Original URL: https://www.theregister.com/2014/06/30/isc_leipzig_roundup/
Rare HPC beauties unveiled: Quivering racks, Lustre clusters and the tiers of a Cray
It's all getting a bit terafloppy in Deutschland
Posted in HPC, 30th June 2014 14:24 GMT
ISC Leipzig The International Supercomputer Show in Leipzig, Germany, was full of fascinating things at the high-end grunt front of the computing business. Here's what attracted this roving hack's eye.
- Bull has started a world-wide reseller agreement with Seagate’s Xyratex ClusterStor business for supercomputer storage. The ClusterStor arrays will be integrated with bullx supercomputers. ClusterStor uses Lustre. Bull has signed a deal with “ Deutsches Klimarechenzentrum (DKRZ) … to design and install a 3-petaflop bullx B720 supercomputer with 45 petabytes (45PB) of ClusterStor CS9000 parallel storage capable of 1TB/s performance.” DKRZ is a climate computing centre in Hamburg.
- Cray has a tiered storage scheme for Lustre. It’s called Tiered Adaptive Storage (TAS) and provides hierarchical storage management (HSM) embracing solid state, disk and tape storage tiers. TAS uses Lustre APIs and this TAS connector runs on any Linux or Lustre environment. Data access is transparent to users, irrespective of the storage tier. File migration policy can be based on quotas and the management tools should be familiar to system administrators and SAM-QFS users.
Cray has also joined the OpenStack Foundation. It will contribute to OpenStack and work to integrate open source capabilities into future Cray products and services "to benefit the supercomputing industry".
- The US NCSA (National Centre for Supercomputing Applications) has deployed a compute and storage environment as part of the Illinois Campus Cluster Program, using Direct Networks' SFA storage array. It’s called a “storage condo” because it will be used by 22 different researchers and faculties across the campus, each putting in some money to pay for it. The initial 1PB or so capacity could double in the near future. Read a case study here.
- Fast LTA, with LTA standing for Long Term Archive, had a Silent Brick Library array on display.
This was a rack mount shelf with carousels containing twelve 1TB 2.5" disk drives from three different manufacturers to minimise failures happening inside a manufacturer's batch of disks. Data is written to the library linearly, from carousel to carousel, and striped across the drives in a carousel (or brick). Erasure coding is the data protection mechanism and each brick is self-contained in that sense. Once a brick is full then power to it can be switched off, coming on again when data access is needed. A rack can hold up to 60TB of uncompressed data with 12 bricks.
Silent Brick with open disk drive carousel
If you need faster data access than a tape archive can provide yet minimise on power consumption then this looks pretty good.
- Panasas says Genedata has chosen its scale-out NAS ActiveStor 14 array for RNA/DNA sequence data analysis work. Genedata, which develops and markets software for statistical analysis and scientific interpretation of extremely large biological datasets, says the kit solves an IO bandwidth problem and so keeps its computing cores busier.
- Samsung's large stand had a 3D V-NAND focus with a canned presentation being cycled through on a display. We snapped two slides from this:
30nm-class V-NAND
The stand guy said that 2D Planar NAND was hitting a scale-down wall in terms of geometry and vertical stacking was the way to go. He talked of Samsung's technology progressing to 96 cell layers and a 1Tb die. We interpret the chart to mean a 48-layer 256Gb or 512Gb die is due in 2015. A 512Gb die must surely mean the V-NAND geometry has to change from its current 30nm-typre geometry to a 2Xnm class geometry though.
General V-NAND roadmap
- There was an AMD stand at the show and it featured the FirePro graphics Processing Unit (GPU) technology: very useful for accelerating a CPU by offloading graphics-style processing to it – less expensive than using more CPUs.
Mac Pro
AMD said Apple’s Mac Pro used this technology inside its shiny black pipe box interior. AMD has produced a W9100 FirePro technology product. delivering 2.62 teraflops of double precision performance and supporting up to six 4K displays. There is also the W8100, with a 2560 Stream Processor inside, which can support four such displays. It has OpenCL support.
AMD FirePro W8100 shootout against NVIDIA Quadro K5000
The big target for AMD is NVidia and it says its tech delivers more power for less cost: according to the company, 23 times more GPU grunt than NVidia’s Quadro K5000.
This story is being written on a 27” iMac which I used to think was gorgeous. After seeing what AMD GPUs can do to an x86 box running 4K displays I’m awash with desire. These things just fly and the colour rendering, resolution and screen real estate make me think the iMac is just plain dull now. Surely this kind of technology will become affordable - just - for mainstream desktop users … please.
Water-cooled hardware seen on IBM HPC stand at Leipzig
One other thing: water-cooling of compute shelves was pretty common at the show, and to see old-style pipe plumbing used to cool extremely high-tech semiconductor chips seemed a weird technological generational mix, like coming across cavemen using iPads. ®