[slurm-users] ticking time bomb? launching too many jobs in parallel

Paul Edmon pedmon at cfa.harvard.edu
Fri Aug 30 14:22:33 UTC 2019

A QoS is probably your best bet.  Another variant might be MCS, which 
you can use to help reduce resource fragmentation.  For limits though 
QoS will be your best bet.

-Paul Edmon-

On 8/30/19 7:33 AM, Steven Dick wrote:
> It would still be possible to use job arrays in this situation, it's
> just slightly messy.
> So the way a job array works is that you submit a single script, and
> that script is provided an integer for each subjob.  The integer is in
> a range, with a possible step (default=1).
> To run the situation you describe, you would have to predetermine how
> many of each test you want to run (i.e., you coudln't dynamically
> change the number of jobs that run within one array)., and a master
> script would map the integer range to the job that was to be started.
> The most trivial way to do it would be to put the list of regressions
> in a text file and the master script would index it by line number and
> then run the appropriate command.
> A more complex way would be to do some math (a divide?) to get the
> script name and subindex (modulus?) for each regression.
> Both of these would require some semi-advanced scripting, but nothing
> that couldn't be cut and pasted with some trivial modifications for
> each job set.
> As to the unavailability of the admin ...
> An alternate approach that would require the admin's help would be to
> come up with a small set of alocations (e.g., 40 gpus, 80 gpus, 100
> gpus, etc.) and make a QOS for each one with a gpu limit (e.g.,
> maxtrespu=gpu=40 ) Then the user would assign that QOS to the job when
> starting it to set the overall allocation for all the jobs.  The admin
> woudln't need to tweak this except once, you just pick which tweak to
> use.
> On Fri, Aug 30, 2019 at 2:36 AM Guillaume Perrault Archambault
> <gperr050 at uottawa.ca> wrote:
>> Hi Steven,
>> Thanks for taking the time to reply to my post.
>> Setting a limit on the number of jobs for a single array isn't sufficient because regression-tests need to launch multiple arrays, and I would need a job limit that would take effect over all launched jobs.
>> It's very possible I'm not understand something. I'll lay out a very specific example in the hopes you can correct me if I've gone wrong somewhere.
>> Let's take the small cluster with 140 GPUs and no fairshare as an example, because it's easier for me to explain.
>> The users, who all know each other personally and interact via chat, decide on a daily basis how many jobs each user can run at a time.
>> Let's say today is Sunday (hypothetically). Nobody is actively developing today, except that user 1 has 10 jobs running for the entire weekend. That leaves 130 GPUs unused.
>> User 2, whose jobs all run on 1 GPU decides to run a regression test. The regression test comprises of 9 different scripts each run 40 times, for a grand total of 360 jobs. The duration of the scripts vary from 1 and 5 hours to complete, and the jobs take on average 4 hours to complete.
>> User 2 gets the user group's approval (via chat) to use 90 GPUs (so that 40 GPUs will remain for anyone else wanting to work that day).
>> The problem I'm trying to solve is this: how do I ensure that user 2 launches his 360 jobs in such a way that 90 jobs are in the run state consistently until the regression test is finished?
>> Keep in mind that:
>> limiting each job array to 10 jobs is inefficient: when the first job array finishes (long before the last one), only 80 GPUs will be used, and so on as other arrays finish
>> the admin is not available, he cannot be asked to set a hard limit of 90 jobs for user 2 just for today
>> I would be happy to use job arrays if they allow me to set an overarching job limit across multiple arrays. Perhaps this is doable. Admttedly I'm working on a paper to be submitted in a few days, so I don't have time to test jobs arrays thoroughly, but I will try out job arrays more thoroughly once I've submitted my paper (ie after sept 5).
>> My solution, for now, is to not use job arrays. Instead, I launch each job individually, and I use singleton (by launching all jobs with the same 90 unique names) to ensure that exactly 90 jobs are run at a time (in this case, corresponding to 90 GPUs in use).
>> Side note: the unavailability of the admin might sound contrived by picking Sunday as an example, but it's in fact very typical. The admin is not available:
>> on weekends (the present example)
>> at any time outside of 9am to 5pm (keep in mind, this is a cluster used by students in different time zones)
>> any time he is on vacation
>> anytime the he is looking after his many other responsibilities. Constantly setting user limits that change on a daily basis would be too much too ask.
>> I'd be happy if you corrected my misunderstandings, especially if you could show me how to set a job limit that takes effect over multiple job arrays.
>> I may have very glaring oversights as I don't necessarily have a big picture view of things (I've never been an admin, most notably), so feel free to poke holes at the way I've constructed things.
>> Regards,
>> Guillaume.
>> On Fri, Aug 30, 2019 at 1:22 AM Steven Dick <kg4ydw at gmail.com> wrote:
>>> This makes no sense and seems backwards to me.
>>> When you submit an array job, you can specify how many jobs from the
>>> array you want to run at once.
>>> So, an administrator can create a QOS that explicitly limits the user.
>>> However, you keep saying that they probably won't modify the system
>>> for just you...
>>> That seems to me to be the perfect case to use array jobs and tell it
>>> how many elements of the array to run at once.
>>> You're not using array jobs for exactly the wrong reason.
>>> On Tue, Aug 27, 2019 at 1:19 PM Guillaume Perrault Archambault
>>> <gperr050 at uottawa.ca> wrote:
>>>> The reason I don't use job arrays is to be able limit the number of jobs per users

More information about the slurm-users mailing list