beowulf in space

Robert G. Brown rgb at phy.duke.edu
Wed Apr 16 12:34:10 EDT 2003


On Wed, 16 Apr 2003, Eray Ozkural wrote:

> On Tuesday 15 April 2003 19:36, Robert G. Brown wrote:
> > SO although I'm certain that they use clusters on spacecraft in at least
> > one sense of the word, I'm equally certain that they are NOT beowulfs,
> > according to the standard definition.
...
> Thus, I think it's more of a software problem. Can you really build "the" high 
> performance platform that successfully and completely abstracts the 
> OS/network/CPU? What would be needed for such a thing? (I'm not thinking 
> 'Java', that's slow, thank you) Such software is surely in line with Beowulf 
> thinking.

Absolutely, although I'd refer to it more precisely as "cluster
computing" thinking and not beowulfs per se.  To be picky, a beowulf is
"single machine" supercomputer built out of COTS (commodity, off the
shelf) components, running an open source operating system,
traditionally linux, and possibly some software such as Scyld or bproc
that flattens PID space or is otherwise designed to promote that image
of "a beowulf" as being a single machine.

All beowulfs are clusters, not all clusters are beowulfs, see
interminable discussions in years past in the archives and Kragen's FAQ.

A spacecraft cluster will simply never be built with real COTS parts --
they are too unreliable and not nearly expensive enough:-).  They might
conceivably be built with "customized" parts that have a COTS origin --
a system homologous to or derived from a COTS design but subjected to a
far more rigorous manufacture and testing regimen.  It might also be
built with one of the beowulf networks since COTS (in the usual sense of
the term) or not in some cases they are the only game in town.

So I wouldn't be incredibly surprised to see a spacecraft containing a
bunch of "intel" or "amd" nodes, interconnected with e.g. SCI (because
it is switchless and hence arguably more robust).  Those nodes, however,
will be built on motherboards and CPUs custom engineered for low power,
radiation hardness, fault tolerance, redundancy, and tested ad nauseam
before ever leaving the earth.  It is cheaper to spend $100K or even
more on each on those nodes ("identical" in function to a $2000
board+network interface here on earth) and be almost certain that they
won't fail than it is to deal with the roughly 10% failure rate per year
observed for at least one component in a lot of COTS systems.

   rgb

Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list