Partitions and Limits

Partitions are a group of nodes together with a set of properties and limits. Most notably, the maximum time a single job is allowed to run (the so-called wall time) on Julia HPC is different, depending on the partition the job is submitted to. If a user does not provide an explicit time limit for a job, the limit is set automatically (roughly a quarter of the partitions maximum time).

Standard time limit partitions

The following partitions are currently provided:

Partition	Description
standard	The default partition. If you don't specify any requirements, your jobs will run here.
gpu	Select this partition if you need nodes with GPU cards
fat	Select this partition if you need a lot of CPUs and/or a lot of shared memory (>375GB)

If you don't set any timelimit, a default runtime of 6 hours will be set. You can extend this runtime up to two days with the -t or --time= option to your job submission. Example line from a job script:

#SBATCH -t 2-00:00:00

Concerning the syntax, see the following line from the sbatch man page:

Acceptable time formats include "minutes", "minutes:seconds", "hours:minutes:seconds", "days-hours", "days-hours:minutes" and "days-hours:minutes:seconds".

Partitions with different time limits

Reduced wall time

There are partitions available with a maximum wall time of 12 hours. Their names carry a short marker:

• short corresponds to standard
• short-gpu and short-fat correspond to gpu and fat, respectively.

Selecting one of the short partitions results in a higher priority for the job.

Extended wall time

There are partitions available with a maximum wall time of 7 days. Their names carry a long marker:

• long corresponds to standard
• long-gpu and long-fat correspond to gpu and fat, respectively.

Selecting one of the long partitions results in a lower priority for the job.

Additionally, there are partitions with a maximum wall time of 30 days. Their names carry an ultra marker:

• ultra corresponds to standard
• ultra-gpu and ultra-fat correspond to gpu and fat, respectively.

Selecting one of the ultra partitions results in an even lower priority for the job.

Specialized Partitions

Jobs with GPUs

Use one of the GPU partitions gpu, short-gpu, long-gpu or ultra-gpu to request a node with a GPU accelerator card:

$ sbatch -p gpu --gpus-per-task 1 myjob.sh

This will request a node with one GPU card (--gpus-per-task 1).

Huge memory jobs

Use one of the fat partitions fat, short-fat, long-fat or ultra-fat to request a fat node:

$ srun -p fat -c 144 --mem=1500G myjob.sh

This requests a node for a (shared memory) job with 144 threads (-c 144) and total RAM amount of 1500 GiB (--mem=1500G).

The Test partition

You can test your jobs by using the special test partition.

$ srun -p test my_test.sh

On this partition users can run (short) tests to see whether their jobs run properly, without having to wait for a long time in the queue for production runs.

However, to avoid abuse of this testing queue, additional conditions apply:

• There are only two (virtualized) nodes available, each with 4 CPU cores and 32GB of RAM. That way, we are able to offer some testing machines without having to take compute nodes out of any production environment.
• The walltime limit for the testing partition is 5 minutes by default and can be extended by the user to a maximum of 1 hour.
• Each user can only have 4 simultaneous jobs (running and pending combined) in the testing queue. Further job submission will be denied until one of the four active jobs finishes.