In this tutorial you will connect to the ML cloud slurm cluster and allocate computing resources to run some very basic example code. This is just a starting point, and you will probably need to work through some or all of the follow-up examples to use slurm for your more sophisticated needs.
Once Slurm access is granted, you can open a terminal and connect to the slurm head node:
ssh <your-slurm-username>@134.2.168.52
Note: You should switch to SSH-key based authentication as soon as possible, as described in the next tutorial and here.
Download the code in your current working directory on the slurm headnode using the following line:
git clone https://github.com/baumgach/tue-slurm-helloworld.git
Change to code directory
cd tue-slurm-helloworld
The file 01_hello_world.sbatch contains instructions for your slurm job. Those are a bunch of options for which GPUs to use etc (lines starting with #SBATCH), followed by the code that must be executed.
Among other things 01_hello_world.sbatch contains the following two lines
#SBATCH --output=logs/hostname_%j.out # File to which STDOUT will be written
#SBATCH --error=logs/hostname_%j.err # File to which STDERR will be written
specifying where the log and error files for each job will be written to. You can see, that this script writes them to a logs directory relative to your working directory, i.e. the folder where you execute your container command. Slurm will not create this folder automatically, so we need to do it manually first using
mkdir logs
Then submit your slurm job using the following command
sbatch 01_hello_world.sbatch
This will allocate resources on the slurm cluster as specified in 01_hello_world.sbatch, and run the code specified there on these resources. Note that the leading #SBATCH do not mean that those lines are commented out. This is just slurm syntax.
Have a look at the outputs written to ~/logs/.
Of course this is a very basic example, and you should later taylor 01_hello_world.sbatch to your needs.
More info on how to manage Slurm jobs can be found here