Yes, use of the computing resources is free of charge for all IGBMC staff. For specific needs, research teams can purchase computing resources that will be integrated into the cluster and for which they will have priority or dedicated access.

All cluster tools are accessible via the module utility. You can list all available tools using the module avail command.

Please visit our Computing tools page to know about how tools are deployed and how you can request a tool.

Scientific computing tools are deployed via the national deployment repository proposed by the Institut Français de Bioinformatique (IFB), to which the IGBMC is a major contributor: https://gitlab.com/ifb-elixirfr/cluster/tools

You can contribute to this repository to propose the deployment of a new calculation tool. This contribution will enable the tool to be deployed rapidly and automatically not only on the IGBMC cluster, but also on all IFB network clusters.

You can also deploy the tool in your personal folder (homedir) or in one of your project spaces (space2) via a Conda environment or Singularity image. To find out more, consult our documentation.

First, check the reason given in the job registry. If the job has been stopped for less than a few minutes, you can view detailed information with the command :

scontrol show jobid=jobid

Otherwise, the accounting database must be queried:

sacct

The last column of the command result gives two values separated by a colon. The first is the return code of the script (only code 0 indicates that everything has ended successfully), the second the signal number if the process was killed by a signal. The State column can give a clue as to the reason for the crash: FAILED indicates a job error, CANCELLED indicates that the job was stopped prematurely.

It is possible to modify the columns presented by sacct, for example :

SACCT_FORMAT="jobid,user,elapsed,state%40,Comment,Timelimit,Start,ReqMem,maxRss,ReqCpus,nodelist" sacct -S 2021-11-08 -E 2021-11-10

où la syntaxe "%N" permet de rallonger la taille de la colonne. C'est nécessaire lorsque le contenu d'une colonne affiche des "+", ce qui signifie que le texte est plus long que la largeur de la colonne.

Les champs "ReqMem" et "MaxRss" permettent de voir si la mémoire demandée par le job a atteint ou dépassé la mémoire reservée, ce qui est une cause fréquente d'annulation des jobs. Dans ce cas, relancez le job avec plus de mémoire (si vous n'avez pas d'information pour calculer le besoin en mémoire, doublez-la à chaque tentative infructueuse, si possible) puis diminuez-la au cours des lancements ultérieurs pour adapter au mieux la demande à l'utilisation réelle.

Les options -S et -E permettent respectivement de spécifier une date de début et de fin de recherche pour les jobs à afficher.

Si votre job affiche un résultat "FAILED", consultez la sortie d'erreur et la sortie standard du job. Par défaut, Slurm place celles-ci dans des fichiers nommés "slurm-numérodujob.out" pour les jobs lancés avec sbatch.

where the syntax %N is used to extend the column size. This is necessary when the contents of a column display +, meaning that the text is longer than the width of the column.

The ReqMem and MaxRss fields show whether the memory requested by the job has reached or exceeded the reserved memory, which is a frequent cause of job cancellation. If this is the case, restart the job with more memory (if you don't have the information to calculate the memory requirement, double it for each unsuccessful attempt, if possible), then decrease it on subsequent runs to adapt the demand to actual usage.

The -S and -E options allow you to specify a search start and end date respectively for jobs to be displayed.

If your job displays a FAILED result, consult the job's error output and standard output. By default, Slurm places these in files named slurm-jobid.out for jobs launched with sbatch.

A number of compute nodes are equipped with GPUs. At the time of writing, these are:

• phantom-node1: 4x Nvidia K80 (Tesla)
• phnatom-node33 : 2x Nvidia A100 (Ampere)
• phantom-node34 : 2x Nvidia A100 (Ampere)

To reserve these resources, place your job in the gpu partition (option -p gpu) and select the number and type of GPUs to allocate to the job.

This is done by specifying the option:

--gres=gpu:gpuname:number 
or
--gres=gpu:number 

You can find out which GPUs are available by running:

sinfo -N -o "%N %P %G" -r | grep -v "(null)"

sinfo is the SLURM command to view informations about SLURM nodes and partitions

-N: display each node

-o “%N %P %G”: output only the node name (%N), the partition associated to the node (%P) and the generic resources available (%G)

-r: only show responding nodes

| grep -v “(null)”: exclude nodes with no generic resources from output

Si vous ne choisissez pas un type de GPU spécifique, le système d’allocation prendra n’importe lequel disponible.

Les cartes A100 peuvent être subdivisées selon le besoin, aussi les GPU disponibles avec ces cartes peuvent varier selon le moment. Leur nommage suit la forme aXg.Ygb, où X est le nombre d’unités de calcul et Y la quantité de RAM associée à ce GPU. Si aucune configuration disponible n’est adaptée au job que vous voulez lancer, merci de contacter le pôle calcul.

If you don't choose a specific GPU type, the allocation system will take any available one on the given partition.

A100 cards can be subdivided according to need, so the GPUs available with these cards may vary from time to time. Their naming follows the form aXg.Ygb, where X is the number of compute units and Y the amount of RAM associated with that GPU. If none of the available configurations is suitable for the job you want to run, please contact the helpdesk.