At 03:32 PM 9/20/2006, John Leidel wrote:
Uh oh... I see another HPL debate looming in the wings. :-)
Once again, we find ourselves debating how to measure synthetic
performance in order to quantitatively compare extremely disparate
architectures. The HPL test(s) fill a need. However, my personal
belief is that we must simply wait until the machine is built and
decide based on /real/ workloads how it performs. The cell is
certainly an interesting, yet complex, approach to computing. My
only fear is that we, as an industry, are not paying close attention
to our common strategic goal : Perform more science per unit
time. So I pose a question, with the advent of such complex
machines, are we taking a step forward or in reverse?
But, science has been done with things like "attached array
processors", notably things like the Floating Point Systems AP120B
FFT box, some 30 years ago, so complexity, to get to a solution,
isn't inherently steps in the wrong direction. (the original AP120
was an integer array coprocessor, designed to do FFTs for "acoustic
signal processing" with ARPA money. The AP120B was apparently
targeted to folks in oil, seismic processing I assume. One might
assert that this is grubby commerce and not science, though.)
Also, is the goal:
perform more science per unit time
OR
perform more science per dollar
OR
some combination of these plus other factors.
This gets into fascinating discussions about are you better of
spending X dollars on computers that run at Y speed today, and
computing for 3 years, or, invest the X dollars, wait 2 years, and
spend >X dollars on computers that run at >Y*3 speed and compute for 1 years.
Personally, I like R.W.Hamming's statement: "The purpose of computing
is insight, not numbers", so on that basis, since science is
hopefully driven more by insight rather than raw crunching, computing
sooner is better than computing later.
Joe Landman wrote:
Craig Tierney wrote:
see it in the near future. The key to winning the Roadrunner
proposal was to get HPL to run at over 1 PF (science be damned). The Cell
heh ...
Maybe it is time to define a unit of measure.
1 HPL = 1 TF
or something like that ...
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
