GWDG HPC Resources

Access GWDG HPC cluster resources for deep learning: GPU nodes, job scheduling, available courses, and compute allocations at University of Göttingen.

Overview

The GWDG (Gesellschaft für wissenschaftliche Datenverarbeitung mbH Göttingen) is a joint data and IT service center for the University of Göttingen and the Max Planck Society. It provides:

• High-performance computing (HPC) clusters
• GPU resources for deep learning
• Training and courses
• Research collaboration

Getting Started

Access Requirements

1. Account Registration - Apply through your institution
2. SSH Key Setup - Required for cluster access
3. Course Completion - Recommended for efficient usage

Cluster Access

# Using SSH key authentication
ssh -i $HOME/.ssh/id_rsa_nhr -l username glogin.hlrn.de

# Example with specific key and user
ssh -i $HOME/.ssh/id_ed25519 u10000@glogin9.hlrn.de

# Add to ~/.ssh/config for easier access
Host gwdg
    HostName glogin9.hlrn.de
    User u10000
    IdentityFile ~/.ssh/id_ed25519
    ServerAliveInterval 60
    ServerAliveCountMax 3

# Now just use: ssh gwdg

# Generate ed25519 key (recommended)
ssh-keygen -t ed25519 -f ~/.ssh/id_gwdg -C "your.email@example.com"

# Or RSA key
ssh-keygen -t rsa -b 4096 -f ~/.ssh/id_rsa_gwdg

# Copy public key to GWDG
# (Submit through GWDG portal or email)
cat ~/.ssh/id_gwdg.pub

GWDG Training Courses

Deep Learning with GPUs

Comprehensive course on GPU computing for deep learning on HPC clusters.

Topics Covered:

• GPU architecture and CUDA basics
• Deep learning frameworks (PyTorch, TensorFlow)
• Batch job submission with GPUs
• Multi-GPU training strategies
• Performance optimization

Scientific Computing on Clusters

Practical introduction to HPC cluster usage.

Topics Covered:

• Cluster architecture
• SLURM job scheduler
• Resource allocation
• Module system
• Data management

:::tip[Highly Recommended] Complete these courses before starting intensive GPU workloads. They cover best practices specific to GWDG infrastructure and will save you time troubleshooting. :::

Common GWDG Workflows

Submit GPU Job

#!/bin/bash
#SBATCH --partition=gpu
#SBATCH --gres=gpu:1
#SBATCH --time=24:00:00
#SBATCH --mem=32G
#SBATCH --cpus-per-task=4
#SBATCH --job-name=my_training
#SBATCH --output=logs/%x_%j.out

# Load modules
module load python/3.10
module load cuda/12.1

# Activate environment
source $HOME/venvs/ml/bin/activate

# Run training
python train.py --epochs 100 --batch-size 64

Interactive GPU Session

# Request interactive session with GPU
srun --partition=gpu --gres=gpu:1 --mem=16G --cpus-per-task=4 --time=2:00:00 --pty bash

# Check GPU availability
nvidia-smi

# Run interactive work
python

Check Job Status

# Your jobs
squeue -u $USER

# Specific job details
scontrol show job JOBID

# Job history
sacct -u $USER --format=JobID,JobName,State,Elapsed,MaxRSS

Data Management

Transfer Data to GWDG

# Upload dataset
rsync -avhP --progress \
  /local/datasets/imagenet \
  u10000@glogin9.hlrn.de:/work/datasets/

# Download results
rsync -avhP --progress \
  u10000@glogin9.hlrn.de:/work/experiments/checkpoints \
  ./local_backups/

# Upload file
scp -i ~/.ssh/id_gwdg large_dataset.tar.gz \
  u10000@glogin9.hlrn.de:/work/datasets/

# Download file
scp -i ~/.ssh/id_gwdg \
  u10000@glogin9.hlrn.de:/work/results/model.pth \
  ./local/

# Connect with sftp
sftp -i ~/.ssh/id_gwdg u10000@glogin9.hlrn.de

# SFTP commands
put local_file.txt         # Upload
get remote_file.txt        # Download
ls                         # List remote files
lcd /local/path            # Change local directory

Storage Locations

# Home directory (limited space)
$HOME

# Work directory (larger quota)
/work/$USER

# Scratch for temporary files
/scratch/$USER

# Check quotas
quota -s

Module System

# List available modules
module avail

# Search for specific module
module avail cuda

# Load modules
module load python/3.10 cuda/12.1 cudnn/8.9

# List loaded modules
module list

# Unload module
module unload cuda

# Purge all modules
module purge

Best Practices

1. Test Locally First - Debug on small datasets before cluster submission
2. Use Batch Jobs - Don’t run long jobs on login nodes
3. Monitor Resources - Use seff JOBID to check efficiency
4. Clean Up - Remove old data from scratch regularly
5. Checkpointing - Save progress frequently for long jobs

:::caution[Login Nodes] Login nodes are shared resources. Never run computationally intensive tasks on login nodes. Always submit batch jobs or request interactive sessions. :::

Troubleshooting

Job Won’t Start

# Check queue
squeue -p gpu

# Check job priority
sprio -j JOBID

# Explain why job is pending
squeue -j JOBID --start

Out of Memory Errors

# Check actual memory usage
seff JOBID

# Request more memory in job script
#SBATCH --mem=64G

Connection Issues

# Test connection
ssh -v u10000@glogin9.hlrn.de

# Check key permissions
chmod 600 ~/.ssh/id_gwdg
chmod 644 ~/.ssh/id_gwdg.pub

Additional Resources

• GWDG Official Documentation
• SLURM Guide - General HPC usage
• Multi-GPU Training - Distributed training
• Backup & Sync - Data transfer strategies

:::note[Institution-Specific] This guide focuses on GWDG infrastructure. Other HPC centers may have different configurations, but general principles apply. :::