This is an old revision of the document!
LOEWE-CSC
The LOEWE-CSC is a general-purpose computer cluster based on several CPU and GPU architectures. The system was installed in Industriepark Höchst in late 2010 and has been constantly upgraded since then.
Cluster performance:
- CPUs performance (dp)1): 226 TFlop/s (peak)
- GPUs performance (dp): 597 TFlop/s (peak)
- Cluster performance HPL (2010): 299.3 TFlop/s
- Energy efficiency Green500 (2010): 740.78 MFlop/s/Watt
Hardware:
- 848 compute nodes in 36 water-cooled racks, two server rooms,
- thousands of CPU cores and hundreds of GPGPU accelerators,
- 64-128 of RAM per node and over 2 PB aggregated disk capacity,
- QDR and FDR InfiniBand interconnects,
- a parallel scratch filesystem with a capacity of 764 TB and an aggregated bandwidth of 10 GB/s.
| Nodes | CPU | Cores per CPU | Cores per Node | Threads2) per Node | RAM per Node | GPU |
|---|---|---|---|---|---|---|
| 600 | 2xAMD Opteron 6172 | 12 | 24 | 24 | 64 GB | AMD HD 5800 |
| 198 | 2xIntel Xeon E5-2670 v2 | 10 | 20 | 40 | 128 GB | n/a |
| 50 | 2xIntel Xeon E5-2630 v2 | 6 | 12 | 24 | 128 GB | 2xAMD S10000 |
InfiniBand Network
Most of the compute nodes (the ones in the main server room) are connected through a mixed 4X QDR-FDR InfiniBand 2:1 blocking fat-tree network (FDR at the spine level, FDR and QDR at the edge level). In the second (smaller) server room there are two non-blocking InfiniBand “islands” (one per cabinet), one running at 4X QDR and one at 4X FDR. The connection between the rooms is a 20 Gb/s Ethernet. Compute jobs are scheduled in a way that avoids mixed allocations, i.e. jobs are either kept within the fat-tree or on an island.
