hiebeler@cs.rpi.edu (Dave Hiebeler) (06/02/90)
There has been some discussion recently on alt.fractals and
comp.arch about the Connection Machine, how fast it is for
various things, how easy it is to use, and so on. I've added
comp.theory.cell-automata (and will send a copy to comp.parallel) as
well, since this topic has more than a little relevance to those
groups too. Please try to restrict follow-up articles to the most
appropriate group(s) only, although these topics do span many
disciplines.
I'd like to point out a good recent article about the CM, and also
mention some comments by Toffoli and Margolus on the generality of
various types of computers.
First, I'd like to mention that there is an article
"Computational Physics on the Connection Machine" by Bruce Boghosian,
in the Jan/Feb 1990 issue of "Computers in Physics". It's a good
overview of the CM, but unlike earlier articles of that type, Bruce
makes some comments about the CM based on his own and other people's
opinions, over the past few years of using the machine.
In addition to the quick intro to the machine, he talks about how
some of the software primitives on the CM correlate to basic
mathematical operations. He describes several example applications
that people have used the CM for, and how to avoid some potential
problems in adapting a simulation to the CM (such as deciding what
type of geometry to use, and how to implement different kinds of
boundary conditions).
He also mentions why it is often *easier* to implement a simulation
on the CM than on traditional machines. Overall, the article is
definitely worth checking out.
****
Someone mentioned in a previous article that the CM isn't
good at being used as a "general purpose processor", that it's only
"good for solving certain kinds of problems, not all computable
problems". This is an important thing to say, but I think it is
very misleading.
Toffoli (either in one of his articles with Margolus, or in a
presentation at the CA conference, I don't remember which) discussed
this issue as it relates to CAMs (Cellular Automata Machines). I
think it applies equally to the CM. Basically, his belief (as I
understand it) is that a CAM is just as general-purpose as a
traditional-architecture computer, but that it is centered on a
different class of applications. That is, you can make a
computation-universal cellular automaton, so in principle you could
lay out a Sun in a cellular automaton array, and simulate it on a CAM
(or the CM). It would be *very* inefficient, however. Similarly, you
can run cellular automata based physics simulations on a Sun, but it
is *also* very inefficient.
The point of it all is that neither architecture is any more
"general purpose" than the other; they are just centered on different
areas, and do well on different things. Since CAMs and CMs are pretty
new, right now we have many applications for traditional machines but
not as many for the new machines. As time goes on, that will change.
It would be foolish to continue to run cellular automata simulations
(and many other computational physics applications) on traditional
architectures; it would simply be a waste of resources. In my
opinion, this is a *very* important topic. See Boghosian's article as
well; he does mention why supercomputers (even vector machines such as
Crays) won't pass the teraflops barrier, but massively parallel
machines such as the CM (and cellular automata based machines, such as
the CAMs) will be able to.
While I'm at it, another good article to check out is:
"Cellular Automata Machines", by Margolus and Toffoli,
in the book Lattice Gas Methods for Partial Differential Equations,
editors Doolen et. al., Addision-Wesley, 1990.
It discusses the architecure of CAM-8, as they envisioned it in
mid-1989 or so. (The final CAM-8 may of course differ from the
article, as changes may be decided upon as they get farther along in
implementing it).
The Lattice Gas PDE book is a good book to have on your shelf
anyway, and the paperback version only costs about $25.
All opinions expressed in this article are my own; my employers may
or may not agree with any of it.
--
Dave Hiebeler / Computer Science Dept. / Amos Eaton Bldg. /
Rensselaer Polytechnic Institute / Troy, NY 12180-3590 USA
Internet (preferred): hiebeler@turing.cs.rpi.edu Bitnet: userF3JL@rpitsmts
"Off we go, into the wilds you ponder..."