Helix/Hardware: Difference between revisions
S Richling (talk | contribs) |
S Richling (talk | contribs) |
||
| Line 83: | Line 83: | ||
|} |
|} |
||
The lines in blue are relevant for [https://wiki.bwhpc.de/e/Helix/Slurm Slurm batch scripts]. |
The lines marked in blue are relevant for [https://wiki.bwhpc.de/e/Helix/Slurm Slurm batch scripts]. |
||
=== Intel Nodes === |
=== Intel Nodes === |
||
Revision as of 15:05, 9 August 2024
System Architecture
The bwForCluster Helix is a high performance supercomputer with high speed interconnect. The system consists of compute nodes (CPU and GPU nodes), some infrastructure nodes for login and administration and a storage system. All components are connected via a fast Infiniband network. The login nodes are also connected to the Internet via Baden Württemberg's extended LAN BelWü.
Operating System and Software
- Operating system: RedHat
- Queuing system: Slurm
- Access to application software: Environment Modules
Compute Nodes
AMD Nodes
Common features of all AMD nodes:
- Processors: 2 x AMD Milan EPYC 7513
- Processor Frequency: 2.6 GHz
- Number of Cores per Node: 64
- Local disk: None
| CPU Nodes | GPU Nodes | ||||
|---|---|---|---|---|---|
| Node Type | cpu | fat | gpu4 | gpu4 | gpu8 |
| Quantity | 355 | 15 | 29 | 26 | 3 |
| Installed Working Memory (GB) | 2048 | 256 | 256 | 2048 | |
| Available Memory for Jobs (GB) | 236 | 2010 | 236 | 236 | 2010 |
| Interconnect | 1x HDR100 | 1x HDR100 | 2x HDR100 | 2x HDR200 | 4x HDR200 |
| Coprocessors | - | - | 4x Nvidia A40 (48 GB) | 4x Nvidia A100 (40 GB) | 8x Nvidia A100 (80 GB) |
| Number of GPUs | - | - | 4 | 4 | 8 |
| GPU Type | - | - | A40 | A100 | A100 |
The lines marked in blue are relevant for Slurm batch scripts.
Intel Nodes
Some Intel nodes (Skylake and Cascade Lake) from the predecessor system will be integrated. Details will follow.
Storage Architecture
There is one storage system providing a large parallel file system based on IBM Spectrum Scale for $HOME, for workspaces, and for temporary job data.
Network
The components of the cluster are connected via two independent networks, a management network (Ethernet and IPMI) and an Infiniband fabric for MPI communication and storage access. The Infiniband backbone is a fully non-blocking fabric with 200 Gb/s data speed. The compute nodes are connected with different data speeds according to the node configuration.