[gclist] Re: GC performance

[Basile STARYNKEVITCH]

>>>>> "Stefan" == "Stefan Monnier" <stefan.monnier@lia.di.epfl.ch> writes:

[[quoting me, Basile]]

    >> My overall non-expert opinion is that using a GC and allocating
    >> everything (without even using a stack) on a GC-ed heap is
    >> worth it. The advantage of reifying call frames (as
    >> continuations or closures) is significant for other reasons
    >> (call/cc, powerful control structures, ease of debugging, etc),
    >> so even a 25% penalty is IMHO acceptable.

    Stefan> I agree in that the simplicity of the heap-allocation
    Stefan> makes it very attractive.  But the 25% estimate you quote
    Stefan> can be misleading: depending on your cache architecture,
    Stefan> the slowdown can get much higher because of cache misses.

[[I knew a little bit about it - Appel and others mentions it - but it
is not yet an issue for me, because most of the computers I can use
today don't have big L2 caches]]


[[I'm skipping interesting stuff by Stefan]]

    Stefan> [...]  Has anybody tried to avoid most of
    Stefan> the space cost of copying collection by having several old
    Stefan> generations and only collect one at a time (and hence only
    Stefan> need 1 additional such old generation, which can be chosen
    Stefan> much smaller than the typical 50% of the whole memory). I
    Stefan> feel like this would pose interesting problems to reduce
    Stefan> inter-generation references.


I actually did code a GC which don't condemn all old generations, but
only some of them. Since I didn't tune the GC, and more importantly,
since it is not yet used because I'm still ending the design of the
langage and implementing the compiler, I don't have any figures.

(and when it will run, I don't know in which journal could I publish
any results, if it is worthwhile - suggestions welcome by email).


--

N.B. Any opinions expressed here are solely mine, and not of my organization.
N.B. Les opinions exprimees ici me sont personnelles et n engagent pas le CEA.

Please cite a **pertinent part** of my mail in all answers
Veuillez citer une **partie pertinente** de mon courrier dans vos reponses



----------------------------------------------------------------------
Basile STARYNKEVITCH   ----  Commissariat a l Energie Atomique (civil)
DRN/DMT/SERMA * CEA/Saclay bat.470 * 91191 GIF/YVETTE CEDEX * France
fax: (33) [1] 69.08.85.68; phone: 69.08.40.66; homephone: 46.65.45.53
email: basile.starynkevitch@cea.fr (or else basile@soleil.serma.cea.fr);  
I speak french, english, russian. Je parle francais, anglais, russe.
----------------------------------------------------------------------