Re: [Gcl-devel] address@hidden: Re: Windows versus Linux performance results]

[Camm Maguire]

Greetings, and thanks for your report!

Well, my first hunch is the difference in GCL versions -- significant
work has gone in since 2.5.3. If you'd like to try Mike Thomas' latest
mingw build of 2.6.1, you can get it (temporarily) at

http://www.cs.utexas.edu/users/boyer/gcl

As you may know, the sequence of compromises on ftp.gnu.org, savannah,
and debian.org have made it necessary to pause a bit in GCL
development.  For example, I have patches ready to support gprof
profiling, which cannot yet be committed as these services are not yet
reenabled. They are quite small, and should provide useful feedback if
the newer gcl version does not remove the performance discrepancies.
If you wish to try them, instructions can be found at

http://mail.gnu.org/archive/html/gcl-devel/2003-12/msg00069.html

We are also in the process of reviewing or default memory layout.  In
fact, if I recall, we may have made the switch from a 32M default
maxpages setting to a 128M between 2.5.3 and 2.6.1.  These will likely
make a great difference in performance.  You can find good sample
settings for gcl/maxima at

http://mail.gnu.org/archive/html/gcl-devel/2003-12/msg00077.html

I'd like to call particular attention to the setting

(si::set-gmp-allocate-relocatable t)

-- I'd be interested to see how this affects your results.

In any case, we are in the process of reviewing all of this so that
more performance can be had out of the box.  We are temporarily in a
suspended state pending restoration of cvs and autobuilder services. 

Take care,

"Matt Kaufmann" <address@hidden> writes:

> Camm --
> 
> An ACL2 user, Jared Davis, has reported running the same tests on the same
> machine, both using GCL, that take 55 minutes on Linux but 90 minutes on
> Windows.  However, the GCL versions are different.  I just thought I'd pass
> this info along to you In case this surprises and interests you.  Details are
> below (you'll notice that Jared took some care in the direction of making this
> a fair comparison).
> 
> -- Matt
> From: "Jared C. Davis" <address@hidden>
> Subject: Re: Windows versus Linux performance results
> To: "Matt Kaufmann" <address@hidden>
> Date: Sat, 13 Dec 2003 12:54:06 -0600
> 
> Yes, both tests are using GCL.  The Windows test was with GCL 2.5.3, and 
> the Linux test was with GCL 2.6.1 (which I don't even see on the GCL web 
> site.  I guess it is from CVS, it says November 6...  It was from 
> Debian's apt-get unstable branch.)
> 
> Feel free to forward it to whomever, I can rerun the test with the 
> older/newer GCL if that will make a difference.
> 
>       Jared
> 
> Matt Kaufmann wrote:
> > Thanks, Jared.  I'm assuming that you used GCL in both cases; is that 
> > correct?
> > Which version?  Do you mind if I pass along your message to Camm Maguire, 
> > the
> > primary GCL maintainer?  He might be interested and he might be able to 
> > offer
> > some insights.
> > 
> > Thanks --
> > -- Matt
> >    Date: Sat, 13 Dec 2003 00:46:32 -0600
> >    From: "Jared C. Davis" <address@hidden>
> >    User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.5)
> >     Gecko/20031105 Thunderbird/0.3
> >    X-Accept-Language: en-us, en
> >    X-WSS-ID: 13C46A47697516-01-01
> >    Content-Type: text/plain;
> >     charset=us-ascii;
> >     format=flowed
> > 
> >    Hrmn, seems that the regression suite only took 55 minutes on Linux but 
> >    took 90 minutes on Windows, both on this same machine.  I am not sure 
> >    what would be causing the disparity, and honestly don't have much sense 
> >    of where to begin in tracking down the problem.
> > 
> >    Some thoughts...
> > 
> >      1. I ran the Linux test after setting up hdparm to use about the least
> >     efficient drivers it could, transferring only 3MB/sec from the hard
> >     disk.  For this reason I do not believe that poor Windows drivers
> >     are at fault.
> > 
> >      2. The GCL I am running for Linux was obtained through apt-get, whereas
> >     the GCL I am running for Windows was compiled myself.  I would
> >     expect the self-compiled version to have taken advantage of more
> >     optimizations.
> > 
> >      3. Both ACL2 images were built from the ACL2 sources.  I don't see how
> >     there should be any difference here.
> > 
> >      4. Perhaps the unix emulation or whatever allows these unix programs
> >     to run on Windows is causing the slow down.  I don't have any stats
> >     as far as the performance of "other" programs compares.
> > 
> >    Anyway, thought I'd let you know.  Of course, the performance is still 
> >    within a factor of 2, which honestly might be good enough.  Maybe there 
> >    are "native" lisps that a Windows user might get better performance from?
> > 
> >        Jared
> > 
> > 
> >    Matt Kaufmann wrote:
> >    > Thanks, Jared.  If you want to do the regression, great!  The basic 
> > one is done
> >    > by standing in the acl2-sources directory and invoking:
> >    > 
> >    > make certify-books-fresh
> >    > 
> >    > If you want to fetch all the workshop books, then a longer test is:
> >    > 
> >    > make regression-fresh
> >    > 
> >    > But the shorter one would be plenty interesting.  Thanks!
> > 
> > 
> ----------
> 
> 
> 
> 
> _______________________________________________
> Gcl-devel mailing list
> address@hidden
> http://mail.gnu.org/mailman/listinfo/gcl-devel
> 
> 
> 

-- 
Camm Maguire                                            address@hidden
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah