Network
ICARUS uses a redundant 100 GbE Ethernet network to connect compute nodes, login nodes, storage, and the rest of the infrastructure.
Each tested compute and login node uses two 100 Gbit/s network links connected to separate high-speed switches. The links are combined using LACP, providing up to 200 Gbit/s aggregate network capacity per node and redundancy if one link becomes unavailable.
High-Speed Network
| Component | Configuration |
|---|---|
| High-speed switches | 2 × NVIDIA/Mellanox MSN4600C |
| Network speed per link | 100 Gbit/s |
| Links per node | 2 |
| Aggregate capacity per node | 200 Gbit/s |
| Network adapters | NVIDIA/Mellanox ConnectX-6 Dx |
| Link aggregation | IEEE 802.3ad / LACP |
| Jumbo frames | MTU 9216 |
| RDMA support | Available over Ethernet |
The 200 Gbit/s figure is aggregate capacity; a single connection may use only one 100 Gbit/s link.
Compute and Login Networks
Compute nodes communicate through the private ICARUS network. Login nodes connect this internal network with the NTUA network and provide the main user access point to the cluster.
MPI and other communication libraries use the provided network; users do not need to configure interfaces.
Additional Network
| Component | Count | Specification |
|---|---|---|
| NVIDIA/Mellanox MSN4600C | 2 | 64-port 100 GbE switches |
| HPE Aruba CX 6300M | 4 | 48-port 1 GbE switches |
The 100 GbE network is used for high-bandwidth system communication, while the additional Ethernet infrastructure supports management and auxiliary network services.