Re: [Gcl-devel] Boyer benchmark results

[Camm Maguire]

Greetings!

"Paul F. Dietz" <address@hidden> writes:

> Camm Maguire wrote:
> 
> > 5) Paul, I was also a bit concerned with the posted results when you
> >    indicated that the clisp/cmucl ration should be 10x.  I just
> >    thought it worth noting that Peter gets 6.2x below, and with my
> >    rerun results without the printing I get 4.2x.  My machine is a
> >    dual 2.4Ghz Intel Xeon, and I'm wondering if we are seeing effects
> >    of differing memory bandwidths.
> 
> 
> I'd make a small wager that the problem is cache.  The Xeon has a really
> big L3 cache, doesn't it?  And I bet Boyer fits entirely within it.  That
> could affect the results quite substantially.
> 

Yes, to flesh this out a bit, boyer would have to fit in the xeon
cache, *and* cmucl would have to do something considerably more memory
friendly than gcl/clisp.  I can think of two possibilities for the
latter 1) copying gc for cons pages, improving locality, and 2)
emission of explicit athlon prefetch instructions. Can anyone confirm?
In any case, this points out an area in which gcl can improve in the
future, and is thus worth the observation.  We have a hybrid gc
strategy at the moment -- copying for 'relocatable' pages holding
strings, bignums, ... and unmoving collection for other types of
pages, including cons.  Though obviously more complicated to implement
given the fact that the copied data can contain pointers, perhaps a
set-cons-allocate-relocatable following the current gmp example might
be of eventual interest.

Take care,

>       Paul
> 
> 
> _______________________________________________
> Gcl-devel mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/gcl-devel
> 
> 
> 

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