side solution, am I right? I
cannot create a qos as a user?
I'm trying to implement a user-side solution so that the user can limit the
number of jobs.
I use two clusters (or two sets of clusters to be exact), and this is
valuable on both for (slightly) different reasons:
1) A small-ish cluster of about 140 GPUs that does not use fair share, and
where they do not want to set a hard limit on number of jobs per users
because it may change over time, especially around conference dedalines.
But they do want users to have the ability to self-manage the number of
jobs they run, hence my script doing this from the user side
2) A large cluster that uses fairshare. But the fairshare is at the
association level. So within an association, we want users to be able to
limit the number of jobs they run, similar to cluste 1.
I hope I've understood your suggestions and replied on-topic. My apologies
if I've misunderstood anythig.
On Tue, Aug 27, 2019 at 12:53 PM Brian Andrus <toomuchit at gmail.com> wrote:
> Here is where you may want to look into slurmdbd and sacct
> Then you can create a qos that has MaxJobsPerUser to limit the total
> number running on a per-user basis:
> Brian Andrus
> On 8/27/2019 9:38 AM, Guillaume Perrault Archambault wrote:
> Hi Paul,
> Your comment confirms my worst fear, that I should either implement job
> arrays or stick to a sequential for loop.
> My problem with job arrays is that, as far as I understand them, they
> cannot be used with singleton to set a max job limit.
> I use singleton to limit the number of jobs a user can be running at a
> time. For example if the limit is 3 jobs per user and the user launches 10
> jobs, the sbatch submissions via my scripts may look this:
> sbatch --job-name=job1 [OPTIONS SET1] Dependency=singleton my.sbatch
> sbatch --job-name=job2 [OTHER SET1] Dependency=singleton my.sbatch
> sbatch --job-name=job3 [OTHER SET1] Dependency=singleton my.sbatch
> sbatch --job-name=job1 [OTHER SET1 Dependency=singleton my.sbatch
> sbatch --job-name=job2 [OTHER SET1 ] Dependency=singleton my.sbatch
> sbatch --job-name=job3 [OTHER SET2] Dependency=singleton my.sbatch2
> sbatch --job-name=job1 [OTHER SET2] Dependency=singleton my.sbatch2
> sbatch --job-name=job2 [OTHER SET2 ] Dependency=singleton my.sbatch2
> sbatch --job-name=job2 [OTHER SET2 ] Dependency=singleton my.sbatch2
> sbatch --job-name=job1 [OTHER SET2 ] Dependency=singleton my.sbatch 2
> This way, at most 3 jobs will run at a time (ie a job with name job1, a
> job with name job2, and job with name job3).
> Notice that my example has two option sets provided to sbatch, so the
> example would be suitable for conversion to two Job Arrays.
> This is the problem I can't obercome.
> In the job array documentation, I see
> A maximum number of simultaneously running tasks from the job array may be
> specified using a "%" separator. For example "--array=0-15%4" will limit
> the number of simultaneously running tasks from this job array to 4.
> But this '%' separator cannot specify a max number of tasks over two (or
> more) separate job arrays, as far as I can tell.
> And the job array element names cannot be made to modulo rotate in the way
> they do in my above example.
> Perhaps I need to play more with job arrays, and try harder to find a
> solution to limit number of jobs across multiple arrays. Or ask this
> question in a separate post, since it's a bit off topic.
> In any case, thanks so much for answer my question. I think it answer my
> original post perfectly :)
> On Tue, Aug 27, 2019 at 10:08 AM Paul Edmon <pedmon at cfa.harvard.edu>
>> At least for our cluster we generally recommend that if you are
>> submitting large numbers of jobs you either use a job array or you just for
>> loop over the jobs you want to submit. A fork bomb is definitely not
>> recommended. For highest throughput submission a job array is your best
>> bet as in one submission it will generate thousands of jobs which then the
>> scheduler can handle sensibly. So I highly recommend using job arrays.
>> -Paul Edmon-
>> On 8/27/19 3:45 AM, Guillaume Perrault Archambault wrote:
>> Hi Paul,
>> Thanks a lot for your suggestion.
>> The cluster I'm using has thousands of users, so I'm doubtful the admins
>> will change this setting just for me. But I'll mention it to the support
>> team I'm working with.
>> I was hoping more for something that can be done on the user end.
>> Is there some way for the user to measure whether the scheduler is in RPC
>> saturation? And then if it is, I could make sure my script doesn't launch
>> too many jobs in parallel.
>> Sorry if my question is too vague, I don't understand the backend of the
>> SLURM scheduler too well, so my questions are using the limited terminology
>> of a user.
>> My concern is just to make sure that my scripts don't send out more
>> commands (simultaneously) than the scheduler can handle.
>> For example, as an extreme scenario, suppose a user forks off 1000 sbatch
>> commands in parallel, is that more than the scheduler can handle? As a
>> user, how can I know whether it is?
>> On Mon, Aug 26, 2019 at 10:15 AM Paul Edmon <pedmon at cfa.harvard.edu>
>>> We've hit this before due to RPC saturation. I highly recommend using
>>> max_rpc_cnt and/or defer for scheduling. That should help alleviate this
>>> -Paul Edmon-
>>> On 8/26/19 2:12 AM, Guillaume Perrault Archambault wrote:
>>> I wrote a regression-testing toolkit to manage large numbers of SLURM
>>> jobs and their output (the toolkit can be found here
>>> <https://github.com/gobbedy/slurm_simulation_toolkit/> if anyone is
>>> To make job launching faster, sbatch commands are forked, so that
>>> numerous jobs may be submitted in parallel.
>>> We (the cluster admin and myself) are concerned that this may cause
>>> unresponsiveness for other users.
>>> I cannot say for sure since I don't have visibility over all users of
>>> the cluster, but unresponsiveness doesn't seem to have occurred so far.
>>> That being said, the fact that it hasn't occurred yet doesn't mean it won't
>>> in the future. So I'm treating this as a ticking time bomb to be fixed asap.
>>> My questions are the following:
>>> 1) Does anyone have experience with large numbers of jobs submitted in
>>> parallel? What are the limits that can be hit? For example is there some
>>> hard limit on how many jobs a SLURM scheduler can handle before blacking
>>> out / slowing down?
>>> 2) Is there a way for me to find/measure/ping this resource limit?
>>> 3) How can I make sure I don't hit this resource limit?
>>> From what I've observed, parallel submission can improve submission time
>>> by a factor at least 10x. This can make a big difference in users'
>>> For that reason I would like to keep the option of launching jobs
>>> sequentially as a last resort.
>>> Thanks in advance.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the slurm-users