GPU server CATA

GPU server commissioned by CATA. This server is integrated into the Kultrun server at the Universidad de Concepcion.

To request an account, please email lkrapp@udec.cl. The email subject depends on your current account status:

Email subject for Users with a Kultrun account: CATA account
Email subject for Users without a Kultrun account: CATA account + Kultrun

For support requests, please also include Daniel Benavides: dbenavid@desem.cl.

All emails must include a justification for the need for GPU computing based on the NVIDIA A4000 capabilities.

GPU Server Access

Details on how to connect and utilize the GPU server resources.

1. SSH Access Point

Access to the server is granted via SSH. You can connect to the main Kultrun server using the following command in your terminal. Replace user with your Kultrun username.

ssh user@kultrun.astro-udec.cl

2. Resource Management with Slurm

The GPU server is not accessed directly but through the Slurm Workload Manager, which manages all computing jobs on the Kultrun cluster.

To submit jobs specifically to the GPU server, you must specify the **"cata"** partition in your Slurm batch script. This ensures your job is directed to the correct resources.

Important:

Ensure your job submission script includes the line #SBATCH --partition=cata.

To request "N" GPUs add #SBATCH --gres=gpu:N.

MPI CUDA GPU script


!/bin/bash
                                                                
#SBATCH --job-name=cataA4000
                                                    
#SBATCH --account=cata
                                                          
#SBATCH --partition=cata
                                                        
#SBATCH --time=01:00:00
                                                         
#SBATCH --nodes=1
                                                               
#SBATCH --ntasks-per-node=16
                                                    
#SBATCH --gres=gpu:1
                                                            
#SBATCH --output=en1.out
                                                        

module load openmpi cuda

mpirun -np 1 ./execode