We are building and expanding a GPU Cluster as part of the Lamarr Institute. At the moment, the cluster consists of nine nodes with the following basic configuration:

2 AMD Epyc 7713 Processors (64 Cores * 2 Threads each ⇒ 256 Cores per Node, 248 available for jobs)
2TB System-RAM
28TB local SSD storage
8x A100 SXM4 HGX w/each 80GB GPU-Memory, NVLink (600GB/s) connected
10GbE Network

The nodes itself cannot be accessed directly, we have prepared a login node for this purpose, which can be utilized as a staging area for your code and data.

All nodes and the login node share a common data storage area (332TB, cephfs), which is mounted as /home/ and set as the standard user homes. Shared datasets could be placed there and made available to selective groups of users via standard Posix file permissions/user groups.

All nodes have 28TB local SSD storage available, mounted under /data. Please remove your data from there after you are finished with your calculations.

The workload management software/scheduler is SLURM. Please refer to the user documentation of this software for further information about SLURM:

https://slurm.schedmd.com/documentation.html

Full container support is enabled. We recommend the usage of enroot as container backend for SLURM. If you need a personalized library or OS setup around your software, containers will be your way of choice.

As a general rule, realistic resource allocation for your jobs is mandatory. You are only allowed to request the resources from the workload manager that you are actually needing to complete your calculations. Wasting precious machine time and making other users wait in line needlessly will lead to revocation of access to the cluster.

At this time, there is only one active partition available. It is called “batch”, the timelimit for jobs in this partition is set to a maximum of 72h and a default of 3h.

At the moment the cluster is available to all Principal Investigators of the Lamarr Institute. If you want to gain access, please contact us per email and request an account for you and/or your colleagues. Please do not forget to specify the nature your personal association to the Lamarr Institute, so we can verify the validity your access requests.

As soon as you received the confirmation email from us, you can access the cluster through the login-node using the SSH protocol. All necessary additional information regarding the login will be contained in this email.

Obviously you will not get administrative privileges on the cluster to setup your personal software environment. But you can import containers to setup any libraries or frameworks you need to use. We tested the cluster setup using enroot from NVidia, which is integrated into SLURM quite well. Any other container environment supported by pyxis should work, too, but was not tested by us yet.

In enroot, you import first import external container images, but those images are immutable. You have to create a container and start it read/write to make persistent changes to it.

First you import an image (i.e. ubuntu from docker registry):

$ enroot import -o ~/ubuntu.sqsh docker://ubuntu

then you create an container from the image (ubuntu.sqsh):

$ enroot create --name mycontainer ~/ubuntu.sqsh

Now you start the container and install everything you need:

$ enroot start -r -w mycontainer

When you are finished, create a new sqsh Image from your container:

$ enroot export --output ~/mycontainer.sqsh mycontainer

Now you can run software via SLURM inside the container like this:

$ srun --container-image ~/mycontainer.sqsh /script-in-the-container

Of course you can create your custom image (mycontainer.sqsh) on any host you like, it does not have to be the login node. You can transfer the .sqfs file from anywhere to the login node (i.e. via scp/sftp) and start your container via srun just like shown above.

If you want to share bigger datasets with other users on the cluster you can request shared storage on the cluster. It will be located under /home/Shared/. Please specify the usernames of all users that should be able to access this data (and any other hepful informations for us regarding your reequest) in an informal request email. We will then create a user group for the specified users, create the directory and set the necessary permissions up for you.

If you want to allocate a certain number of GPUs for your job, use the parameter '–gres=gpu:COUNT' in your srun/squeue calls, where COUNT is the requested number of GPUs for your job.

If you want to receive email notifications about the status of your jobs, you can specify a recipient email address at job submission, i.e.:

sbatch –mail-type=ALL –mail-user=someuser@somedomain.com …

Please specify a full (with domain!) and valid email address here.

Please remember that interactive jobs often leave things unnecessarily idle (especially if forgotten and/or unterminated), so please think about your impatient colleagues in the queue behind you and try to avoid interactive jobs whenever possible.

There is a (federated) matrix chatroom for all users of the Lamarr GPU Cluster, please feel free to join if you like:

#lmgpu-tech:matrix.informatik.uni-bonn.de