Denied reservation

Users/Site Administrators
1

Hello,

My environment:

  • OpenPBS version: 20.0.0
  • Queues with distinct priority levels: LowPriorityQ, workq, HighPriorityQ
  • Execution hosts with: 1, 2 or 3 gpus (ngpu)
  • Attributes:
    server:
    resv_enable = True
    acl_resv_host_enable = False
    acl_resv_group_enable = False
    acl_resv_user_enable = False
    vnodes:
    resv_enable = True

Problem:
LowPriorityQ has a large number of jobs requiring ngpu=1 and workq has one queued job that requires ngpu=3 and this job never run.

Possible solution:
Create a reservation for ngpu=3 to run the job, but PBS deny the reservation submission:
pbs_rsub -R 202205271600.00 -D 24:00:00 -W delete_idle_time=60 -l select=ngpus=3 -I 10
R5.cgpusub DENIED

Why the reservation submission was denied?

Thannk you in advance!

2

Nodes cannot be in use by other jobs to be reserved. If you want you can convert your difficult job into an asap reservation using pbs_rsub -Wqmove.

If you don’t like the time of the corresponding reservation then nodes are hogged by jobs with overly long wall time limits.

BTW: if you make all your ‘easy’ jobs topjob_ineligible and set strict_ordering and backfill enabled and have a decent backfill depth the system will on each cycle calendar the N top jobs to avoid pushing them into the future, I.e. it will make such asap reservations during each cycle. But it relies on a decent job sort order and realistic wall time (or soft wall time) limits.

3

Hi Alexis,

The strict order with backfill solution worked great for me!

Thank you for your help!