High Performance Computing infrastructures description
Computing clusters
You can get a list of the characteristics of each cluster by running the recap.py command on the cluster front end.
Dahu is currently the leading platform for high performance computing (HPC) and data analysis (DA).
Head node: dahu.univ-grenoble-alpes.fr or dahu from the bastions.
- HPCDA platform
- Omnipath 100Gb network
- Local SSD and HDD scratch space
- Acces to Bettik, Silenus and Mantis
The path for the scratch1 space is /var/tmp.
The path for the scratch2 space, when it exists, is /var/tmp2.
The Bigfoot platform currently includes all machines with GPU cards.
Head node: bigfoot.univ-grenoble-alpes.fr or bigfoot from the bastions.
- IA and Deep Learning platform
- 3 nodes with 4 Tesla V100 GPUs with NV-link per node
- 4 nodes with 4 Tesla V100 GPUs with PCIe link per node
- 5 nodes with 2 Tesla A100 GPUs including one available as 7 independant MIGs
- Omnipath 100Gb network
- 35 “Virgo” nodes with a T4 GPU available during night time (reserved for training during the day) in collaboration with UGA teaching services
- Acces to Bettik, Silenus and Mantis
The path for the scratch1 space is /var/tmp.
The path for the scratch2 space, when it exists, is /var/tmp2.
Head node: luke.univ-grenoble-alpes.fr or luke from the bastions.
- Heterogeneous platform for specific project requirements
- 10Gbe network
- Visualisation node
- Computing nodes with reserved access for certain projects, excluding generalist “ciment” and best-effort nodes
- Acces to Bettik and Mantis (not Silenus)
This platform is uniquely intended to host computing resources for operations that cannot be performed on existing platforms.
Access to the machines, except for the generalist “ciment” nodes and best-effort mode, is restricted to the teams that funded them.
CiGri, lightweight computing grid
CiGri (Ciment Grid) is a lightweight computing grid based on top of Gricad’s OAR clusters.
Head node: cigri from the bastions.
- Computing grid giving access to all clusters from a single entry point.
- Allows efficient management of parametric job campaigns (bag of tasks / embarassingly parallel).
- Automatic and transparent resubmission.
Storage infrastructures
- High performance distributed storage
- Accessible as local filesystem from Luke and Dahu
- Filesystem mounted on
/bettik - User folder and file creation and management
- Default access rights to be adjusted by the user
- Usage information on the CiGri grid server under “Bettik BeeGFS Storage”
- Very high performance distributed storage (scratch) on Omnipath network
- Accessible as a local filesystem from Dahu and Bigfoot
- Filesystem mounted on the
/silenus directory - Creation and management of folders and files by users
- Default rights to be adjusted by the user
- Usage information on Silenus
Mantis, cloud storage
- Cloud block mode storage
- Accessible from all clusters
- Access rights management and sharing with other users
- Accessible from the IDRIS AdaPP machine
- Usage information on the CiGri grid server under “Mantis iRODS Storage”