eugene@pioneer.arpa (Eugene Miya N.) (08/27/87)
Hugh LaMaster (just down stairs and down the hall) wrote:
>2) Be careful to compare oranges and oranges. [Or apple to apples].
Great... I'm literary trying to write down just what this person-hood
(Mother-hood went out years ago) statement means. Yeah we want
consistency. But, it is not as simple to achieve as people think.
One thing I have found in certain ego benchmarks is that when you get an
"apples-to-apples" comparison, people say, "You can't compare Golden
Delicious to Macs." The problem may ultimately be unsolveable.
What I am working on is along the lines of what I call "equivalent
machines." Such that I can substitute, by linear transformations, one
measure for another. Consider that Dongarra's linpack has what many
consider constrains: 64-bit fixed size (orignally 100x100 dense system).
floating point "intensive." 64-bit is typically derailed by some
marketing types by, "Well, in case you don't need the precision, we
offer 32-bit..." The means then becomes what's a 32-bit derived-number
versus a 64-bit rate. The logic might go: well the ratio is 1:2, so
that should be the ratio between any two 32 to 64-bit comparisons,
right? The problem is things like this don't scale. I am also
thinking about what I might call, a "minimally equivalent machine"
and a "minimally equivalent benchmark," and I'm trying to think of this
stuff as I go along. Give it a try.
Try comparing a 36-bit computer to a 32-bit computer. Set explicitly
aside certain consistency assumptions: like identical cycle time. What,
you've never work on a 36-bit [32-bit] computer. Well go give it a try!
You might get your thinking warped, but seriously. Tell me how you
think you would do some of these apples-to-oranges comparisons [note:
both use the same DNA, therefore, it's just the number of genomes
right?].
From the Rock of Ages Home for Retired Hackers:
--eugene miya
NASA Ames Research Center
eugene@ames-aurora.ARPA
"You trust the `reply' command with all those different mailers out there?"
"Send mail, avoid follow-ups. If enough, I'll summarize."
{hplabs,hao,ihnp4,decwrl,allegra,tektronix,menlo70}!ames!aurora!eugene