On Sat, 3 Apr 1999, Casioqv wrote:
> It is, I find complete systems in the newspaper for that much.
>
> On Sat, 03 Apr 1999, [EMAIL PROTECTED] wrote:
> >On Thu, 1 Apr 1999, Casioqv wrote:
> >
> >> I find that 486 100's are currently the most cost effective. At about $50/box
> >> that is $500/ghz + networking.
> >
Gee, I thought this was an April Fool's joke but it keeps going;-).
Seriously now...
a) Used 486 systems may be of use to a very few persons, mostly for
"having fun and learning", but they aren't generally available and
aren't generally available for $50 where they are available, and aren't
a very good choice for a serious numerical computation.
b) Raw clock scaling is not a good predictor of performance across
processor generations. The 486 (P4) is two to three generations behind
current P6's (Celerons, PII's and PIII's). A 400 MHz Celeron system
probably has as much raw speed (especially in floating point) as all ten
486's put together (that's just 2.5x clock speed scaling over 2.5
processor generations) and certainly costs less. A "GHz" of CPU
performance is as meaningless as a "bogogip" (the same thing, actually).
What matters is how long it takes one's (often floating point intensive)
code to finish.
c) In addition to much better floating point performance, P6-family
processors have better caches (speed and size) and sit on a faster
memory bus accessing a better memory design (SDRAM). I would bet that a
486, even at 100 MHz, is not fast enough to push a 100BT NIC at
capacity, for example. There was evidence of this in the netperf
database, if I recall correctly.
d) The networking cost is non-trivial. Let's be generous and assume
$20/NIC plus $50/switched port. That's another $700. Heck, make it
just $500. That makes the true ten-node "system" cost $1000 or more.
For $1000, one can build TWO 400 MHz Celeron systems with 64 MB of SDRAM
each, your choice of a 3 GB HD each or an 8 GB in one to serve the
other, and a small stack of 100BT NIC's. No need for a switch with only
two nodes. I'd bet that on nearly any task, either one of them would
tie or beat all ten 486's working on embarrassingly parallel code.
Of course it's downhill from embarrassingly parallel code that fits in
the presumed tiny memories of the $50 486's for the ten boxes...and one
probably has an extra $200-500 to spend for a proper comparison since,
as Uwe also noted, there are a lot more costs to consider for a ten node
system than for just one, two, or three. Our estimated ten-system
network cost is heavily lowball at $500 to begin with. Then one needs
five times as much electricity and five times as much cooling and five
times as much shelf or floor space and five times as much miscellaneous
"stuff". Ten 7' RJ45 cables add up to another $70 or so (over the
counter price from a CHEAP place -- more like $150 from Best Buy), for
example, and then there are a couple of $20 power strips required and it
takes more human time to set up and run all those systems and (being
old) a ten node 486 system is going to be MUCH less stable than a one or
two or three node systems made with brand new 400 MHz Celerons, and
then...
I think that if one looked at REAL cost, three 400 MHz diskless Celeron
nodes could be built and directly interconnected for what it costs to
buy 12 486's and network them together, even at $50/each for the 486
systems. The new Celeron SMP systems drop the comparative marginal cost
per CPU-Hz even more. I cannot imagine that a three node 400 MHz
Celeron beowulf wouldn't outperform the 12 node 486 beowulf by a factor
of two or more and would be MUCH easier to program (it would yield most
of the speed advantage for ordinary von Neumann code).
Given the choice, I know which one >>I'd<< take...
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:[EMAIL PROTECTED]
