Machine Specifications
- Nick Maggio (Unlicensed)
- Mark Allen
- Rob Yelle (Unlicensed)
Information on the hardware that makes up the Talapas Cluster
Compute Hardware
The Talapas HPC cluster contains three categories of nodes: standard nodes, GPU nodes, and large memory nodes. Nodes that are owned by the University are made available to all researchers via our compute club. Nodes that are owned by PIs, or condo nodes, have restricted access as dictated by the PI.
Shared Nodes
Qty | Node Type | Processors (total cores) | Memory | Local Storage | Networking | Accelerator |
|---|---|---|---|---|---|---|
96 | Standard Nodes | dual E5-2690v4 (28 cores) | 128GB | 200GB SSD | Single Port EDR InfiniBand | N/A |
24 | GPU Nodes | dual E5-2690v4 (28 cores) | 256GB | 200GB SSD | Single Port EDR InfiniBand | Dual NVIDIA Tesla K80 |
8 | Large Memory Nodes | quad E7-4830v4 (56 cores) | 1TB, 2TB, or 4TB | dual 480GB SSD | Single Port EDR InfiniBand | N/A |
Condo Nodes
Qty | Node Type | Processors (total cores) | Memory | Local Storage | Networking | Accelerator |
|---|---|---|---|---|---|---|
82 | Standard Nodes | dual Gold 6148 (40 cores) | 192GB, or 384GB | 240GB SSD | Dual Port EDR InfiniBand | N/A |
Storage Hardware
All compute resources are connected to our DDN GRIDScaler 14k storage appliance via the EDR InfiniBand interconnect. The DDN appliance runs GPFS and provides over 1.5 PB of usable storage.
Appliance | Enclosures | Drives | Filesystem | Usable Space |
|---|---|---|---|---|
DDN GS14k | 5 SS8462 84-slot enclosures | 10 800GB Mixed Use SSD Drives (metadata) 362 6TB 7200 RMB 12Gb/s SAS 4Kn drives 21 800GB Mixed Use SSD Drives (fast tier) | GPFS | 1,579 TiB |
Interconnect
All compute nodes and the DDN storage controllers are connected via a high speed EDR InfiniBand network providing 100Gbit/s throughput. The network is arranged in a "fat-tree" topology with compute nodes connected to leaf switches and all leaf switches connected to core switches. Currently, the network is configured with a 2:1 overall blocking ratio, i.e. there are twice the number of connections up from the nodes to the leaf swithces as there are from the leaf switches to the core switches. This allows us to scale economically while providing unblocked communication between the 24 compute nodes that share a common leaf switch. The two DDN storage controllers each have dual connections to the InfiniBand core switches
