- 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 -

NEC Cluster Hardware and Architecture (laki + laki2): Difference between revisions

From HLRS Platforms
Jump to navigationJump to search
No edit summary
 
(4 intermediate revisions by 2 users not shown)
Line 45: Line 45:
* '''Node Upgrades''' (2016/17)
* '''Node Upgrades''' (2016/17)
** 360 nodes Dual Intel [[hsw|'Haswell']] [http://ark.intel.com/de/products/81908/Intel-Xeon-Processor-E5-2680-v3-30M-Cache-2_50-GHz E5-2680v3]
** 360 nodes Dual Intel [[hsw|'Haswell']] [http://ark.intel.com/de/products/81908/Intel-Xeon-Processor-E5-2680-v3-30M-Cache-2_50-GHz E5-2680v3]
*** 2.5 Ghz, 12 Cores per processor, 20 Threads
*** 2.5 Ghz, 12 Cores per processor, 24 Threads
*** 4 memory channels per processor, DDR4 2133Mhz memory
*** 4 memory channels per processor, DDR4 2133Mhz memory
*** 344 nodes with 128GB RAM (''hsw128gb12c'')  
*** 344 nodes with 128GB RAM (''hsw128gb12c'')  
Line 58: Line 58:
** QDR Mellanox ConnectX-2 IB HCAs (40gbit)
** QDR Mellanox ConnectX-2 IB HCAs (40gbit)
** 4 nodes have additional a 4TB local scratch disk
** 4 nodes have additional a 4TB local scratch disk
* '''(Nov. 2017) 10 Additional GPU graphic nodes with'''
** Nvidia Tesla P100 12GB
** 2 sockets ech 8 cores (Intel E5-2667v4 @ 3.2GHz)
** 256GB memory
** 3.7TB /localscratch, 400GB SSD /tmp


* '''network''': [http://de.wikipedia.org/wiki/Infiniband InfiniBand] Double Data Rate
* '''network''': [http://de.wikipedia.org/wiki/Infiniband InfiniBand] Double Data Rate
Line 93: Line 100:
* Large Memory nodes (144GB, 256GB)
* Large Memory nodes (144GB, 256GB)
* Pre-Postprocessing node with very large memory (1TB)
* Pre-Postprocessing node with very large memory (1TB)
* Visualisation nodes with Nvidia Quadro FX5800
* Visualisation nodes with Nvidia Quadro FX5800 or Nvidia Tesla P100
* Different memory nodes (<s> 12GB, 24GB,</s> 32GB, <s>48GB</s>, 64GB, 128GB, 256GB)
* Different memory nodes (<s> 12GB, 24GB,</s> 32GB, <s>48GB</s>, 64GB, 128GB, 256GB)



Latest revision as of 16:22, 16 August 2018

Hardware

  • ~150 compute nodes are of type NEC HPC-144 Rb-1 Server (see NEC Products)
    • dual CPU compute nodes: 2x Intel Xeon X5560 Nehalem EP "Gainestown" (5000 Sequence specifications)
      • 4 cores, 8 threads
      • 2.80 GHz (3.20 Ghz max. Turbo frequency)
      • 8MB L3 Cache
      • 1333 MHz Memory Interface, 6.4 GT/s QPI
      • TDP 95W, 45nm technology
      • "Nehalem" microarchitecture
    • compute node RAM: triple-channel memory
      • standard: 12 GB RAM (nehalem/mem12gb)
      • 20 nodes upgraded to 24GB (mem24gb), 48GB (mem48gb) or 144GB (mem144gb) RAM
        • 2 nodes of 144GB Memory nodes have additional a 6TB local scratch disk installed
        • 1 nodes of 144GB Memory nodes have additional a 2TB local scratch disk installed
    • 16 compute nodes have additional Nvidia Tesla S1070 GPU's installed.

  • Pre- & Postprocessing node (smp node)
    • 8x Intel Xeon X7542 6-core CPUs with 2.67GHz (8*6=48 Cores)
    • 1TB RAM
    • shared access
  • Visualisation node (vis)
    • 5 nodes each with 8 cores Intel W3540 and 24GB memory (4 for laki and 1 for laki2)
      • Nvidia Quadro FX5800
  • Node Upgrades (2012/2013)
    • 128 nodes Dual Intel 'Sandy Bridge' E5-2670 (204 for laki and 124 for laki2)
      • 2.6 Ghz, 8 Cores per processor, 16 Threads
      • 4 memory channels per processor, DDR3 1600Mhz memory
      • 96 nodes with 32GB RAM (sb/mem32gb)
      • 30 nodes with 64GB RAM (mem64gb)
      • QDR Mellanox ConnectX-3 IB HCAs (40gbit)
  • Node Upgrades (2014/2015)
    • 80 nodes Dual Intel 'Haswell' E5-2660v3
      • 2.6 Ghz, 10 Cores per processor, 20 Threads
      • 4 memory channels per processor, DDR4 2133Mhz memory
      • 76 nodes with 128GB RAM (hsw128gb10c)
      • 4 nodes with 256GB RAM (hsw256gb10c)
      • QDR Mellanox ConnectX-3 IB HCAs (40gbit)
  • Node Upgrades (2016/17)
    • 360 nodes Dual Intel 'Haswell' E5-2680v3
      • 2.5 Ghz, 12 Cores per processor, 24 Threads
      • 4 memory channels per processor, DDR4 2133Mhz memory
      • 344 nodes with 128GB RAM (hsw128gb12c)
      • 16 nodes with 256GB RAM (hsw256gb12c)
      • QDR Mellanox ConnectX-3 IB HCAs (40gbit), 144 of the 128GB nodes have fdr IB, (fdr)
  • Additional large memory nodes
    • 10 nodes Quad Socket AMD Opteron 6238
    • 2.6 Ghz, 12 cores per processor
    • 4 memory channels per processor, DDR3 1600Mhz memory
    • 256GB RAM (mem256gb)
    • QDR Mellanox ConnectX-2 IB HCAs (40gbit)
    • 4 nodes have additional a 4TB local scratch disk
  • (Nov. 2017) 10 Additional GPU graphic nodes with
    • Nvidia Tesla P100 12GB
    • 2 sockets ech 8 cores (Intel E5-2667v4 @ 3.2GHz)
    • 256GB memory
    • 3.7TB /localscratch, 400GB SSD /tmp




Architecture

The NEC Cluster platform (laki and laki2) consists of several frontend nodes for interactive access (for access details see Access) and several compute nodes of different types for execution of parallel programs.


Compute node types installed: * Intel Xeon 5560 (nehalem)

  • Intel Xeon E5-2670 (Sandy Bridge)
  • AMD Opteron 6238 (interlagos)
  • Intel E5-2680v3 and E5-2660v3

* Nvidia Tesla S1070 (consisting of C1060 devices)

  • Large Memory nodes (144GB, 256GB)
  • Pre-Postprocessing node with very large memory (1TB)
  • Visualisation nodes with Nvidia Quadro FX5800 or Nvidia Tesla P100
  • Different memory nodes ( 12GB, 24GB, 32GB, 48GB, 64GB, 128GB, 256GB)


Features

  • Operating System: ScientificLinux 6.9 (internal test was done on Windows HPC Server 2008)
  • Batchsystem: Torque/Moab
  • node-node interconnect: Infiniband + GigE
  • Global Disk 500 TB (lustre) for laki + 500TB (lustre) for laki2
  • Many Software Packages for Development