Guys, here are some papers relating to this sort of thing.  I found them 
through the swiki, so someone must have interest in this area.

http://www.psrg.lcs.mit.edu/publications.html#cgc

Rob

At 03:37 AM 1/25/2002, Cees de Groot wrote:
>John M McIntosh <johnmci at smalltalkconsulting.com> said:
> >a) In case of an issue you can't kill the VM, go back and restore
> >from a change set because changes you make *are* persistent.
> >
>Make backups ;-)? I'm fully aware that this may be an issue, and I don't
>have a solution. Well, maybe one - some copy-on-write trickery but that'd
>quickly get really OS-dependent.
>
> >b) If your machine crashes I'm not quite sure what the state of your
> >image would be, are there pages outstanding to be written not yet
> >written?
> >
>That would need a sort of image checker.
>
> >c) If you wander around writing bits, how much IO would that
> >generate, especially if you do a full GC and compaction.
> >
>That's a real issue, especially if the image doesn't fit in physical RAM.
>Probably the allocator and GC would need to be modified to apply some
>heuristics so that objects that 'belong' together move 'closer' together (now,
>just define these two terms ;)).
>
>OTOH, it doesn't seem like a very hard patch, and it would be fun to try it
>out just for the heck of it. Can't do it myself, first have to build that
>textsearching thingy for the mailing list...
>
>--
>Cees de Groot               http://www.cdegroot.com     <cg at cdegroot.com>
>GnuPG 1024D/E0989E8B 0016 F679 F38D 5946 4ECD  1986 F303 937F E098 9E8B