Your first hour on NVwulf: log in, bring your data over from SeaWulf, load software, and run a first GPU job.

On this page: Log in · Move data from SeaWulf · Load software · First GPU job · Interactive GPU session · Tips

Need an account? See Getting access to NVwulf.

1. Log in

Use SSH (Duo required):

ssh -X <netid>@login.nvwulf.stonybrook.edu

Or use your browser with Open OnDemand for NVwulf, which has Jupyter, RStudio, VS Code, AI chat, and science apps ready to launch (see the full app list). Windows setup and Duo tips are in Connecting to the cluster.

2. Move data from SeaWulf

NVwulf has its own storage, separate from SeaWulf. Copy what you need with rsync, running the command on NVwulf:

# One file
rsync -v <netid>@login.seawulf.stonybrook.edu:~/test.txt ~/test.txt

# A whole folder, including subfolders
rsync -av <netid>@login.seawulf.stonybrook.edu:~/test/ ~/test/

For very large datasets, Globus may be faster.

3. Load software

module avail                  # see what's available
module load miniconda/3       # load a module
module load slurm             # needed for job commands

More in Using modules.

4. Your first GPU job

See the GPU partitions (h200x4, h200x8, b40x4, plus -long and debug- versions):

sinfo

Save this as test_job.slurm. It asks for 1 GPU, 8 CPUs, and 25 GB of memory for 5 minutes:

#!/bin/bash
#SBATCH --job-name=test-tf
#SBATCH --output=res.txt
#SBATCH --partition=h200x4
#SBATCH --nodes=1
#SBATCH --ntasks=8
#SBATCH --cpus-per-task=1
#SBATCH --mem=25g
#SBATCH --gpus=1
#SBATCH --time=05:00

module load tensorflow/2.19.0
python tf_test_nn_training.py

Submit it and watch it:

sbatch test_job.slurm
squeue -u $USER

When it finishes, the output is in res.txt. To check the GPU was really used, add nvidia-smi as a line before your program.

Interactive GPU session

For testing and debugging, get a shell on a GPU node:

srun -p h200x4 --gpus=1 --cpus-per-task=8 --mem=25g --time=01:00:00 --pty bash

Type exit when you're done so the GPU goes back to others.

Tips

  • Ask only for the GPUs you'll use. Idle GPUs are expensive and slow down everyone's queue.
  • Start in a debug partition for short tests before a big run.
  • Use the job script builder to make NVwulf scripts with the right partition limits.
  • Read Good conduct on the RCI clusters before running large jobs.
Applies to NVwulf