Re:
> "Increase dramatically" may be too dramatic a phrase. About the best you will
> experience after OCTCOMPing a program will be ca. 30% increase in
> performance, which means that you'll have to be looking for it if you want to
> measure it. A 30% increase or decrease in performance is well within the
Well, Wirt and Gavin's numbers appear to be off SIGNIFICANTLY.
The OCTCOMP is *VERY* good for CM code ... and has been for many,
many years.
Note that Wirt said "the best you will experience", so I'll
first look programs that are 100% CPU bound, which is decidedly not typical.
For example, 1000 loops of doing the simple Sieve of Eratosthenes
in CM SPL on a 3000/968 takes 102.624 CPU seconds.
The OCT'd version takes 12.993 second.
That's 77% less time, or perhaps a 689% increase in performance
(old time / new time * 100 - 100)
This is not unusual, and I've posted similar numbers over the last 15 years :)
So....
what about a program that *isn't* CPU bound?
How about EDITOR.PUB.SYS?
I took CATALOG.PUB.SYS, appended a copy to itself, and saved that
as my test file, FOO.
I then did:
run EDITOR (CM)
TEXT FOO
(and ignore the timings)
TEXT FOO
CPU: 4500 millisecs; elapsed = 5 seconds
and
run OCT'd copy of EDITOR
TEXT FOO
(and ignore the timings)
TEXT FOO
CPU: 1947 millisecs; elapsed = 2 seconds
In short, a major performance boost, resulting in a
savings of 56%, or a performance improvement of 131%
(that's 4500/1947 * 100 - 100)
(I used the CPU times here, because of the lack of precision
in the elapsed times)
Now...as to why the original poster's performance didn't
increase...
One can certainly pose various theories, including accidents
in testing (e.g., perhaps the software had already been
OCT'd) ... to things like "the application spends most of
its time in an NM XL". Only hands on testing/analysis/investigation
would reveal more information, I suspect.
Stan Sieler [log in to unmask]
www.allegro.com/sieler/wanted/index.html www.allegro.com/sieler
* To join/leave the list, search archives, change list settings, *
* etc., please visit http://raven.utc.edu/archives/hp3000-l.html *
|