[Beowulf] What's the category of Beowulf among Clusters?

[Robert G. Brown]

On Fri, 19 Dec 2008, Lux, James P wrote:

>
>
>
> On 12/19/08 11:50 AM, "Robert G. Brown" <rgb at phy.duke.edu> wrote:
>
>> On Fri, 19 Dec 2008, Lux, James P wrote:
>>
>>> Beowulf -> the definition evolves.. But.. I'd say it's a bunch of
>>> interconnected commodity computers intended to form a single computational
>>> resource. Some would add "running open source software" Commodity is
>>
>> Some would, and it absolutely is in the original definition and perhaps
>> even worth fighting for, or about.  Not necessarily ONLY open source
>> software, but the operating system should be open source if nothing
>> else.
>
> I agree philosophically, but one has to wonder if open source OS is an
> essential part of beowulf-ness, or just happened to be an enabling aspect

Absolutely.  Read the original statement and definition on the
beowulf.org website and it explains why.  As have I a time or two.  Too
many things upon which performance critically depends are in the OS, and
who wants to rely on Microsoft for bugfixes?  These are the folks that
left a critical exploit in Explorer for six months just last year, so
that every Windows machine on the planet at that time had at least
spyware infecting it at the end of it.

> (partly because Don B. was cranking out drivers for Linux.. Another enabling
> aspect).  That is, beowulfs wouldn't have existed without open source OS,
> but neither would they have existed without cheap commodity PCs, which are
> hardly open source.  And I think the real enabler was the realization that
> you could use the proverbial "pile o' PCs" to do real work.

It was (and is) both, or all three:  Commodity PCs, Commodity network,
and Open Source Operating System.  Microsoft can sell a cluster product
if they want, but they cannot call that product a beowulf cluster.


> Is a group of vector processors a matrix processor (or tensor processor,
> depending on the rank?)?  Or isn't it just a bigger vector processor?

I think it is either a tensor processor of some rank or possibly a
graded division algebra, personally.  Row AND column indices on your
parallelization...

> I think parallelism, in some sense (even if EP) is an essential part of
> beowulfness.  As is "potentially single task-ness"..

I would certainly never argue, but at the very first PVM talk I ever
attended, back in the fall of 1992, the demonstration by Vaidy involved
a pile of (IIRC) a mixture of DECs, Suns, and (no kidding) a Cray.  And
of course I would argue that Geist, Dongarra et. al. are the ones that
REALLY "invented" the HPC commodity parallel cluster.  Lots of people
were doing massively parallel computing on piles of workstations before
"the beowulf", myself among them.

The only real difference was that linux let one use REALLY cheap PCs,
and to be honest the difference in performance between the P5/Pentium
and a Sun workstation was so large that there wasn't much of a
PRICE/performance advantage in a cluster of PCs vs a cluster of the
cheaper Suns. The P6 -- the Pentium Pro -- was another matter.  A dual
PPro gave you a competive number of raw FLOPS compared to (say) an
inexpensive Sun and let you buy a LOT more aggregate FLOPS/dollar.

The point being that parallelizing vector processors has a longer
history that the beowulf (CM5, anyone?) and is certainly "permitted" by
the model, as long as they are commodity vector processors.

> But I still wonder about the prevalence of clusters in Geatland. I'm
> rereading the historical document now, and I find mentions of halls of
> warriors (hmm. Parallelism, commodity, open source, all potentially working
> on a single task.. But that puts our man Beowulf really functioning as the
> head node, and nobody refers to the army as Beowulf.  Finally, when it comes
> to really useful stuff (slaying of Grendel, etc.), B does it himself)  Maybe
> the clustering stuff is in that digression in the middle of the wedding,
> which I always skip over.

There you've got me.  But hey, I named my first efforts "distributed
parallel supercomputers" and look where it got me:-) Not exactly a
household word, eh?  Where everybody (with a bit of nerd in them
everybody, admittedly) knows what a beowulf is, sort of.

    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