[KLUG Members] Mosix

[members@kalamazoolinux.org]

Bruce Smith (bruce@armintl.com) wrote:
>Awhile back when some KLUG members were setting up a web server cluster
>with TurboLinux, some people said it would be fun to create a Beowulf
>cluster.  The problem with that is, even if we did create a Beowulf
>cluster, we wouldn't have any software to run on it!
A (more or less) true statment; remember that these sorts of tools are
developed for folks who feel that rolling up sleeves and diving in (and
usually DEEP) to coding to solve their problems comes with the job. The
class of problems people customarily tackle with supercomputers are very
much **NOT** in the category of "off the shelf" software. In reality, no
one in this end of things cares a wit that there's no software, part of 
their mission is to create the stuff.

Software packages that do exist are fairly tightly focused on particular
problems, like stellar structural models and fluid dynamics over laminar
wings. Attempting to genrealize this kind of code is a non-starter; no 
one's waiting for the next release, they're too busy writing it for their
own needs. This population needs good optimizing compilers and a few other
raw materials (coffee, raw meat, air, maybe a little music, etc.) to grind
out what they need.

>"While Beowulf clusters are extremely powerful, they aren't for
>everyone.  The primary drawback to these types of clusters is they 
>require specially-designed software (written to hook into PVM or MPI) 
>in order to take advantage of the cluster. ..."
That's right (or other tools, like C-Linda). If you want this kinda
power, you're buying into programming a lot of stuff yourself. Generally
the notion of "applications" fall apart at this level; a lot of activity
that's subsidiary to really solving the HARD problem is done with more 
convential computers. For example, the results of some really heavy-duty
calculations are written off to files, and then graphics, reports, and
other (simpler) analysis on the results are done with software that is
retty much off-the-shelf.

>IBM takes a different approach to CPU type clusters which doesn't have
>this limitation, and makes a group of computers look like a big SMP box.
This would dramatically simplify the development of tightly coupled 
clustering sofware. Supercomputer and parallel proccesing folks are 
normally quite skeptical of such claims (having been burned on such
claims by poor- or non-performance on a number of occasions)

>Interesting reading for anyone interested in clustering.
I'll get 'round to it shortly and pass it on, Thanks!

							Regards,
							---> RGB <---