<---- template headericclude ----->
Setting up a Cluster for Scientific Computing
FedoraForum.org - Fedora Support Forums and Community
Results 1 to 3 of 3
  1. #1
    Join Date
    Jun 2009
    Posts
    32
    Mentioned
    0 Post(s)
    Tagged
    0 Thread(s)

    Setting up a Cluster for Scientific Computing

    I want to set up a cluster for scientific computing (mainly statistical stuff with R). I have a few conceptual questions.

    First, is there a difference between a Beowulf cluster and a cluster that has single-system image ("SSI," e.g. using openSSI or LinuxPMI)? If so, what's the difference?

    Second, if there is a difference between Beowulf and SSI, which one is better for scientific computing?

    Third, does using Eucalyptus make sense for scientific computing or is this more suitable for IO-oriented operations such as web service or databases?

    Fourth, is there some other technology that I should pay attention to that I have not named here?

    Thanks!

  2. #2
    Join Date
    Jun 2009
    Posts
    472
    Mentioned
    0 Post(s)
    Tagged
    0 Thread(s)
    First, LinuxPMI is NOT a SSI, it is the old openMosix standard taken up again and is distinctly different

    Here is a good very simple explanation of Linux PMI

    http://en.wikipedia.org/wiki/LinuxPMI

    I think what you are looking for in terms of comparisons is best summarized here :

    http://en.wikipedia.org/wiki/Single-system_image

    Second: which one is better depends on what your input is , how you code, and what output you want

    Third: By Eucalyptus, do you mean the Eucalyptus Public Cloud (EPC) ? If so , If they can give you an API to run R i dont see why you could not use a cloud service to run stats.

    Fourth : have you considered options other than R?

  3. #3
    Join Date
    Jun 2009
    Posts
    32
    Mentioned
    0 Post(s)
    Tagged
    0 Thread(s)
    Thanks for your reply.

    First:
    Okay, I just read this and it seems that LinuxPMI is multi-system-image software, which looks like SSI, but from the perspective of each node. So it is sort-of-SSI (see also here). Still, I'm not 100% sure what the difference between Beowulf and SSI is (but see also second point below).

    Second:
    The way I understand things is that Beowulf is different from SSI because with Beowul you have a master node and slave nodes (which you do not have with SSI). So if you want to tailor your programs using MPI, then using a Beowulf is better (this is probably called explicit parallelism). If you just want to run some programs without explicit parallel programming, use something like LinuxPMI, but the speed gains might be lower than with Beowulf (i.e. implicit parallelism). Please correct me if I'm wrong (which might very well be the case).

    Third:
    By Eucalyptus I mean this. My point is not if one can use it with R in a cloud (most likely yes), but whether or not it would result in a meaningful speed improvement without having to explicitly code R in a parallel way (i.e. implicit parallelism).

    Fourth:
    I'm in general open to alternatives to R. But since R is the lingua franca in statistics, I want to focus (at least for now) on R. See also this R task view. Is there a particular alternative that you might have in mind, i.e. an alternative that would work in some way better than R in a cluster?

    Thanks.
    Last edited by martih; 2nd March 2010 at 07:52 PM. Reason: added clarification to point three.

Similar Threads

  1. Cluster Computing How to?
    By Anji_fedora in forum Using Fedora
    Replies: 7
    Last Post: 12th August 2009, 05:20 PM
  2. Computing Cluster
    By HENRYLINUX in forum Using Fedora
    Replies: 1
    Last Post: 17th November 2007, 11:18 PM

Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •  
[[template footer(Guest)]]