Gould UNIX systems

[Bill Mitchell]

Several weeks ago I posted a short note giving some personal observations
about Gould UNIX systems.  Since then, I've had an opportunity to do some
very brief and informal (about two hours worth) testing on a 32/97 and I
thought I'd give a quick report of what I found.

I should say that since there seems to be ample evidence that the Gould
processors run as fast as Gould claims on a wide variety of applications,
my main interest was in finding anomalous data points.

Unfortunately, the dial-in line on the 32/97 I was using was extremely noisy
and I was able to transfer only some very short test programs to the 97.
Because of this, the bulk of my testing was done via standard programs.

The Gould system was a single processor 32/97 with 4M of physical memory
running Gould's "baseline 8.1" which is a 4.1c more or less.  The tests
were also run on a Vax-11/780 with 4M of physical memory running 4.2.  I
was on the only person on the 97; the Vax tests were run with a few other
users on.

In no particular order, here are some results that are representative of
what I found:  (Times are in seconds and are averaged over several runs.)

	nroff /usr/dict/words
		97:   31.6 user/  .8 system
		Vax: 140.2 user/ 5.0 system

	dd if=x of=y ibs=1  (where x is a 50kbyte file)
		97:  1.1 user/ 19.0 system
		Vax: 7.3 user/ 62.2 system
		
	gdata 10000 | cat | cat | cat >/dev/null (gdata 10000 generates 10000
						  50-character lines)
		97:  3.0 user/ 1.6 system
		Vax: 6.1 user/ 1.7 system
		
	fork 1000 (Does 1000 fork and waits of a child that dies first thing.)
		97:   .4 user/ 32.1 system
		Vax: 2.2 user/ 33.5 system

	ex <ldup.in (opens a new file, puts a 50 character line in the buffer
			and then repeatedly duplicates the buffer with 1,$t
			until it contains 32k lines.)
		97:   8.2 user/  7.1 system
		Vax: 27.0 user/ 16.0 system

As a rule, the 97 ran in the 3-4 times a Vax range for processes that
didn't do many system calls.  The presence of system calls seemed to
disproportionately slow down the 97 with the forking process being the
extreme example.


Although my 97 testing is perhaps best characterized as "cursory", the
system really does have a nice feel to it and I'm convinced that the
performance figures that Gould quotes are representative and not
"optimistic".  The only complaint that I have with the 97 and associated
software is the debugger problems I mentioned in an earlier article.  If
those problems could be corrected then I don't think I would have any
reservations whatsoever about recommending the 97.

					Bill Mitchell
					whm.arizona at csnet-relay
					{noao,mcnc,utah-cs}!arizona!whm