The OSIRIM platform consists of a 640 core computing cluster distributed over 17 computing servers + 28 Nvidia GTX 1080 TI cards (3584 cuda cores each) distributed over 7 servers. This cluster is structured as follows :
- User connection nodes
These nodes are used to validate the programs before launching them on the compute cluster. These interactive nodes are shared among all users and should not be used for running long jobs.
- Compute nodes (co2-nc01, …, co2-nc10)
These nodes are servers dedicated to calculations. A process running on a compute node accesses data hosted on the storage area, performs a process, and records the result on that area.
The technical characteristics of the 10 nodes of physical computations (10 x 512 GB of RAM and 64 cores) are the following ones :
- 10 IBM servers X3755 M3
- 4 AMD Opteron 6262HE processors with 16 x 1,6Ghz cores
- 512 Go RAM
- 2 x 300 Go RAID1 SSD disk
- 2 x 10Gb/s network
- Compute nodes (co2-ng01, …, co2-ng07)
These nodes are servers dedicated to GPU calculations. Each of the 7 servers is equipped with 4 Nvidia Geforce GTX 1080TI graphics cards.
The technical characteristics of the 7 nodes are the following ones :
- 7 DELL T630 servers
- 2 Xeon 2640 V4 processors (20 threads)
- 2 x 400 Go RAID1 SSD disk
- 1 x 10 Gb/s network
- 4 GPUs Nvidia GTX 1080 TI (3584 cuda cores each)
A process running on a compute node accesses data hosted on the storage area, performs a process, and records the result on that area. The data storage on the OSIRIM platform is provided by an EMC Isilon bay composed of :
- 12 X400 nodes of 36 SATA drives of 3 TB each
- each node is connected to the network via a trunk of 2 links 10Gb/s
- the useful volume is about 1 Po
OneFS is the OS of the storage array that integrates both the file system, volume management, and data security.
The set consists of a single distributed file system, with a single namespace, which has the ability to present data to servers following multiple protocols. NFS and HDFS are the access protocols used on OSIRIM.