Computing environment on Dalia

⚠ INFORMATION
This page was translated by an AI (LLM) with a cursory human check and is awaiting full review.

Disk spaces

There are two types of storage space on Dalia:

• A personal space HOME per user: /lustre/home/<login>
  • The environment variable $HOME is set when you log in and points to this space;
  • The quotas per user are 3 GB and 12,000 inodes and can be increased if needed.
• A workspace per user and per project: /lustre/work/<project_group>/<login>
  • Currently, there is no environment variable pointing to this folder. We recommend setting a variable (e.g. $WORK) in the ~/.bashrc to make things easier;
  • A workspace common to all members of the same project is also available: /lustre/work/<project_group>/commun;
  • The workspace quotas (users + common) depend on the project. In total, the machine has 500 TB.

info

• By default, the number of inodes allocated will be proportional to the storage volume requested for the project, up to 3,000 inodes per GB.
• For each space, there are quotas in volume (GB) and number of files/directories (inodes). You will be blocked as soon as you exceed one of the limits.

Software environment

Basic software stacks (compilers, CUDA libraries, etc.) are accessible via modules, as on Jean Zay. To list them, you can run the command:

module avail

warning

Some modules are loaded by default when you log in and are visible via the command module list executed before any other command module. This behaviour can be modified in your .bashrc but a module purge may therefore generate unwanted side effects, such as loss of access to Slurm.

The installation of software environments on Dalia is then the responsibility of the users. It is possible to work in containers or conda environments according to your preference.

danger

Whether you build a container or a conda environment, it is crucial that the installations are compatible with the ARM architecture of Dalia's compute nodes which is different from the Intel architecture of the login nodes.

Apptainer Containers

It is recommended to work in containers. The Apptainer containerisation solution is available on Dalia.

info

To avoid quota problems, you can redirect the cache directory used by Apptainer to the workspace associated with your project (~WORK). To do this, simply define the variable APPTAINER_CACHEDIR as follows:

export APPTAINER_CACHEDIR=/lustre/work/<project_group>/<login>/<cache_directory>

Conda environments

You can, if you wish, install your own conda environment on Dalia.

It is recommended to perform this procedure directly on a compute node so that the ARM architecture is automatically detected during installation.

Example script for installing a conda environment with Miniforge:

# Connexion à un nœud de calcul
srun --ntasks=1 --cpus-per-task=36 --gres=gpu:1 --hint=nomultithread --pty bash

# Téléchargement de miniforge
cd /lustre/work/<project_group>/<login>/
wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-aarch64.sh"

# Installation de miniforge
bash Miniforge3-Linux-aarch64.sh

warning

When running the script Miniforge3-Linux-aarch64.sh, you will be asked if you accept the default installation location /lustre/home/<login>/miniforge3. To avoid saturating your HOME, it is necessary to specify a path pointing to your workspace instead, for example /lustre/work/<project_group>/<login>/miniforge3.