Pegasus Job Queues

Pegasus queues are organized using limits like job size, job length, job purpose, and project. In general, users run jobs on Pegasus with equal resource shares. A user’s current or recent resource usage lowers the priority applied when LSF assigns resources for their new jobs.

The general queue is available for jobs requiring 15 or less cores.

The parallel queue is available for jobs requiring 16 or more cores. Submitting jobs to this queue *requires* resource distribution -R "span[ptile=16]" which makes sure as few nodes as possible are used.

The bigmem queue is available for jobs requiring more than 32 GB of memory per node. Submitting jobs to this queue requires project membership. Do not submit jobs that can run on the general and parallel queues to the bigmem queue.

Warning

Jobs using less than 1.5G of memory per core on the bigmem queue are

in violation of acceptable use policies and the IDSC account responsible for those jobs may be suspended (Policies).

Pegasus Job Queues

Queue Name

Cores per node

Memory

Wall time default / max

Description

general

16 (15 for jobs)

up to 32 GB per node

1 day / 7 days

jobs up to 15 cores, up to 32 GB memory

parallel

16

up to 32 GB per node

1 day / 7 days

parallel jobs requiring 16 or more cores, up to 32 GB memory per node.
requires resource distribution -R “span[ptile=16]”

bigmem

16

up to 250 GB per node

4 hours / 5 days

jobs requiring nodes with more than 32 GB of memory, max. of 64 cores allowed per job.

debug

16

up to 32 GB

30 mins / 30 mins

job debugging, max. of 64 cores allowed per job.

interactive

16 (15 for jobs)

up to 250 GB per node

6 hours / 1 day

interactive jobs
max 1 job per user

gpu

xx

320 max

1 day / 7 days

gpu debugging restricted queue

phi

xx

320 max

1 day / 7 days

phi debugging restricted queue