Orchid GPU cluster

This article provides details on JASMIN’s GPU cluster, named ORCHID.

GPU cluster spec  

The JASMIN GPU cluster is composed of 16 GPU nodes:

• 14 x standard GPU nodes with 4 GPU Nvidia A100 GPU cards each
• 2 x large GPU nodes with 8 Nvidia A100 GPU cards

Request access to ORCHID  

Before using ORCHID on JASMIN, you will need:

1. An existing JASMIN account and valid jasmin-login access role: Apply here     
2. Subsequently (once jasmin-login has been approved and completed), the orchid access role: Apply here     

The jasmin-login access role ensures that your account is set up with access to the LOTUS batch processing cluster, while the orchid role grants access to the special LOTUS partition used by ORCHID.

Holding the orchid role also gives access to the GPU interactive node.

Note: In the supporting info on the orchid request form, please provide details on the software and the workflow that you will use/run on ORCHID.

Test a GPU job  

Testing a job on the JASMIN ORCHID GPU cluster can be carried out in an interactive mode by launching a pseudo-shell terminal Slurm job from a JASMIN scientific server e.g. sci-vm-01:

srun --gres=gpu:1 --partition=orchid --account=orchid --pty /bin/bash
srun: job 24096593 queued and waiting for resources
srun: job 24096593 has been allocated resources

# you are now on gpuhost16

The GPU node gpuhost016 is allocated for this interactive session on LOTUS

Note that for batch mode, a GPU job is submitted using the Slurm command ‘sbatch’:

sbatch --gres=gpu:1 --partition=orchid --account=orchid gpujobscript.sbatch

or by adding the following preamble in the job script file

#SBATCH --partition=orchid
#SBATCH --account=orchid
#SBATCH --gres=gpu:1

Note 1: gpuhost015 and gpuhost016 are the two largest nodes with 64 CPUs and 8 GPUs.

Note 2: CUDA Version: 11.6

Note 3: The Slurm batch partition/queue orchid has a maximum runtime of 24 hours and the default runtime is 1 hour. The maximum number of CPU cores per user is limited to 8 cores. If the limit is exceeded then the job is expected to be in a pending state with the reason being QOSGrpCpuLimit

GPU interactive node  

There is an interactive GPU node gpuhost001.jc.rl.ac.uk, with the same spec as other Orchid nodes, that you can access via a login server to prototype and test your GPU code prior to running as a batch job.

ssh -A gpuhost001.jc.rl.ac.uk

# you are now on gpuhost001

Software Installed on the GPU cluster (to be updated)  

• CUDA drivers 10.1, and CUDA libraries 10.0, 10.1 and 11.4
• CUDA DNN (Deep Neural Network Library)
• NVIDIA container runtime (see notes below)
• NGC client (GPU software hub for NVIDIA)
• Singularity 3.7.0 - which supports NVIDIA/GPU containers
• SCL Python 3.6