GC performance - was Re: [gclist] GC topics

[Darius Blasband]

> 
> I think there are some major differences in our assumptions.  If 
> you are working with a closed system, then it is true that it is 
> not hard to implement precise GC.  How often do you get to do this?  
> Your point is perhaps purely academic.
> 
Not really. It is a production issue for real world systems, in a closed
world indeed. 

> > - While it can be a pragmatic issue, I don't believe that the ability
> >   to plug a GC in an existing system tells much about how easy it
> >   can be to implement one. 
> 
> As it happens, it is also easy to write a conservative GC.
> 
Well, I guess I must be dump, or misinformed, so please help me. My
initial idea of a conservative GC was about a system which does
scan memory, and uses heuristics to determine what might be valid pointers.
If this assumption is true, it maybe is easy, but seems like very low-level
stuff to me. How efficient can such a test be ?

It maybe is easy to write, but how easy would it be to optimize ?  What 
happens with portability ? If I am wrong regarding what such GC systems 
are, please let me know.

> > - Quite amazingly, no one seems to mention the possible advantages of 
> >   precise GC.
> 
> If there's no hope of using precise GC (because of uncooperative 
> compilers and/or libraries) then its advantages, possible or actual, 
> are not useful.  Furthermore, it appears that most of the advantages 
> are less dramatic than theory would predict.  At least, that's what 
> I get from reading papers by Wilson and by Boehm.  There's also system 
> costs -- if I use a conservative GC, I can, in practice, make use 
> of an already-written compiler that performs a number of optimizations. 
> If I use precise GC, that's probably not an option.  The GC may be 
> faster, but if the rest of the code is slower (because it is underoptimized) 
> what's the point?                    

Well, the point is pretty easy to make: our compiler generates intermediate
C source files, and there is no contradiction between using a precise GC
and still have very decent performance.

How dramatic these advantages are ? Really, I don't know. If my assumption
about the conservative GC is right, then I really cannot expect such a 
system to be even comparable with a precise GC.

Regards,

Darius