Skip to content

GPUs

LXPLUS

The easiest way to access a GPU is by using the LXPLUS service:

ssh yourusername@lxplus-gpu.cern.ch
The GPUs are usually an NVIDIA T4.

You also access A100 on nodes with AlmaLinux8:

ssh yourusername@lxplus8-gpu.cern.ch

SWAN

You can also access GPUs on the SWAN Service. You will have to create a ticket to be granted access, you will be pointed to the link for this when you try to create a session.

Using SWAN you have access to GPUs, using CERN's modified version of jupyter notebooks - you can run code interactively, thus it is very good for prototyping.

A detailed view of the software available on SWAN is listed on the LCG Release website.

Kubeflow

The CERN Kubeflow service is focused on deep learning workflows. It contains notebooks, tools for ML pipelines and hyper-parameter optimization.

LXBATCH

You can view nodes with GPUs by specifying a constraint to condor_status, for example, to view all nodes with more than 2 GPUs:

condor_status -constraint 'GPUs > 2'

We can also filter by GPU type, however, due to the current migration process, this command is different for A100 GPUs and the rest:

  • On A100 

     condor_status -constraint 'regexp("A100", CUDADeviceName)'

  • On everything else 

     condor_status -constraint 'regexp("V100", GPUs_DeviceName)'

To increase your priority in the queue you can subscribe to the np-comp group.

To require a GPU when submitting a job to the HTCondor LXBATCH cluster, we can specify in our .sub the following requirement: 

request_GPUs = 1

Info

Currently only machines with A100 GPUs will be assigned to your job by default. This is due to an ongoing upgrade of the nodes to AlmaLinux9, thus to run a job using V100 or T4 you must specify:

requirements = TARGET.OpSysAndVer =?= "AlmaLinux9"
If the job can run on both versions then: 
requirements = ( TARGET.OpSysAndVer =?= "AlmaLinux9" || TARGET.OpSysAndVer =?= "CentOS7")