Re: [Gcl-devel] GCL memory allocation and GC problems

[Camm Maguire]

Greetings!

"Vadim V. Zhytnikov" <address@hidden> writes:

> Well, notice that experimental value (really purely experimental I have not 
> any idea how to
> estimate optimal holesize theoretically)
> of 10000 pages for hole is absolute value, note relative to
> MAXPAGE.  Since now holesize is proportional MAXPAGE one can get
> larger holesize by building gcl with larger MAXPAGES.
> BTW, maybe we could change default MAXPAGE to some larger
> value. As far as I see there is nothing terrible wrong with it.
> Current value 128*1024 - 512Mb is based on typical RAM for
> modern PC.  As for hole size as some fraction of MAXPAGE.

512M seems good to me, though if anyone has arguments for a higher
value I'd certainly listen.  There is a certain amount of static
memory overhead for the maxpage setting.

> Too big holesize compared to maxpages is not good either.
> Do you remember some my other tests with very large hole
> size (almost whole MAXPAGES) - gcl just stops allocating new
> cell.  I don't understand precisely why - probably it is

There will be some granularity issues as a new hole allocation is
attempted near the end of maxpages.  We can make this more robust post
2.6.2, but in any case, I think the way we should think about this is
how many core expansions would we like before we allocate all of
maxpages?  Too few and one loses fine granular control over the core
size, too many and one spends ages in GC.

> related to the way how GCL extends hole if required
> (there is funny formula in alloc.c which counts available
> pages).  After some thinking I suppose that value 64*MAXPAGE/1024
> may be a good choice.
> 

Yes, you may have noticed that I've settled on a slightly more generous
value of MAXPAGES/10.  If anyone sees a problem please let me know,
but these settings have apparently sped up acl2 and axiom to some
degree as well as maxima.

Take care,

> 
> -- 
>       Vadim V. Zhytnikov
> 
>        <address@hidden>
>       <address@hidden>
> 
> 
> 
> 
> _______________________________________________
> 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