lua-users home
lua-l archive

[Date Prev][Date Next][Thread Prev][Thread Next] [Date Index] [Thread Index]


In my particular situation, new objects tends to come in big clumps,
as well as collectable objects. I mark a whole huge table as
collectable, but have quite a long time (in the order of minutes)
before I'll need that memory for the next wave of new objects. So a
long slow collection is fine, as long as eventually the memory gets
collected some time before the deadline. I can see exactly how fast
the collection is going, because I know there are no new objects being
introduced.

What I did do was modify the GCSTEPMUL define down from 8 to 2 to slow
the collection and improve performance (lots of little collects rather
than fewer big collects) - I think it would be nice if those
parameters #defined at the top of lgc.c were exposed to the external
lua_gc interface so they could be tweaked on the fly. I'm not sure of
the exact implications of all of them, but more tunability seems like
a good thing in this situation.


On Fri, 28 Jan 2005 09:49:42 +0200, Asko Kauppi <asko.kauppi@sci.fi> wrote:
> 
> Does this all mean that 5.1 will/is not so well equipped for scarce
> memory situations (embedded devices)?
> 
> 28.1.2005 kello 04:15, skaller kirjoitti:
> 
>   On Fri, 2005-01-28 at 05:32, Chris Chapman wrote:
> >> Just a quick query to satisfy my curiosity about the new incremental
> >> garbage collector: if I have a state with a significant amount of
> >> collectable objects, does incrementally collecting with lua_gc("step")
> >> guarantee that it will definitely be collected at some point?
> >>
> >> The reason I ask is that the behaviour I'm seeing from the lua VM is
> >> that the GCthreshold remains constant at some high value, even though
> >> the garbage collector has been through several sweeps. If I call
> >> lua_gc and make it do a full (non-incremental) collect, it manages to
> >> clear up all of the collectable data. If I don't do a full collect,
> >> the data is never collected. No combination of parameters to SETPACE,
> >> SETINCMODE or STEP seems to make a difference.
> >>
> >> I would guess that I'm just misusing the system, but all I'm doing is
> >> starting the incremental collection at load time (GCRESTART), then
> >> constantly calling SETPACE, SETINCMODE and then STEP (in that order)
> >> once every tick.
> >>
> >> Any pointers to my obvious mistake?
> >
> > You probably didn't make one. Tuning an incremental
> > gc is a difficult feedback problem. It is unlikely --
> > and unnecessary --- that it guarrantee to eventually
> > collect garbage.
> >
> > What one would like to guarrantee is that the amount
> > of garbage will converge to some reasonable number
> > when the program is in an equilibrium state, and
> > that there will be some reasonable kind of upper bound on
> > total garbage (at all times).
> >
> > There has been discussion on this recently.
> > It seems the collector feedback 'circuits', and not
> > just the control parameter values, still need to be fiddled
> > with to achieve reasonable performance -- something which
> > is quite hard to define in the first place.
> >
> > Roughly speaking, the collector needs to collect >= garbage
> > generated over a period of time, which means when you start
> > spawning a lot of memory, it needs to ramp up automatically.
> > However it also needs to know how much garbage exists and
> > how much it collects to be self tuning -- and the first
> > piece of information cannot be available except after
> > a full collection.
> >
> > To put this another way, you have a source generating
> > both signal and noise and want to filter out the noise,
> > but unfortunately .. you don't know what the signal
> > looks like .. until after the noise is removed.. :)
> >
> > This can work using prediction based on prior performance,
> > but it isn't easy to get right, particularly since computer
> > programs don't obey any simple physical laws :)
> >
> > --
> > John Skaller, mailto:skaller@users.sf.net
> > voice: 061-2-9660-0850,
> > snail: PO BOX 401 Glebe NSW 2037 Australia
> > Checkout the Felix programming language http://felix.sf.net
> >
> >
> >
> 
>