Re: SIMD/Vector benchmarks

In article <Eb6dnYgVtuHbSbfZnZ2dnUVZ_uudnZ2d@xxxxxxxxxxx>,
David DiNucci <dave@xxxxxxxxxx> wrote:
"How I Learned to Stop Worrying and Love Asynchrony". (It can be found,

I seem to recall there was also a similar subtitle to some movie
somewhere, too. :-)

Actually, I can think of a few CS papers with similar subtitles.
That along with XYZ Considered Harmful, The Next 700 XYZ, The Art of XYZ,
The Science of XYZ, Games XYZers Play, if you catch my drift.

If you guys are driving to that meeting in Greece,
maybe I can hitchhike along again.

While the founding Chair of the SC'xy conferences now has an electric
wheel chair, he was not invited to Salishan this year, and I don't plan
to attend Asilomar (reading Terje? I have to give a talk about
Antaractic science that evening of the RAT session). So Greece is
likely out. I'm surprised you didn't mention Germany (Ken Miura was
just visiting and he's going to that; I just got bac from Germany
and I am awaiting the Austrian postal system to send my trinkets back).

This is why I used developer than user. When you aren't certain
yourself who goes into a code in a physical simulation, asynchrony is
a time variable that they don't want to fool with.

Nevertheless, I stand by my translation/substitution. If the developer
doesn't like speed and the user does, then it sounds like it's time to
the two can be the same
get new developers--or at least find ways to make the old ones happy
while making the user happy. I suggest that giving developers better
models and development tools for asynchrony is easier than convincing
the users that they don't really like speed.

Get better scientists. More Nobel laureates.
Actually I am hoping that the lack of speed makes them develop better
intellectual tools than computers. We are discussing diagrams (schematics
for instance) vs. text based equations on another mailing list.

tolerate message passing.

There are many things to like about message passing, but passing
messages isn't among them.

Too much copying.

%A Marc Mezard
%T Passing Messages Between Disciplines
%J Science
%V 301
%N 5640
%D 19 September 2003
%P 1685-1686
%K physics/computer science (AI) perspectives,
information theory error correction, belief propagation (BP),
discrete optimization satisfiability, statistical physics spin glasses,
%X Nice short paper on the meeting of 2 disciplines.
It glosses over certain global/long distance topics, but quite nice.


Most of those good things (from latency hiding to clean interfaces) can
be found in interfaces I've propounded over the years (e.g. CDS,
ScalPL), minus some of the bad things (like gratutitous copying in
shared memory environments).

The problem Dave as noted by Perlis and
55. A LISP programmer knows the value of everything, but the cost of nothing.
generalized to most computer people is that implementation costs arent
trivial (scaling issues).


While this reference may appear "old" to some, the algorithms are older
and not much improved, and still fairly relevant (no Moore's law for
algorithms).

"No Moore's law for algorithms" sounds good, but I wonder if it's really true.

Let me know if you can come up with basic factoring algorithms faster
than Euclid.

How has the number of operations in an algorithm increased over time?

As a start:
Take any basic algorithm, 1-d say, Newton, on small machines you ran 1-D.
When you use regular data structures like arrays (we can get to
semistructured data structures later), you start to add integer adds and
multiplies w/o increasing the basic FP calculation. Those don't count.
FP not in direct solution: doesn't count. That's a skewed metric.
I have to agree with chums at the NSA that time to solution is more important.

If an algorithm can be considered a proof, what about the one
used in the 4-color map theorem compared to previous proofs (say,
diagonalization)?

Well actually, I need to get that code from the UIUC guys in the
historic software issue.

That all begs the "what's an algorithm?" question.

Get down here, and I will see if Don will have dinner with you.
I know one when I see one.

In any case, as long as Moore's law applies to platforms, it's worth
reconsidering the algorithms that run on them now and then.

Hardware only. Still in many cases algorithms of the 1890s (not a
transposition). Look Moore's law is based on scribing fine lines
(in the submicron level) and that's Newton's inverse squared law.
It's not double every 18 months, it's quadruple every 3 years.
Double is only used by managers and journalists.

--
.

Relevant Pages

  • Re: "Algorithms" in Molecular Biology?
    ... same thing as asking "does the Moon calculate its orbit?" ... biological science, lies with understanding nature's ALGORITHMS. ... people could be compelled to follow social laws. ...
    (sci.bio.evolution)
  • Re: Buying Books
    ... The basic distinction is between science and religion. ... Algorithms are ... Empire of the East is all about demons and gods and such, ...
    (rec.arts.sf.written)
  • Re: Buying Books
    ... The basic distinction is between science and religion. ... Algorithms are ... I suppose Amber is the best case for fantasy, but even there, ...
    (rec.arts.sf.written)
  • Re: Computer algorithm resources
    ... like me who is reasonably familiar with math but not computer ... I expect that there are also computer science discussion groups to ... For algorithms and data structures I ...
    (sci.math)
  • Re: "Algorithms" in Molecular Biology?
    ... biological science, lies with understanding nature's ALGORITHMS. ... but he is more than willing to censor my posts when I ... people could be compelled to follow social laws. ...
    (sci.bio.evolution)