Re: Installation

jcduell_at_lbl_dot_gov
Date: Thu Nov 17 2005 - 17:14:57 PST

  • Next message: Eric Frederich: "Re: Installation"
    On Thu, Nov 17, 2005 at 11:13:26AM -0500, Eric Frederich wrote:
    > Hello,
    > 
    > I am new to parallel processing. I am reading through the
    > documentation and have come to understand that when you compile UPC
    > programs then are compiled into an intermediate form, then sent to
    > different nodes where they are compiled into C or C++.  Please tell me
    > if this asumption is correct. I can have a cluster with an AMD
    > Thunderbird, and AMD Athlon XP, and an Intel P4 because they will be
    > sent the source and will be compiled locally during the run?
    > 
    > Also, I downloaded the UPC package and read through the INSTALL file
    > that came with it. I think I will be able to install it but I was
    > wondering a couple of things.  1, do I need to build this on each node
    > in my network? I think the answer is yes, but I wanted to ask anyway.
    > 2, doesn't there need to be a daemon process running on each node to
    > listen and accept requests for computation?
    > 
    > The only thing I saw about multiple hosts was in the man page where it
    > mentioned to use either an evironment variable with spaces between
    > host names or a file with one host per line and use an environment
    > variable to point to that file. I didn't see anything about how to
    > install UPC on these machines.
    > 
    > Could someone please give me an explaination on how UPC is supposed to
    > work with multiple compute nodes.
    
    Eric:
    
    Do you have NFS installed between your nodes?  If not, you will want it.
    Once you have a shared filesytem between the nodes, the question of
    whether you need to install UPC on all of them becomes easy--just
    install it once, and it's on all the nodes.  You also need NFS to easily
    run UPC jobs over UDP.  Every node needs to be able to run the
    executable, and NFS is the easiest way to do this.
    
    Although you have different processor types (AMD vs. Intel), I believe
    all the processors you mention are all x86 compatible, so there's
    nothing stopping you from running the same UPC program on all of them.
    
    A clarification:  Berkeley UPC doesn't push C code to your nodes.  Your
    UPC code will be translated to C, and then compiled to a regular
    executable, on the same machine as where you type 'upcc'.  So things
    work much like 'gcc', or any other compiler.  One difference is that by
    default, you'll wind up using our public translator at
    upc-translator.lbl.gov, but this happens "under the covers", so unless
    you've lost network connectivity, you'll never notice.
    
    You don't need to run any daemon services on your compute nodes for
    Berkeley UPC.  If you run over UDP (which sounds like the best network
    for you), 'ssh' will be used to push jobs out to the nodes.
    
    One last thing you need to make sure of is that you are running a
    similar operating system setup on each node.  In particular, the
    standard C libraries must be compatible.  The safest bet is to run the
    exact same operating system version on each machine.
    
    -- 
    Jason Duell             Future Technologies Group
    <jcduell_at_lbl_dot_gov>       Computational Research Division
    Tel: +1-510-495-2354    Lawrence Berkeley National Laboratory
    

  • Next message: Eric Frederich: "Re: Installation"