Newsgroups: sci.physics.plasma
From <@yonge.cs.toronto.edu:wayne@cs.toronto.edu> Wed Jan 4 20:26:44 1995
From: wayne@cs.toronto.edu (Wayne Hayes)
Subject: Re: speed of cray vs unix work stations

johncobb@uts.cc.utexas.edu (John W. Cobb) writes:
>Of course any answer depends on your particular code, etc. But in my
>experience, the correct number is 10 or less. In terms of $/flop, the
>smaller the machine, the cheaper. Of course their is a limit on size
>below which you cannot run your code, or below which the speed becomes
>excessively slow.

Yup, it certainly depends on the specific code. I'm using a Fujitsu
VPX/240, which is a vector supercomputer, I think similar to a Cray.

On code which is 100% vectorizable (a simple little benchmark I wrote),
it is about 1000 times faster than a Sun SPARCstation IPC, or about 70
times faster than the top-of-the-line SPARC at U of Toronto (I think
it's an SS10/5---something, v8 sparc, 256MB RAM). I know that RS6000's
are alot faster than the original SPARC, but I don't know how they
compare to the more recent SPARCs. Anyway, I'd expect that on highly
vectorizable code, a speedup of 50 is about right. But not all code is
highly vectorizable, though. Code that I recently wrote with 90%
vectorization ran about 10 times faster on the VPX than on said
high-end SPARC. If the vectorization of your code is less than about
90%, it's probably not worth running on a vector supercomputer. Fire
up PVM and run it on a bunch of workstations, if your algorithm has
course-grained parallelism.

--
"There's more to see than can ever ||>> Wayne Hayes wayne@cs.utoronto.ca <<
be seen, More to do than can ever be || Astrophysics, Computer Science, Non-
done." -- from ``The Circle of Life'', || Linear Systems & Chaos & Shadowing,
in Walt Disney's _The Lion King_ || Thinking, Hiking, Biking, and Sex.