Skip to content

Running Jobs on GannettPeak

Assumptions

This page only covers what is unique about running jobs on the MedicineBow Cluster. All other information about running jobs is covered on Running HPC Jobs With Slurm Pages

Running Jobs on GannettPeak is very similar to MedicineBow, with both interactively with ‘salloc’ and batch jobs with ‘sbatch’ using Slurm and Open OnDemand. However, there are some differences to be aware of.

Maximum Walltime

    username@ai4wy-log2:~$ salloc -A projectname -t 25:00:00
    salloc: error: Job submit/allocate failed: Requested time limit is invalid (missing or exceeds some limit)
    username@ai4wy-log2:~$ salloc -A projectname -t 1-1:00:00
    salloc: error: Job submit/allocate failed: Requested time limit is invalid (missing or exceeds some limit)

Partitions

At time of writing, the GannettPeak cluster has only two hardware partitions, ai4wy-1 for single CPU/GPU nodes and ai4wy-2 for the dual CPU/GPU nodes. Requests for multiple GPU resources can get distributed between the single and dual GH200 node configurations. For example, if a job required 3 GPU resources:

    username@ai4wy-log2:~$ salloc -A projectname -G 3
May allocate 2 or 3 nodes depending on the availability of the single or dual GH200 configured nodes.

However, since the maximum number of GPU resources in a single node is two. Using the Slurm directive ‘--gpus-per-node=’ will only allow values of ‘1' and '2’. Please refer to the GannettPeak Hardware Profile for more information.

Required Slurm Directives on GannettPeak

Just like on MedicineBow, the Slurm directive ‘--account=’ is required to submit any job on GannettPeak. However, since every node in the GannettPeak cluster has GPU resources, none of the NVIDIA or CUDA commands will work unless the ‘--gres=gpu:x’ Slurm flag is used. It is highly recommended to use both the ‘--account=’ or ‘-A’ and ‘--gres=gpu:x’ or ‘-G’ Slurm directives for every job.

Interactive Applications with Open OnDemand

At the time of writing, interactive applications like Jupyterlab or Interactive XFCE desktop are not functional. They will be rolled out in the following weeks.