- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -
- Information contained in the HLRS Wiki is not legally binding and HLRS is not responsible for any damages that might result from its use -
HPE Hawk Hardware and Architecture: Difference between revisions
(Created page with "=== Node/Processor === Compute nodes as well as login nodes are equipped with AMD EPYC 7742 64-Core Processor detailed information will be provided later. Please check for...") |
|||
| (26 intermediate revisions by 2 users not shown) | |||
| Line 1: | Line 1: | ||
=== Node/Processor === | === Node/Processor === | ||
==== hawk compute nodes ==== | |||
* CPU type: AMD EPYC 7742 | |||
* 2 CPU's / node | |||
* 64 Cores / CPU | |||
* CPU frequency: 2.25 GHz | |||
* 256 GB / node | |||
* 5632 nodes (until 2024-06-18), 4096 nodes (from 2024-06-18) | |||
With respect to details of the processor deployed in Hawk, please refer to this [https://kb.hlrs.de/platforms/upload/Processor.pdf Slides]. | |||
==== Pre- and post processing ==== | |||
Within HLRS simulation environment special nodes for pre- and post processing tasks are available. Such nodes could be requested via the [[Batch_System_PBSPro_(Hawk) | batch systems]] using the smp queue. | |||
Available nodes are | |||
4 nodes 2 TB Memory, 2 Socket AMD EPYC 7702 64-Core, shared usage model | |||
1 Node 4 TB Memory, 2 Socket AMD EPYC 7702 64-Core, shared usage model | |||
more specialized nodes e.g. graphics, vector, DataAnalytics, ... are available in the [[NEC_Cluster_Hardware_and_Architecture_(vulcan)|Vulcan cluster]].<BR> | |||
If you need such specialized nodes on vulcan cluster for pre- or postprocessing inside your project located on hawk resources, please ask your project manager for access to vulcan. | |||
==== AI nodes ==== | |||
For AI compute jobs there is a special part of nodes with GPU's available: | |||
* number nodes: 24 | |||
* CPU TYPE: 2 Socket AMD EPYC 7702 64-Core | |||
* memory per node: 1TB | |||
* GPU's per node: 8 | |||
* GPU type: NVIDIA A100 | |||
* GPU memory: 20 nodes with 40 GB, and 4 nodes with 80 GB | |||
* Node to node interconnect: Dual Rail InfiniBand HDR200 | |||
* local disk capacity per node: 15TB | |||
Nodes can be used via the [[Batch_System_PBSPro_(Hawk) | batch systems]] by requesting the special node_type's (rome-ai, nv-a100-40gb, nv-a100-80gb). | |||
see also [[Big_Data,_AI_Aplications_and_Frameworks]] | |||
<br> | <br> | ||
=== Interconnect === | === Interconnect === | ||
Hawk deploys an Infiniband HDR based interconnect with a | Hawk deploys an Infiniband HDR based interconnect with a 8-dimensional hypercube topology. Please refer to [https://kb.hlrs.de/platforms/upload/Interconnect.pdf here] with respect to the latter. Infiniband HDR has a bandwidth of 200 Gbit/s and a MPI latency of ~1.3us per link. The full bandwidth of 200 Gbit/s can be used if communicating between the 16 nodes connected to the same node of the hypercube (cf. [https://kb.hlrs.de/platforms/upload/Interconnect.pdf here]). Within the hypercube, the higher the dimension, the less bandwidth is available. | ||
Topology aware scheduling is used to exclude major performance fluctuations. This means that larger jobs can only be requested with defined node numbers (64, 128, 256, 512, 1024 | Topology aware scheduling is used to exclude major performance fluctuations. This means that larger jobs can only be requested with defined node numbers (64, 128, 256, 512, 1024) in regular operation. This restriction ensures optimal system utilization while simultaneously exploiting the network topology. Jobs with a node number of < 128 nodes are processed in a special partition. Jobs from 2048 nodes and more are processed at special times (called XXL days). Please ask if needed. <br> | ||
With respect to further details, please refer to the [https://kb.hlrs.de/platforms/upload/Interconnect.pdf Slides] already referenced above. | |||
<br> | <br> | ||
=== Filesystem === | === Filesystem === | ||
Available lustre filesystems on hawk: | |||
* ws11: | |||
** available storage capacity: 15 PB | |||
** lustre devices: 2 MDS, 2 MDT, 20 OSS, 40 OST | |||
** performance: ~200 GiB/s | |||
Additional an central HOME and project fileserver is also mounted on hawk. | |||
Some special nodes have a local disk installed which can be uses as localscratch. | |||
See also [[Storage_(Hawk)| Storage (Hawk)]] | |||
<br> | <br> | ||
Latest revision as of 10:57, 27 November 2024
Node/Processor
hawk compute nodes
- CPU type: AMD EPYC 7742
- 2 CPU's / node
- 64 Cores / CPU
- CPU frequency: 2.25 GHz
- 256 GB / node
- 5632 nodes (until 2024-06-18), 4096 nodes (from 2024-06-18)
With respect to details of the processor deployed in Hawk, please refer to this Slides.
Pre- and post processing
Within HLRS simulation environment special nodes for pre- and post processing tasks are available. Such nodes could be requested via the batch systems using the smp queue. Available nodes are
4 nodes 2 TB Memory, 2 Socket AMD EPYC 7702 64-Core, shared usage model
1 Node 4 TB Memory, 2 Socket AMD EPYC 7702 64-Core, shared usage model
more specialized nodes e.g. graphics, vector, DataAnalytics, ... are available in the Vulcan cluster.
If you need such specialized nodes on vulcan cluster for pre- or postprocessing inside your project located on hawk resources, please ask your project manager for access to vulcan.
AI nodes
For AI compute jobs there is a special part of nodes with GPU's available:
- number nodes: 24
- CPU TYPE: 2 Socket AMD EPYC 7702 64-Core
- memory per node: 1TB
- GPU's per node: 8
- GPU type: NVIDIA A100
- GPU memory: 20 nodes with 40 GB, and 4 nodes with 80 GB
- Node to node interconnect: Dual Rail InfiniBand HDR200
- local disk capacity per node: 15TB
Nodes can be used via the batch systems by requesting the special node_type's (rome-ai, nv-a100-40gb, nv-a100-80gb).
see also Big_Data,_AI_Aplications_and_Frameworks
Interconnect
Hawk deploys an Infiniband HDR based interconnect with a 8-dimensional hypercube topology. Please refer to here with respect to the latter. Infiniband HDR has a bandwidth of 200 Gbit/s and a MPI latency of ~1.3us per link. The full bandwidth of 200 Gbit/s can be used if communicating between the 16 nodes connected to the same node of the hypercube (cf. here). Within the hypercube, the higher the dimension, the less bandwidth is available.
Topology aware scheduling is used to exclude major performance fluctuations. This means that larger jobs can only be requested with defined node numbers (64, 128, 256, 512, 1024) in regular operation. This restriction ensures optimal system utilization while simultaneously exploiting the network topology. Jobs with a node number of < 128 nodes are processed in a special partition. Jobs from 2048 nodes and more are processed at special times (called XXL days). Please ask if needed.
With respect to further details, please refer to the Slides already referenced above.
Filesystem
Available lustre filesystems on hawk:
- ws11:
- available storage capacity: 15 PB
- lustre devices: 2 MDS, 2 MDT, 20 OSS, 40 OST
- performance: ~200 GiB/s
Additional an central HOME and project fileserver is also mounted on hawk. Some special nodes have a local disk installed which can be uses as localscratch.
See also Storage (Hawk)