Connect with us

Super Computer Help

Discussion in 'General Electronics' started by Eric Uren, Jul 25, 2003.

  1. Eric Uren

    Eric Uren Guest

    To whoever can give me a hand,

    I am a summer intern at a company called AT Systems. I was
    recently assigned the task of building a large parallel computer. I
    have thirty mother boards each containing the following: P266
    processor, 128 MB of RAM, 128 IDE, Compact Flash Drive, and Ethernet
    and USB Ports. Our engineering department is attached to the factory,
    so having a custom built tower, power supply, monitor and such is not
    a problem. Any advice, help, links, etc. would be greatly appreciated
    on the first steps I should take.

    Eric Uren
    AT Systems
     
  2. Nirodac

    Nirodac Guest

    This technology has already been done. It would be easier to buy something
    off the self from IBM, Dell or HP. The software is the key issue here.
    But, given your hardware, I would suggest connecting all the computer boards
    together using the Ethernet links and a controllable hub / switch, or
    network load sharing device. The OS will have to be able to divided up the
    tasks, send that portion to a given computer node, then retrieve the data
    once it's been acted on. Like I said, the software is the key ingredient
    here. There is hardware out there that will take network traffic, and share
    it across multiple computers NOT running parallel software. This is "load
    sharing" not parallel tasking, where each computer operates independently.
    Parallel computing is where multiple computers work on (share)one task. So
    you need to know if it's true parallel tasking you want or load sharing.
    I would suggest using a version of Telnet to access each motherboard, once
    your software is up and running. It'll cut down on screens, keyboards and
    KVM's. Although you'll need one keyboard, mouse and screen, for the initial
    setup of EACH node.

    Although it applies only to UNIX, this site might give you some insight on
    how the rest of the industry does it.

    IBM has a web site that carries "red books", these are IBM documents on how
    there systems operate, hardware and software wise.

    Web Site=http://publib-b.boulder.ibm.com/Redbooks.nsf/Portals/UNIXRedbooks

    Remember this is for UNIX RISC(p series) site, they may have a similar site
    for Intel based (e series) servers. This is intended to give you some
    insight into parallel computing, not, to build a system for you.

    I assume you know enough on how to power these mother boards, current loads,
    etc.

    Good luck.
    Ray
     
  3. Me

    Me Guest

    Build a big blade server. Put rails on each side of all the motherboards
    and make a cage that they will slide into standing on end. Put fans in the
    front and back to flow air through.
     
  4. For starters, look through Scientific American around the Summer or
    Fall of 2001. You're looking for an article (large article, split
    across 2 issues IIRC) on the "Stone Soup-er Computer" that was done at
    Oak Ridge Labs. The article includes enough links & bibliography to
    get you started. Your project sounds quite similar to what they did.

    Also do a web search on "KLAT2". That's "Kentucky Linux Athlon
    Testbed 2", another effort to build a supercomputer in your garage.

    These are all "Cluster" machines, an architecture that started
    becoming popular around 6 - 8 years ago. There's also a website
    called "500 Fastest Cluster Computers in the World" or something like
    that which is exactly what the title suggests. It also includes links
    & biblio.

    The hardware architecture is pretty straightforward: tie everything
    together with the fastest Ethernet cards, switches & cabling you can
    get (fairly cheap) or use a proprietary networking product called
    "Myrinet" (not so cheap).

    Do another websearch on "Beowulf" (or sometimes spelled "Beowolf") and
    discard all the references to midevil legends. What's left is about a
    zillion links to a Linux-based operating system for cluster computers.
    I believe several variants are available under GPL terms. ("Beowulf"
    and "cluster" might be a better choice of seed for your search
    engine.)

    Then comes the hard part . . .

    Unless you have some piece of software specifically created for a
    cluster computing environment, you might as well have 30 individual
    machines. In fact, you may be better off with 30 individual machine
    since they wouldn't have the overheads imposed by the network and the
    cluster OS. Unless a piece of software is aware that its task can be
    split among several processors, it's going to run entirely in one
    processor and no speed-up will be noticed. In my mind, the biggest
    problem with cluster machines (or parallel multiprocessor
    architectures in general) is finding a problem that can take good
    advantage of the multiprocessor environment. It wants to be something
    that requires a lot of independent calculations, and not much
    communication for either data dissemination or collecting results.
    Maybe something involving computations with very large, sparse
    matrices that can be broken down into unlinked operations on smaller
    submatrices.

    Practical comment: with processor clocks now in the 2 GHz range, your
    30 266's have the processing power of only 3 or 4 "modern" processors.
    Multiprocessor motherboards mounting 2 processors are not difficult
    (nor terribly expensive) to get, and they're supported by (some
    versions of) Windoze. And they don't have as much communications
    overhead to deal with. So at the end of the day, your supercomputer
    may not be much better than a desktop box you get from Best Buy or
    CompUSA.

    Don't get me wrong - I'd love to do the project you have. But mostly
    for the bragging rights to say that "I did it!" . . . and with the
    intent that Version 2.0 WILL be based on 33, ATHLON 3GHz
    Multiprocessor motherboards. The sort of machine that might answer
    the question, "Is there a God?" . . .


    ( . . . by replying, "There is now.")
     
  5. Linux boxes can easily be run completely headless (no video card or
    keyboard controller) and use the serial port for console work, and ssh
    or telnet for normal use.

    Parallel computing is very complex. It's not really possible to split
    up the threads of a task and run one on each computer; processes need
    to run on a single machine unless specially designed.

    However, using them as a computing pool is very easy, and has already
    been done. (People request job time, and their job is run on any
    available processor for that time).

    What is this system to be used for?

    Good luck,
    --J.
     
  6. N. Thornton

    N. Thornton Guest


    Hi

    I am most curious as to what computers will be doing in 200 years
    time. I wrote a piece on it, but what do you think they'll do?

    Regards, NT
     
  7. Eric Uren

    Eric Uren Guest

    Well for the software part of your question. We are not really
    concerned with the software issue yet. Because we do not know what the
    college will want to do with the computer, so we are leaving that part
    basically up to them. We are just trying to build the computer and
    make sure it works, then get it off of our hands and use it as a tax
    write off. After much research, emails, internet searches, and
    threads, I think this is what I might do. I think I will use an
    Ethernet connection to link all of the nodes together. I will use a
    Cisco 2950 with 48 ports to be the switch. I think I might use 6x450
    watt power supply's at the base of the tower. Position fans at the
    base to blow upward to cool off the boards. Kind of like a chimney. I
    will make two stacks of 15 boards, and place the adjacent of
    eachother, with the switch in the middle of the two stacks. My tower
    will have the dimensions of roughly 4 feet high, two feet deep, and
    one foot wide. Those are just rough estimates rounded neatly. Any
    responses would be greatly appreciated.

    Eric Uren
    AT Systems
     
  8. I chased down some of the links I sent you after. In Google, "KLAT2"
    returns about 750 responses. There's a decent picture at
    <http://aggregate.org/KLAT2/> and others on the pages linked from this
    page. Another page at
    <http://www.arstechnica.com/cpu/2q00/klat2/klat2-1.html> might give
    you some ideas about physical implementation & packaging.

    " "stone soupercomputer" " returns about 600 links. The Scientific
    American article is at
    <www.sciam.com/2001/0801issue/0801hargrove.html> and looks like it was
    in July & August 2001. Another link for the Stone Soupercomputer
    Project is <stonesoup.esd.ornl.gov/>

    "beowulf cluster" returns over 75,000 links. Start with
    <http://www.beowulf.org/> and use the links on the sidebar.
    <http://www.beowulf-underground.org/> is a good one and
    <http://www.fysik.dtu.dk/CAMP/cluster-howto.html> may be VERY useful
    to you.
     
Ask a Question
Want to reply to this thread or ask your own question?
You'll need to choose a username for the site, which only take a couple of moments (here). After that, you can post your question and our members will help you out.
Electronics Point Logo
Continue to site
Quote of the day

-