The Turbulence Modeling and Simulation team deploys a number of computational resources specific to its activities in addition to those available at LEGI.
The high availability virtualization cluster
Based on 3 DELL R530 (quorum), this high-availability virtualization cluster is running Proxmox-VE and provides all the critical services to support (nearly) H24, 7D/7 availability. This concerns the name resolution, authentication, clusters management ... but also the critical data storage which is distributed and replicated among the nodes (CEPH). It allows hot migrations of virtual machines (without service disruption) and provides resilience, without human intervention, to most technical failures.
The clusters for distributed parallel computing
- Kareline (available since 2016) is a 400-cores (Intel Xeon) cluster built around a fast (40Gbits/s) and low latency infiniband QDR network to specifically meet the needs of distributed highly parallel simulations run by the team.
- Tenibre (available since 2021) is a 256-cores (AMD Epyc2) cluster designed for memory-bound applications (memory bandwidth is close to 205GB/s/socket). The computational nodes are connected by a 100Gb/s network (infiniband HDR100) and data access also relies on this fast network. Moreover this cluster has a dual GPU node (Nvidia A100) to develop hybrid computing approaches and to support the team’s new research topics.
- Osorno (available since 2025) is a cluster of 480 AMD Genoa cores, also specifically sized for codes that are highly dependent on memory bandwidth (more than 400 GB/s/48-core processor). These nodes are interconnected at 100 Gb/s by an HDR100 infiniband network for computing and by a 25 Gb/s Ethernet network to a distributed storage cluster for data.
The management of these three clusters is based on BlueBanquise. They enable the development of our simulation codes and to run the meso-resolution calculations, upstream of the large supercomputers of the national centers.
The nodes dedicated to 3D vizualization and large data processing
- A DELL R730 server for intensive 3D visualizations, configured with two 10-core Haswell-EP Xeons (E5-2660 v3), 512GB of RAM, and an AMD Firepro W7000 GPU. It enables large data volumes to be viewed using remote session solutions.
- A pre and post-processing server, with 4 Xeons gold 6148 (80 cores in total) and 1.5TB of RAM allows to process very large data resulting from simulations at very high resolutions on the GENCI supercomputers.