Some newbie questions

[Robert G. Brown]

On Mon, 11 Jun 2001, Roger L. Smith wrote:

> On Sat, 9 Jun 2001, Gerry Creager N5JXS wrote:
>
> > > > > -Do i HAVE to use Linux
> > > >
> > > > Yes.
> > >
> > > I completely disagree.  Many people nowadays are using the term "beowulf"
> > > to mean "cluster", and you can cluster pretty much anything with a network
> > > card and an implementation of MPI (depending on what you're doing).
> >
> > Er.  Ah.  We've buirned this barn before.  We've decided that for the
> > purposes of this list, "beowulfery" refers to using Linux on (generally)
> > commodity computing platforms bound as a cluster.
>
> My primary point here is that the clustering community shouldn't get so
> caught up in dogma that they mislead self-proclaimed newbies as to the
> options that are available to them.  I realize that this list is named
> "beowulf", but it's in the best interest of those who are trying to
> further beowulf clustering to further all clustering, and to keep in mind
> that not everyone (especially newbies) particularly care about
> philosophical holy wars.

As has also been pounded around before, it isn't just a philosophical
holy war; there are some practical and technical issues associated with
the casual use of the terms.  As you pointed out, clustering was around
and many of us were using e.g. PVM or other parallel program support
mechanisms long before the beowulf project was begun.  However, the
original beowulf was specifically intended to be a COTS supercomputer
and was architected for that purpose with various custom elements (like
channel bonding) that were an original primary focus of this list.  I
think I know this one by heart...:-)

The fair way to present this to a newbie is to simply lay it out
non-judgementally rather than for EITHER side to get religious.  I'd
therefore answer all the questions slightly differently starting with:

  The first thing for any newbie to do is to consult the beowulf FAQ,
  the beowulf HOWTO, and to visit the primary beowulf site
  (www.beowulf.org) and look through the list archives where, armed with
  a search engine like google, they can almost certainly find detailed
  answers to precisely these questions (possibly annotated with
  considerable discussion:-).

I'd then say that beowulfs/clusters don't "have" to use linux, but that
it is a good idea to use linux or an open source OS like freebsd rather
than a closed source/proprietary OS, and that linux is the traditional
OS of true beowulfs and where most of the beowulf development energy has
been focused.

Next, I'd point out that one could in principle build a mixed
architecture "true beowulf" but it would be a lot of work, segue into a
short discussion about granularity and barriers and synchronization, and
point out that if the task(s) to be performed were coarse grained or
embarrassingly parallel that there are numerous other cluster
architectures (e.g. NOW/COW/POP/heterogeneous/kitchen sink) that are
functionally similar to a "true beowulf" that should be considered.

That's fair to all sides (I think, anyway;-) and maximally usefully
informative.  Of course I'd be >>more<< inclined to just refer the
newbie to e.g.

   http://www.phy.duke.edu/brahma/beowulf_online_book/node2.html

since I said it all there in a lot more detail and with a simple path
for them to learn more...;-)

   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