Deucalion
Architecture
Compute nodes
Deucalion has a heterogeneous architecture and is the first in Europe to incorporate arm and x86 microprocessors in the same cluster. The following table will give you insight into Deucalion compute nodes components.
Partition | Compute Nodes | CPU | GPU | Memory | Disk | Network |
---|---|---|---|---|---|---|
arm | 1632 | Fujitsu A64FX 48-core 2.0 GHz | No GPU | 32GB | WD CL SN720 512GB PCIe | NVIDIA ConnectX-6 100 Gb/s |
x86 | 500 | 2x AMD EPYC 7742 64-core 2.25 GHz | No GPU | 256GB DDR4 | 480GB SSD | NVIDIA ConnectX-6 200 Gb/s |
a100-40 | 17 | 2x AMD EPYC 7742 64-core 2.25 GHz | 4x NVIDIA Ampere A100 40 GB | 512 GB DDR4 | Micron 5300 480GB SATA | 2x NVIDIA ConnectX-6 200 Gb/s |
a100-80 | 16 | 2x AMD EPYC 7742 64-core 2.25 GHz | 4x NVIDIA Ampere A100 80 GB | 512 GB DDR4 | Micron 5300 480GB SATA | 2x NVIDIA ConnectX-6 200 Gb/s |
Software
Component | Description | Version |
---|---|---|
Operating System | Rocky Linux | 8 |
Software environment/package system | Modules | 8.5 |
Containers | Singularity | 3.11 |
Resource manager and scheduling | Slurm | 23.11.4 |
Storage
Purpose | Overview | File System | Mount points |
---|---|---|---|
High Speed Storage | Metadata and Hot Pools NVMe – 430TB HDD Datapools – 10PB MDS Servers – 8 Nodes OSS I/O Servers – 32 Nodes Aggregated Performance - 340GB/s in reads, 260GB/s in writes |
Lustre Parallel File System | /share /projects |
NAS Storage | Total SSD net Capacity – 50TB Data Modules – 2 for redundancy Connection type – 8x10 GbE |
NFSv4 | /home /apps |
Network
Segment | Switch model | Units | Ports | Network Operting System |
---|---|---|---|---|
Top of Rack | NVIDIA Mellanox AS4610 | 64 | 48x 1G Base-T 4x 10G SFP+ |
Cumulus Linux |
Core | NVIDIA Mellanox SN2410 | 2 | 48x 25G SFP28 8x 100G QSFP |
Cumulus Linux |
Core | NVIDIA Mellanox SN2010 | 4 | 18x 25G SFP28 4x 100G QSFP |
Cumulus Linux |
ARM Infiniband
Infiniband Fat-Tree with 1:6 blocking factor using 50 switches NVIDIA Mellanox QM8790 with NVIDIA MLNX-OS
Segment | Units | IB links per switch |
---|---|---|
Leaf | 35 | 24 downlinks 15 uplinks |
Spine | 15 | 35 downlinks 5 uplinks |
x86 Infiniband
Infiniand Fat-Tree with 1:1 non-blocking factor using 27 switches NVIDIA Mellanox QM8790 with NVIDIA MLNX-OS
Segment | Units | IB links per switch |
---|---|---|
Leaf | 17 | 20 downlinks 20 uplinks |
Spine | 10 | 34 doanlinks 6 uplinks |