Occigenincludes 34 racks :
The cluster is composed of three “parts” :
Part 1 | Part 2 | Vizualisation | Large | |
Ref. Manufacturer | Bull B720 | Bull B720 | Bull R421-E4 | Bull Sequana X800 |
Processor name | Haswell | Broadwell | Broadwell | Skylake |
Ref. processor | E5-2690V3@2.6 GHz | E5-2690 V4@2.6GHz | E5-2690 V4@2.6GHz | Xéon Platinum 8176@2.1GHz |
Nb of nodes | 2106 | 1260 | 4 | 1 |
Processors by node | 2 | 2 | 2 | 8 |
Freq. of processors | 2.6 GHz | 2.6 GHz | 2.6 GHz | 2.1 GHz |
Cores by processor | 12 | 14 | 14 | 28 |
Cache size L1 |
12 X 32 Ko instr. 12 X 32 Ko data |
14 X 32 Ko instr. 14 X 32 Ko data |
14 X 32 Ko instr. 14 X 32 Ko data |
28 X 32 Ko instr. 28 X 32 Ko data |
Cache size L2 | 12 X 256 Ko | 14 X 256 Ko | 14 X 256 Ko | 28 X 1 Mo |
Cache size LLC | 30 Mo | 35 Mo | 35 Mo | 38.5 Mo |
Nb of memory channels | 4 | 4 | 4 | 6 |
Memory by node |
1053 X 64 Go 1053 X 128 Go |
64 Go | 256 Go | 3 To |
Memory ref. | DDR4-2133P-R | DDR4-2400T-R | DDR4-2400T-R | DDR4-2666V-R |
Network attachment | Infiniband FDR 56 Gbit/s | Infiniband FDR 56 Gbit/s | Infiniband FDR 56 Gbit/s | Infiniband FDR 56 Gbit/s |
Type of GPU | Nvidia Tesla P100 PCIe 12Go | Nvidia Tesla P100 PCIe 12Go | ||
Nb GPU by node | 1 | 2 | ||
Nb total cores | 50544 | 35280 | 112 | 224 |
The computings racks are connected to five racks mounted on a Lustre shared filesystem with a total capacity of 5 PB useful.
Cooling is provided by a high efficiency warm water system directly in the nodes (mode DLC Direct Liquid Cooling).
The request for allocation of computing hours on this cluster is the subject of two campaigns per year(autumn and spring) through the DARI procedure.
The machine is cut into a rack. A rack includes :
In total, the Occigen cluster consists of 3367computenodes and therefore has 86048cores.
The network used to connect the nodes to each other is an Infiniband network (IB 4x FDR). The topology of the network is in the form of a Fat tree pruned. The network is non blocking within the chassis. Each group of 18 nodes that share the same switch in a chassis can be reached without restrictions. The “ascending” links from a switch chassis to the higher-order switches are divided by two. For 18 nodes, only 9 uplinks are used.
There are two types of filesystems :
To store the results more securely, each node of the machine accesses the /store file system. This is also a Lustre filesystem, but with advanced security mechanisms (duplicate storage and tape storage).It must be used to ensure that the results are properly stored.