[Beowulf] I/O bound simulation

[Robert G. Brown]

On Fri, 30 Nov 2007, Mark Kozikowski wrote:

>
> Hello all,
>
> I have been reading the list for quite a while and have
> gleaned a boat load of information.
>
> Thank you.
>
> Now I am going to ask help for a problem I am having.
>
> I have simulation program that runs on between 5 and 15
> nodes. This depends on the desired fidelity of the simulation.
>
> When I approach the higher fidelity levels, the simulation starts
> to choke on the quantity of data being processed.
>
> It appears that the system is failing on I/O. Transferring large
> amounts of time critical data between process elements.
>
> I a running on a mostly standard Red Hat distro, no special
> compiling or running architectures are in place.
>
> Do any of you have suggestions as to how I might start
> getting control of this I/O problem?

Sure.  Start by describing your cluster IN DETAIL.  What kind of
systems, how many processors, how much memory and disk, and above all,
what kind of network(s) and what speed.

This will help you figure out if there are any relatively "simple"
solutions, like getting more/better network bandwidth.

The second aspect of fixing your problem is to analyze it mathematically
and logically fairly completely.  Supposing that you've got all the
bandwidth between hosts that you can reasonably afford, the next
question is whether or not you can reorganize your task to take maximal
advantage of that bandwidth.  This may involve significantly altering
your task organization in possibly non-intuitive ways, or it may involve
organizing your computation so data transfers themselves occur in a
maximally parallel way.  Whether or not this helps you increase your
throughput depends in some measure on how your task is organized and if
there is room for improvement.

    rgb

>
> Thanks for any help.
>
> Mark Kozikowski
> Advanced Acoustic Concepts
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
>

-- 
Robert G. Brown
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone(cell): 1-919-280-8443
Web: http://www.phy.duke.edu/~rgb
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977