Beowulf type Clusters for Engineering apps.

Jared Hodge Jared_Hodge at iat.utexas.edu
Mon Oct 16 11:28:51 PDT 2000


    I don't really know how many others are using Beowulfs for Finite Element
Analysis, but that's what we're doing.  There currently isn't any software that
I know of that is available to do FEA in parallel on a Beowulf, but we are
working on developing our own.  We've been able to run a few sample problems
using our cluster (8 nodes, 350MHz, 100 MHz FSB, 256MB RAM, and 32 bit Myrinet),
and everything looks like it should scale once we finish some fine tuning.  I
believe the plan is to release the software (possibly for free) once we have
completed it, but I'm not making any promises since I don't know if it will be a
general release or just to DoD contractors and such.  Right now the University
of Texas has the copyright on the software (called EMAP3D).
    If your thinking about developing your own code or adapting someone else's,
good luck and let me know how it goes.  If you do, we've been using the PETSc
Library for parrallelization in one of our approaches, and you may want to check
it out.  I will tell you that we had a major bottleneck with 100 Mbps Ethernet
and had to switch to Myrinet.  We may be able to minimize the communication and
get things to where Ethernet would work, but my guess is that with larger
problems comes a huge need for network bandwidth.
    I am confident that the a beowulf will work for FEA, since we've spent quite
a bit of time investigating this (if not, do you know anyone looking for a soon
to be unemployed grad student...).  Beowulfs are really very stable (especially
compared to Microsoft), it's just that when you get a large complex system with
a lot of interdependent parts, there are a lot of things that can go wrong.  The
same is true of any complex system, just look at all the failed Mars landers (no
offense to any of our NASA friends).  Anyway, we're hoping to get the FEA
calculations going on larger problems sometime soon.  We've run it on the Cray
T3E, and I think that we'll get better performance per node on the Beowulf (once
we upgrade to 16 nodes, 800 Mhz, 133MHz FSB, 512MB RAM, and 64 bit Myrinet) on
larger problems than we do on the Cray, because we'll have 4x the memory per
node as on the Cray.
    If you have any questions, let me know.  I'm pretty new to FEA (I have more
of a CS background), but I'd be glad to ask around to people who would have
answers.

Jared Hodge


Prithvish Chakravarti wrote:

> Hi !
>
> I have some questions for the helpful guys on this list.
>
> 1. How many Beowulf clusters are known to be running Finite Element Analysis
> engineering apps successfully ?
>
> 2. Are there any ready to run Linux based FEA apps for Beowulf clusters ? If
> not, how do I go about getting one done ? I mean, could an existing app be
> modified to run on parallel processors ? Or do I need to write one from
> scratch ? I would be grateful for a few pointers or links....
>
> 3. I am new to this list and to Beowulf. I have been trying to follow the
> conversation. I keep hearing that the clusters run the process for a limited
> up time. Why is this ? Is this a prob with the hardware ?
>
> 4. Has anyone attempted to use Beowulf clusters for Industrial mission
> critical apps ? If so, on what platform ? I am looking at Beowulf clusters
> for this purpose and need to know whether it is a feasible option.
>
> Prithvish Chakravarti
>
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf





More information about the Beowulf mailing list