Re: [Gcl-devel] gcl 2.7.0t5 and Maxima 5.9.1

[Vadim V. Zhytnikov]

Camm Maguire writes:
> Greetings, and thanks Vadim for looking into this as always!
>
> "Vadim V. Zhytnikov" <address@hidden> writes:
>
>
> > Hi!
> >
> > I did some little gcl 2.7.0t5 testing with
> > Maxima 5.9.1.  And I'm very impressed
> > by new twc memory consumption.  On all test
> > maximal memory footprint is almost the same as
> > in CMUCL.
> >
> > On the other hand I observe some slowdown.
> > Performance degradation is not so significant
> > and even with it gcl is faster of as fast as CMUCL.
> > I have to stress that slowdown is not solely due to
> > smaller hole and more frequent relocatable pages GC.
>
>
> Others have noticed some slowdown, at times approacing 15%.  My guess

I observe even bigger slowdown.  For example 31 vs 37 seconds
(net time without GBC). It's about 20%.

> is that the compiler work I've been doing has silently eliminated some
> earlier optimizations, inadvertently of course, but in fact the
> compiler now knows a lot more about types, and if everything doesn't
> match, then the code produced can be less optimial at the moment.  It
> would be most helpful if you could break this down using the gprof
> facility or otherwise.  Build atop a gcl build with --enable-gprof in
> t5 and 2.6.6 -- all will be somewhat slower due to the profiling
> overhead, but the rations should be the same.  THen sandwich your test
> between (si::gprof-start) and (si::gprof-quit).  Then please prune it
> down to the extent possible and post your results.  If you see a
> particularly poorly compiled routine, please post the disassembly in
> t5 and 2.6.6 (i.e. (disassemble 'foo)).  This sort of feedback is
> really essential to make sure we don't give up performance as we move
> toward compilance.
I'll do this as soon as time permits ...

 >>I also observe some small problems:
> > 1. It seems that set-hole-size doesn't work with
> > gcl 2.7.0t5 - to be more precise it has no effect on
> > GBC process.
> >
> > 2. I see wrong GBC messages - lot's of GBC messages
> > for STRUCTURE while in reality this is probably
> > 4w strauctures CFUN GBC.
>
>
> Could you send me a small test so I can reproduce here locally?

Test and it's output is in the attachment.
See STRUCTURE and CFUN in (room) and GBC messages.
I also tried various (si::set-hole-size NNN) in
maxima-init.lisp - no effect whatever.


--
     Vadim V. Zhytnikov

      <address@hidden>
     <address@hidden>

test.out.gz
Description: GNU Zip compressed data

:lisp (room)
:lisp (setq si::*notify-gbc* t)
:lisp (si::gbc-time 0)
w:ratsimp((x+y+z)^300)$
showtime:true$
factor(w);
:lisp (room)
:lisp (format t "~%GBC time ~4,2f seconds" (/ (si::gbc-time) 100.))