Running Jobs

To run a command as a job you can either use srun to quickly run it interactively or sbatch to spool it for asynchronous execution.

In either case you always need to supply the following two parameters for the job:

The course or project for which you run the job. You can see the course or project tags that you can use when you log in via ssh.
The maximum runtime you expect this command to run. This can be omitted in which case a default of 60 minutes is used.

One GPU is automatically added to the job for courses and projects that use GPUs.

Using `srun`

For a simple, short-running command you can use srun with the -A and -t switch. For instance, to get the properties of the GPU at your disposal (and at most spend ten seconds) run:

srun -A {tag} -t 00:10 -o nvidia-smi.out nvidia-smi

Running Interactive Jobs

If you need to interact with a program via terminal then you can also use srun with the --pty argument. To get an interactive bash on a node for at most 60 minutes you can run:

srun --pty -A {tag} -t 60 bash

Using `sbatch`

For sbatch you will put your commands in a script and add comment lines at the beginning that contain additional parameters for sbatch. The example from above but using sbatch would require a script containing this:

#!/bin/bash

#SBATCH --time=00:10
#SBATCH --account={tag}
#SBATCH --output=nvidia-smi.out

nvidia-smi

To send the script to the cluster for execution run:

sbatch batch.sh

GPU Selection

For courses and projects that allow multiple GPUs you can request more than one GPU using the --gpus switch, both with srun and sbatch:

srun --gpus 4 ...

You can also request a specific GPU model for your job using --gpus:

srun --gpus 1080ti:1 ...

Currently the following GPUs are available:

1080ti (NVidia GeForce GTX 1080 Ti, 11GB)
5060ti (NVidia GeForce RTX 5060 Ti, 16GB)

Using Modules

To use the modules command in a batch script your first command must be the following (note the dot):

. /etc/profile.d/modules.sh

For example:

#!/bin/bash

#SBATCH --time=00:10
#SBATCH --account={tag}
#SBATCH --output=nvcc.out

. /etc/profile.d/modules.sh
module add cuda/12.9
nvcc --version

Temporary Space and Network Scratch

Each job receives a dedicated temporary directory under /tmp ($TMPDIR). This directory is located on the local SATA SSD. Data placed there is deleted when the job ends. There are also limitations on how much space you can use.

You also have a personal network scratch directory under /work/scratch that is accessible from all nodes, including the login nodes. Old data is automatically deleted, how soon depends on how much you put in there.

Checking the Job Queue

To see if you job is still waiting in the queue run:

squeue

If you see jobs listed then they are either executing or they are waiting, in case of the later there will be a reason given why the job cannot start yet.

This cluster is maximizing energy efficiency and powers down nodes that are idle. It may first need to start a node to run your job if all others are busy. This can delay the start for five minutes.

Output

Terminal output of commands in the job by default go to a file slurm-{job ID}.out in your home directory. The file name for the output can also be set. In the above examples the terminal output of the job will be in the file nvidia-smi.out.

Aborting Jobs

If you already know that a running or spooled job is not going to do the right thing, cancel it using

scancel {job ID}

This will save you GPU time.

Tunneling SSH to Nodes with `cluster-tunnel`

The cluster-tunnel utility (a clone of euler-tunnel) can be used to connect to nodes via SSH from outside. This can be for instance used to run vscode or code-server directly on nodes.

The instructions are the same as provided for the Euler cluster) but all occurrences of the string euler have to be replaced by cluster.