RCI Cluster

A powerful shared compute cluster with 100+ A100 GPUs, accessed via login nodes. Ideal for large experiments, sweeps, and long jobs.

🔐 Access 

ssh username@login3.rci.cvut.cz

Use login nodes for job submission and environment setup only (they don’t have GPUs).

📦 Storage

Refer to storage info. Use /scratch, /home, or /storage/plzen1/home/username as appropriate.

⚙️ Interactive Jobs

Launch a debugging session:

srun --partition=interactive --gres=gpu:1 --mpi=pmix --mem 25G --ntasks-per-node=1 --pty bash -i

📄 Job Submission

Submit with:

sbatch bash_scripts/my_job.sh

Example job script:

#!/bin/bash
#SBATCH --job-name=my_job
#SBATCH --output=output.log
#SBATCH --error=error.log
#SBATCH --gres=gpu:1
#SBATCH --mem=25G
#SBATCH --ntasks=1

module load PyTorch/2.5.1-foss-2023b-CUDA-12.4.0
source path/to/my_env/bin/activate
python train.py

📊 Monitoring Jobs 

squeue | grep username     # Check running jobs
scancel <job_id>          # Cancel job

📚 Array Jobs

Useful for launching multiple jobs with a single script:

#!/bin/bash
#SBATCH --job-name=multi_train
#SBATCH --output=logs/out_%A_%a.log
#SBATCH --error=logs/err_%A_%a.log
#SBATCH --array=0-3
#SBATCH --gres=gpu:1
#SBATCH --mem=32G
#SBATCH --ntasks=1

module load PyTorch/2.5.1-foss-2023b-CUDA-12.4.0
source path/to/my_env/bin/activate

# List of configs to run
CONFIGS=("imagenet" "coco" "voc" "laion")

python train.py --dataset ${CONFIGS[$SLURM_ARRAY_TASK_ID]}

Submit with:

sbatch run_array.sh

Each job will run train.py with a different dataset.

Tutorial on Interactive Python development on RCI GPU nodes

Tutorial

🐞 Debugging

🔄 File Sync

Same options as Non-RCI servers:

  • PyCharm deployment
  • rsync, scp, git

⚙️ Python and Other Software Setup

See Python environment setup.