Renting a Supercomputer

Discussion in 'Hardware' started by aphexcoil, Nov 14, 2002.

  1. Hmm, look at it this way:

    386 < 486 < pentium < pentium II < etc.

    Think of this as your base hardware logic. Rather than determining what POWER LEVEL you'll achieve, determine an outline of the tasks at hand. The beauty of this cluster concept is that you can assign tasks to each node/computer thereby increasing the speed and abilities of the different aspects of your task. Learn to think, multi-dimensional and multi-layered for your simultaneous multi-tasking environment. :)
     
    #51     Nov 15, 2002
  2. Depends on the effiency of your code.

    That is more important than the hardware with the exception of the network speed/latency ...
     
    #52     Nov 15, 2002

  3. That is interesting. Well, let's say I ask a computer to find a list of all prime numbers. So I come up with a POS algorithm. I want to distribute this task to 10 computers. Since I have no way of splitting up this algoritm into multiple ones, the best I can come up with is to send a group of numbers to a computer and have each computer do a batch and send them back to the main computer.

    The server computer acts as the batch giver and when .. let's say computer 8 has just finished finding out its batch of prime numbers, the server sends computer 8 a new batch to check out. All computers are sending back their batches to the server, which is keeping track of distribution of batches and keeping a database of all calculated information.

    I am an Econometrics major with a minor in mathematics -- I am not a CS major but program on the side, so I don't know the deepest details.
     
    #53     Nov 15, 2002
  4. nusrat

    nusrat

    There isn't one that will give you a practically reliable answer, due to the number of other variables. Trust me: I've been employed doing computing performance analysis and capacity planning for along time.

    However, you can get some idea by looking for benchmark results here:
    www.spec.org
    www.cmg.org
    www.beowulf.org
    NG comp.benchmarks

    But first read this:
    Benchmarks - Fact Fiction or Fantasy
     
    #54     Nov 15, 2002
  5. Yeah.... performance gains against a farm of number
    crunchers is going to be tightly coupled to
    the serial/lack of serial nature of the computation at hand.

    For example... if what your crunching can be done
    within the context of a SINGLE day of ES data, then
    your system will scale almost infinitely.

    1000 computers (all identical), will perform this task
    nearly 1000 times faster than a single computer.

    HOWEVER... if your calculations are dependent on
    ALL previous data (the worst case), your seriously
    screwed and will have to be far more creative in how
    you break up the processing so it can be done in parallel.

    Alphe, you better write a cool ass screen saver like
    the Seti one, and get 10,000 home PC's to crunch the
    data for you :D It would be FREE this way.

    Or you could write a virus, and mail it to 1,000,000 idiots,
    which then downloads batches of data, crunches it, and
    returns it to you. All without being caught :D

    peace

    axeman
     
    #55     Nov 15, 2002
  6. Ninja

    Ninja

    As I understand clusters are great for batch program number-crunching. Are they also great for realtime calculations, e.g. analysing the financial markets or would the latency be a problem here?
    I mean a 'do-it-yourself-at home-cluster', based on Pentium computers, linux , and ethernet.
     
    #56     Nov 15, 2002
  7. Not an issue unless your calculations are so intensive
    they take so long that the entry opportunity has already
    passed.

    I at times have my computer watch about 500 stocks
    in parallel and doing computations on the fly.

    Its a 2.0 ghz P4, and bandwidth, not CPU is my bottle neck.

    Assuming I had the bandwidth, I could easily feed this
    to a cluster and watch every stock, tick by tick.


    peace

    axeman


     
    #57     Nov 15, 2002
  8. Ninja

    Ninja

    Yes thats what I mean. T1 line and tick by tick data for ALL stocks and options and then analyse it on a small cluster (3-5 computers). Could this system be fast enough to give me the results in a timeframe of tens of seconds?
     
    #58     Nov 15, 2002
  9. Well it would seem that if everyone goes from human trading to computer trading, then it really just matters who has:

    a) The fastest bandwidth
    b) Access to the most current quotes
    c) The better algorythm / program
    d) The CPU power to run it


    Let the games begin.

    Axeman,

    Great ideas!!!

    Also, there might be a way to tap into the routers upstream and intercept the data-transmissions and see who is buying and selling before the data got to the exchanges.

     
    #59     Nov 15, 2002
  10. cvds16

    cvds16

    #60     Nov 15, 2002