Queue Structure

Job Routing

Batch jobs will be automatically re-routed to different internal queues as following. One can specify the queue name via #PBS -q normal or #PBS -q ai to access the different partitions.

Note

Jobs that do not satisfy any of the conditions below will not be fulfilled.

Queue Name	Queue Server	Internal Queue	Conditions	Walltime Limit
`normal`	`pbs101`	`q1`	ncpus = 1	02:00:01 to 24:00:00
		`q2`	2 ≤ ncpus ≤ 64
		`q3`	65 ≤ ncpus ≤ 127
		`q4`	ncpus = 128
		`q5`	129 ≤ ncpus ≤ 2048
		`q6`	2049 ≤ ncpus ≤ 98304
		`qdev`*	1 ≤ ncpus ≤ 128	00:00:01 to 02:00:00
		`qlong`*	1 ≤ ncpus ≤ 128	24:00:01 to 120:00:00
		`g1`	ngpus = 1	02:00:01 to 24:00:00
		`g2`	2 ≤ ngpus ≤ 3
		`g3`	ngpus = 4
		`g4`	5 ≤ ngpus ≤ 256
		`gdev`*	1 ≤ ngpus ≤ 4	00:00:01 to 02:00:00
		`glong`	1 ≤ ngpus ≤ 64	24:00:01 to 120:00:00
		`l1`	1 ≤ ncpus ≤ 128 memory = 512GB to 1TB	00:00:01 to 24:00:00
		`l2`	1 ≤ ncpus ≤ 128 memory = 1TB to 2TB	00:00:01 to 24:00:00
		`l3`	1 ≤ ncpus ≤ 128 memory = 2TB to 4TB	00:00:01 to 24:00:00
`ai`	`pbs102`	`aiq1`	ngpus = 1	00:00:01 to 24:00:00
		`aiq2`	2 ≤ ngpus ≤ 3
		`aiq3`	ngpus = 4
		`aiq4`	5 ≤ ngpus ≤ 96
		`aidev`	ngpus = 1	00:00:01 to 02:00:00
		`ailong`*	1 ≤ ngpus ≤ 4	00:00:01 to 120:00:00

* Additional limits on the number of running jobs apply. See Running Job Limits.

Running Job Limits

For fair allocation of HPC resources, the following limits are in place, which may be subject to change:

There is a system wide restriction of a maximum of 100 jobs per user, unless otherwise specified at queue level.
qdev: Maximum 2 running jobs per user at a time, up to a total of 256 ncpus.
qlong: Maximum 2 running jobs per user at a time, up to a total of 128 ncpus.
gdev: Maximum 2 running jobs per user. Maximum 16 ngpus running at a time, all users combined.
ailong: Maximum of 1 running job per user.