Resources per team
Shared access to a GPU accelerated 261 nodes HPC partition.
Visualisation session also available.
Data cube access
Shared through mounted disk space
Accessed through Slurm job scheduler.
Through regular IDRIS support.
Containerisation option available through Singularity.
Regular IDRIS support for application support (firstname.lastname@example.org)
IDRIS - Institute for Development and Resources in Intensive Scientific Computing (http://www.idris.fr/eng/info/missions-eng.html) will provide access to the GPU accelerated partition of its main supercomputer called Jean Zay (http://www.idris.fr/eng/jean-zay/).
Jean Zay’s computing resources are currently used by 2700 researchers and engineers and almost 1450 Projects are running across IDRIS resources (30% HPC projects / 70% AI projects).
IDRIS will provide access to the 1044 NVIDIA V100 SXM2 GPUs split over 261 nodes (4 GPUs per node). Each node includes:
2 Intel Cascade Lake SP 6248 sockets
40 cores running at 2.5 GHz each
192 GB of memory
4 GPUs (32 GB HBM per GPU)
Compute nodes have access to a shared full-flash parallel file system based on IBM Spectrum Scale (capacity > 2 PB ; Throughput > 450 GB/s).
Users will also have access to visualisation and pre/post-processing nodes.
Jean Zay’s full specifications can be found here.
Per user resource
Every user will have access to the compute nodes available to the SDC3 project.
Every project will be allocated 10000h of GPU runtime by default, shared between all users of one SDC3 team.
Each job elapse runtime shouldn’t exceed 100h (see here).
If needed, each team project leader will ask for an extension of computing hours to IDRIS for established process.
Cluster nodes are running RedHat
Job scheduler is Slurm
Both Intel and PGI/NVIDIA compilers are provided
Libraries are managed through the module system
A list of all available Scientific software and libraries is provided through the FAQ. They can be listed with the “module avail” command.
Volume of resource
IDRIS can accommodate up to 40 user accounts, that can be dispatched on any number of teams.
GPUs if any
1044 Nvidia Tesla V100 SXM2 GPUs with 32 GB HBM each.
All users will connect on jean-zay.idris.fr, through ssh:
$ ssh <login>@jean-zay.idris.fr
How to run a workflow
Jobs will be executed on computing resources through Slurm. Full documentation (including examples) is provided at the following address: http://www.idris.fr/eng/jean-zay/
Accessing the data cube
Input Data will be stored on an IBM Spectrum Scale file system mounted on every accelerated node. IDRIS will provide the directory path name where data will be stored.
Different storage spaces are available, as detailed here: http://www.idris.fr/eng/jean-zay/cpu/jean-zay-cpu-calculateurs-disques-eng.html .
Installed software is managed through the module environment. Installation of missing libraries is handled through regular support request. User can also install their own libraries.
A list of currently available tools and library is available at the following address: http://www.idris.fr/eng/jean-zay/
Single-node and multi-node container usage through singularity (see http://www.idris.fr/eng/jean-zay/cpu/jean-zay-utilisation-singularity-eng.html for the relevant documentation).
The documentation is available online at the following address: http://www.idris.fr/eng/jean-zay/
The computing resources are accessed through the Slurm job scheduler
Each team will be assigned to an account with a predefined number of hours.
Storage is granted on a per user and per project basis, as detailed here.
IDRIS support (for applicative troubleshooting) can be contacted either by email (preferred) or by phone (see below). The email subject must include SKADC3. Support by phone can be provided at: +33 (0)1 69 35 85 55.
For more details, please refer to http://www.idris.fr/eng/su/assist-eng.html
The IDRIS HPC community is supported by the following people:
Credits and acknowledgements
This work was granted access to the HPC resources of IDRIS under the allocation 20XX-[project number] made by GENCI.