DRAGNET GPU Cluster Specifications

Raw, theoretical specifications of the cluster. More details can be found in the tender offer.

Cluster benchmark numbers (absolute max achievable speeds for applications)

This is what the cluster looks like (except that we have 6x LAG (red) in place):

• 1 head node
• 1 processing node
• 23 worker nodes

dragnet.control.lofar: login, home dirs + file server (installed programs) (NFS), job scheduling

• Intel Xeon Haswell-EP E5-1620 v3, 4 cores (8 hw threads), 3.50 GHz
• 32 GiB RAM
• 2x 1 Gb eth network
• 2x 4 TB storage (RAID-1, i.e. 4 TB netto)

dragproc.control.lofar: central post-processing, database, reliable storage, secondary/backup node for login/file server/job scheduling

• dual Intel Xeon Haswell-EP E5-2630 v3, 16 cores (32 hw threads), 2.40 GHz
• 128 GiB RAM
• 2x 1 Gb eth network
• 10 Gb eth network
• 8x 4 TB storage (RAID-6, i.e. 24 TB netto) on /data

drg[01-23].control.lofar: bulk data processing, receive data from COBALT, scratch storage

• dual Intel Xeon Haswell-EP E5-2630 v3, 16 cores (32 hw threads), 2.40 GHz
• 4x NVIDIA GeForce GTX Titan X, 12 GiB RAM each (all GPUs total: 565.25 TFlop/s (FP32))
• 128 GiB RAM
• 2x 1 Gb eth network
• 10 Gb eth network
• 56 Gb FDR infiniband network
• 4x 4 TB storage (dual RAID-0, i.e. 16 TB netto; all nodes total: 368 TB) on /data1 and /data2 (Note: local per node and not redundant!)

• infiniband
• 10 Gb ethernet
• dual 1 Gb ethernet

• Mellanox MSX 6025F-1SFS - 36 port unmanaged FDR infiniband switch
• 5x 56G LAG to COBALT infiniband switch (~280 Gb)

• Supermicro SSE-X3348S Ethernet Switch; Ports: 48x 10 Gb, 4x 40 Gb, 2x 1 Gb
• 6x 10G LAG to LOFAR Core 10 Gb network

• 2x Supermicro SSE-G2252 Ethernet Switch

• 4 years support contract on the complete system (delivery was on Thu+Fri 9+10 July 2015)
• delivery + installation