MR>Have you personally evaluated Pentium Pro machines?  I'm open to being
MR>wrong.
Uh oh, I know better than this. Benchmarking is one of those religious
issues that I keep promising myself to stay away from.  Anyway yes, my
OS/2 server and my NT server are dual PPRO's and I have clients that
have directly compared the 233MHz P-II against 200MHz PPRO's in a
variety of applications.  Your hidden meaning is correct and well taken,
in that, some of what I have said is second hand or magazine speak.  I
will be looking at replacing my development system, a lowly PPRO 150, in
the next several months with a dual P-II, for the same reasons as Peter;
compile times are just becoming too costly.  I am thinking of running
another copy of Warp server barefoot for the OS/2 boots.
The real issue here is benchmarks have lost any kind of real credibility
in my mind and the minds of others that try to get real work done with
these little boxes. Many are starting to notice that unless they make
a drastic jump, the megahertz on like architecture is starting to bunch
up. The thrill of super zoom is gone. That probably explains the
overclocking phenomenon. The kick of a new system doesn't quite have the
same effect that it used to, if, you have been making incremental
upgrades.
If you locate the March 98 issue of Dr. Dobb's Journal there is a
fairly lengthy article on how a Pentium 166MHz system can be made to
outperform a P-II 300MHz system using the standard benchmarks around.
The obvious indictment being that the only real benchmark worth a darn
is the use of your day to day work.  That work could be CPU intensive,
memory intensive, I/O intensive or any crosspoint variation of the
three.  Therefore any system can be made to look better and actually is
better than things like Specmark would indicate.
A perfect example is that in no case was the P-II 300 more than 1.65
times faster than the Pentium 166 in the initial Dr. Dobbs baseline
comparisons.  However in your own investigations, if I understood you
correctly, you find the P-II twice as fast matching a Pentium 200
against a P-II 266.  Your personal tests were less of a frequency
mis-match than the Dr. Dobbs tests pitting a Pentium 166 against a P-II
300 indicating that your real work needs are specialized or different in
some way.
As far as the SDRAM goes, I have personally watched tests where EDO or
FPM memory was removed from a system and replaced with SDRAM and
virtually no measurable improvement was seen in the real job mix.
Obviously it was not a linear memory access intensive job mix and
caching came more into play than anything else.  One must remember that
SDRAM access for the first word of an unopened page is basically the
same as it is for FPM/EDO memory.  Therefore if you are constantly
invalidating cache lines against closed pages, SDRAM's increased
performance has no benefit.  The more jobs that are really multi-tasked,
the less the perceived improvement of SDRAM will be noticed, because of
the potential less than linear use of memory.
People always benchmark and then try to compare it to a system that is
close but not exactly the same in some way.  The results in the Dr.
Dobb's article should speak volumes about why this should not be done.
They apparently felt that the article is controversial enough that it
was on their web site.  I don't know if it still is because they
generally just show the article titles and allow source code
downloading, expecting you to buy the hardcopy to read the text.
--Lynn
 * SLMR 2.1a * It looks like an optical illusion, but it isn't.
--- DB 1.39/004485
---------------