[slurm-users] Scheduler does not reserve resources
jeremy.lapierre at uni-saarland.de
Wed Jan 19 09:05:24 UTC 2022
We indeed have overlooked this. The problem is that in general our jobs
need more than 2 days of resources, that's why we select a wall time in
the batch scripts equal to the max wall time allowed by the partition.
One thing we could try is to set the wall time at ~46h for the "light"
jobs in the batch scripts and let 48h for the "heavy" jobs, this way not
all jobs will have the same time limit.
Configuring node list for "light" and "heavy" jobs could do the trick. 2
things that could probably be a problem then are (i) even "heavy" jobs
having very low Priority would have access to resources at the expense
of "light" jobs with higher priority and (ii) regular intervention would
be needed. But maybe there is no other solution.
I thank you a lot for your inputs !
Am 2022-01-19 01:46, schrieb Rodrigo Santibáñez:
> Hi Jeremy,
> If all jobs have the same time limit, backfill is impossible. The
> documentation says: "Effectiveness of backfill scheduling is dependent
> upon users specifying job time limits, otherwise all jobs will have the
> same time limit and backfilling is impossible". I don't know to
> overcome that...
> However, without changing SchedulerType, you could hold pending jobs
> except for the job you want to execute, then release all jobs when the
> desired job is allocated. Also, you could define a node or list of
> nodes available for all jobs excluding nodes for the job of interest,
> then remove the configuration when the latter is allocated. I preferred
> to do the second because the "heavy" job and the "light" jobs will be
> allocated, and I have not to be aware of the queue outside office hours
> (Again, easier to do in a low utilized cluster).
> About "PLANNED", I wasn't aware, and it is a feature of SLURM 21.08.
> Could be that why you don't see it in your cluster?
> On Mon, Jan 17, 2022 at 2:02 PM Jérémy Lapierre
> <jeremy.lapierre at uni-saarland.de> wrote:
>> Hi Rodrigo and Rémi,
>>> I had a similar behavior a long time ago, and I decided to set
>>> SchedulerType=sched/builtin to empty X
>>> nodes of jobs and execute that high-priority job requesting more than
>>> one node. It is not ideal, but the
>>> cluster has low load, so a user that requests more than one node
>>> doesn't delay too much the execution
>>> of other's jobs.
>> I don't think this would be ideal in our case as we have heavy loads.
>> Also I'm not sure if you mean that we should switch to
>> SchedulerType=sched/builtin permanently or just the time needed for
>> the jobs causing problem to be allocated ? Also we have some other
>> experiences on another cluster and slurm should normally reserve
>> resources we think.
>>> Backfilling doesn't delay the scheduled start time of higher priority
>>> but at least they must have a scheduled start time.
>>> Did you check the start time of your job pending with Resources
>>> reason? eg.
>>> with `scontrol show job <id> | grep StartTime`.
>> Yes, the scheduled start time have been checked as well, and this time
>> is updated through time such that jobs asking for 1/4 of a node can
>> run on a freshly-free-1/4th-node. This is why I'm saying that the jobs
>> asking for several nodes (tested with 2 nodes here) are pending
>> forever. It is like slurm never wants to have unused resources (which
>> also makes sense, but how can we satisfy "heavy" resources request
>> then ?). On another cluster using slurm, I know that slurm reserves
>> nodes and the node state of those reserved nodes becomes "PLANNED" (or
>> plnd), this way jobs requesting for more resources than available at
>> the time of submission can later be satisfied. This never happens on
>> the cluster which is causing issues.
>>> Sometimes Slurm is unable to define the start time of a pending job.
>>> typical reason is the absence of timelimit on the running jobs.
>>> In t his case Slurm is unable to define when the running jobs are
>>> when the next highest priority job can start and eventually unable to
>>> if lower priority jobs actually delay higher priority jobs.
>> Yes we always set up the time limit of our jobs to the max time limit
>> allowed by the partition.
>> Thanks for your help,
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the slurm-users