Subject: Re: [SWIPL] integrating YAP/GNUP binaries with SWI
Date: Sunday 26 October 2003 10:04
From: Jan Wielemaker <jan@swi.psy.uva.nl>
To: David Reitter <r8@davids-welt.de>, Jan Wielemaker <jan@swi.psy.uva.nl>, SWI-Prolog list <prolog@swi.psy.uva.nl>

Dave,

> thanks for your hints.
> 
> > Did you use the profiler?  If there are clear hotspots you may consider
> > rewriting them in C (after a close inspection of the Prolog code for
> > optimisations :-).
> 
> The profiler was pretty helpful, and I also rewrote some stuff in C, but it
> turned out that the native version of small pieces of code (which get called
> frequently) is only marginally faster. SWI is doing a good job here. A lot
> of my stuff is based on the unification of large feature structures, and I
> am using prolog's unification mainly. Usually, they succeed (where they may
> fail, I have implemented some fail-fast algorithm). Also, there is a bit of
> backtracking involved, and I hope that the compiled version will be faster
> in these areas. (As said before, I can only compile the kernel, as other

I think I've got bad news for you.  An interesting piece of software
is the peirera benchmark suite for Prolog.  It isn't very realistic,
but it contains a 34 tests that all try to analyse only one particular
aspect of Prolog performance.  With all 34 tests normalised to 1
second `SWI-Prolog CPU time (not optimised)' yap is overall a little
over twice as fast.  Studying the tests in detail, it is mostly
calling predicates that is a lot faster in yap (upto 8 times).
Unification of large terms is roughly equal (yap beats SWI by 20%).
SWI is 5 times faster asserting facts.  Most of the test of the tests
are a bit hard hard to interpret as the difference in calling overhead
affects all the tests and some work should be done to correct the
figures.

On the other hand, the requirement to unify large terms where most
unifications succeed hints to possibly poor design of the
datastructures.  Maybe you can try to describe the problem and design
you used.  There are some really smart and experienced people on this
mailinglist!

> parts of the project use libraries such as http, and the database. SWI-pl is
> doing an excellent job at speeding up development!)

Thats where it is designed for :-)

	--- Jan


----------------
* To UNSUBSCRIBE, please use the HTML form at

    http://www.swi-prolog.org/mailinglist.html

or send mail to prolog-request@swi.psy.uva.nl using the Subject: "unsubscribe"
(without the quotes) and *no* message body.

** An ARCHIVE of this list is maintained at

    http://www.swi.psy.uva.nl/projects/SWI-Prolog/mailinglist/archive/