philosophical question - limits of beowulf defn

David S. Greenberg dsg at super.org
Mon Jul 17 12:36:10 PDT 2000


I'd like to add a pragmatic spin to the question of when is a beowulf no longer
a beowulf.  My concern is making sure that a natural community continues for
clustered machines.
A few years back many of us at the DOE labs become worried that a too narrow
definition of Beowulf meant that the sorts of machines we wanted/needed to
build (see Cplant at Sandia, Little Blue Penguin at LANL, RoadRunner at UNM,
Chiba City at ANL) were excluded and issues we care about were ignored.  Thus
Pete Beckman and I started the Extreme-Linux forum with its own webpage,
mailing list, workshops, and conference tracks.   This has indeed helped us
focus on issues of large system scalability and highest performance.
However, had we left the beowulf fold entirely we would now be missing out on
the move toward much better beowulf system admin (witness the bproc based work
at Skyld, the turnkey type systems from VALinux and HPTi, etc.).  We would also
be less able to contribute back pieces which are suitable for any
installation.  Bill's ability to project the future of beowulf's would also
suffer.
I hope we can continue to welcome anyone who has questions and ideas related to
ANY of Walt's four points in the general context of clustering.  When
particular threads become long and detailed about a particular facet, eg. the
details of Monte Carlo, the advantages of a particular interconnect, or the
organization of some Linux distribution's /etc directory then it can break off
into private email.  With a little luck we can even get a summary back later.
This way we all hear what everyone is up to.
David

Ward William E PHDN wrote:

> -----Original Message-----
> From: Walter B. Ligon III [mailto:walt at parl.ces.clemson.edu]
> Sent: Monday, July 17, 2000 11:10 AM
> Subject: Re: philosipical question
>
> > A Beowulf system has the following characteristics:
>
> > 1) designed for parallel (high-performance) computing
>
> > 2) dedicated nodes - private network
>
> > 3) COTS hardware
>
> > 4) open-source system software
>
> > It seems the question is this: is Myrinet or SCI considered COTS?  Depends
> > on what COTS mean.  I've heard it as "Commercial Off-The-Shelf" and as
> > "Commodity-Off-The-Shelf".  These products would fit the first definition
> > and not the second.
>
> > The original spirit of Beowulf was based on the second definition.  The
> idea
> > was to get the best price/performance possible from mass-produced parts.
> > At the same time I don't think anyone intended to exclude innovation in
> > new hardware, especially in the network.  I don't think most people today
> > would exclude Myrinet or SCI as a Beowulf - but as has already been
> pointed
> > out, this treads on the grounds of philosophical argument - I'd rather not
> > get into THAT fray!
>
> > Walt
>
> Walt, while I agree with what you are saying in practice, I have to
> disagree...
> COTS definitely needs to be defined as COMMODITY Off the Shelf, and not
> Commercial.
> I can buy  (realistically quickly) an Origin from SGI "Commercially Off the
> Shelf".
> No one here is going to argue that an Origin != Beowulf.  One may argue "But
> an
> Origin is also a 'Complete Turnkey' system... not a Beowulf, that is built
> up
> of parts into a Cluster machine."  You would be right, in part... but one
> would be wrong.  VA-Linux sells 'Complete Turnkey' systems... plug in your
> hardware and go, and you have a Beowulf... or would one argue then that
> "Part
> of the definition of a Beowulf is that it would also have to be a CLUSTER
> machine"?  ASCI has done some interesting things with Origin, SP, and
> Starfire
> clusters...  The last argument would be "Ah, but it doesn't use Open Source
> System software!", which of course is poppycock as well, since there are
> MIPS
> ports, for example, of Linux such that you COULD run a Beowulf on a MIPS
> architectured machine... and MPI and PVM are also available for IRIX, as
> well, which is another part of the system software.
>
> The best definition I've heard in the end is the based on the  "Duck"
> definition... "If it walks like a duck, and quacks like a duck, and acts
> like
> a duck, and looks like a duck, it's a duck." i.e., "If it looks like a
> Beowulf,
> and works like a Beowulf, and performs like a Beowulf at Beowulf costs, it's
> a Beowulf."
>
> In the purest terms, SCI and Myrinet machines are NOT Beowulfs.... but in a
> more
> "real" term, they are simply bleeding edge High Performance markers for
> where
> Beowulfs are  headed, and every bit as "Wulfish" as any other FE based
> design.
>
> Bill Ward
> My opinions, and I'm sticking to them ;)
>
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf





More information about the Beowulf mailing list