Re: [Gcl-devel] Re: [Maxima] Maxima GCL init file update

[Vadim V. Zhytnikov]

Camm Maguire ?????:
> Hi Vadim, and thanks!
>
> 1) As you know, we are working on making some of these out-of-the-box
>    gcl defaults as well.  This init script will be just fine in any
>    case.

Right, maybe we'll be able to incorporate _some_ of these
or similar defaults in gcl 2.6.1 but this init can be used
with older GCL versions as well.  In any case I just want to
have them in forthcoming Maxima 5.9.1 release whatever
situation will be with GCL 2.6.1.

> 2) If you can be a bit more specific about which tests, and even which
>    routines or subsystems, are behind the "1.1-2" times slower figure
>    you cite, it would be most helpful.  Even try my latest gprof
>    patch, run a test which shows significant slowdown vs. CMUCL, and
>    profile it.  I'd like to get this figure to about "0.8" or so :-). 

Mostly tests with large polynomials - ratsimp, expand, factor.
On all these tests GCL is slower due to larger percent of GC
time even with new memory allocation parameters.
And I don't think that further allocation parameters
tweaking could give any serious performance boost
(only at the cost of large memory allocation which
is not the thing we want).
Further enhancement require more intricate
improvements like (1) generate less garbage as you
suggested (2) improve GC timing if there is any
possibility (profiling...).



> Take care,
>
> "Vadim V. Zhytnikov" <address@hidden> writes:
>
>
> > Hello!
> >
> > I'm about to upload into CVS new maxima-init.lisp.
> > It affects GCL Maxima only and performs some
> > GCL memory allocation tuning providing significant
> > performance boot especially on very complicated
> > problems (up to factor of 10 and more).
> > According to my tests with such init file
> > GCL Maxima is just 1.1 - 2 times
> > slower than CMUCL Maxima.  New init file works
> > quite well with Maxima 5.9.0 as well.
> >
> > But there is one issue I want to discuss.
> > Initial GCL Maxima memory footprint is now
> > about 11-13 Mb.  Partially due to larger
> > ANSI GCL image partially due to initial
> > memory preallocation. Is this is OK?
> > In principle one can reduce initial memory
> > footprint to 4-7Mb by removing initial
> > preallocation altogether but this slows Maxima
> > down by factor of 2 on small and middle
> > size problems.  Any ideas are greatly
> > appreciated.
> >
> > New maxima-init.lisp:
> > ============================================================
> > ;; Pre-allocating memory and more aggressive allocate growth
> > ;; parameters substantially improves GCL performance
> > #+gcl
> > (progn
> >
> >   #+gmp (si::set-gmp-allocate-relocatable t)
> >
> >   ;; If you really need smaller initial memory footprint
> >   ;; comment these pre-allocations out
> >   (si::allocate-relocatable-pages 400 t)
> >   (si::allocate 'cons   1000 t)
> >   (si::allocate 'fixnum 100  t)
> >   (si::allocate 'symbol 200  t)
> >   (si::allocate 'array  100  t)
> >   (si::allocate 'string 200  t)
> >   (si::allocate 'cfun   400  t)
> >
> >   ;; It is not recommended to make hole size less than 1000
> >   (si::set-hole-size 4000)
> >
> >   (si::allocate-growth 'cons    1 10000 66 33)
> >   (si::allocate-growth 'fixnum  1 10000 66 33)
> >   (si::allocate-growth 'symbol  1 10000 66 33)
> >   (si::allocate-growth 'package 1 10000 66 33)
> >   (si::allocate-growth 'array   1 10000 66 33)
> >   (si::allocate-growth 'string  1 10000 66 33)
> >   (si::allocate-growth 'cfun    1 10000 66 33)
> >   (si::allocate-growth 'sfun    1 10000 66 33)
> >
> >   (si::allocate-growth 'relocatable  1 10000 66 33)
> >   ;(si::allocate-growth 'contiguous   1 10000 66 33)
> >
> >   )
> > =====================================================
> >
> >
> > --
> >      Vadim V. Zhytnikov
> >
> >       <address@hidden>
> >      <address@hidden>
> >
> > _______________________________________________
> > Maxima mailing list
> > address@hidden
> > http://www.math.utexas.edu/mailman/listinfo/maxima


--
     Vadim V. Zhytnikov

      <address@hidden>
     <address@hidden>