Re: thoughts about garbage collection

[urs@cs.stanford.edu (Urs Hoelzle)]

> Maybe someone already thought of this, but I'll give it a shot.  Has anyone 
> gathered stats on the amount of time spent scanning old-to-old references in 
> cards per scavenge?  Depending on that number, I have a suggestion that might
> speed up scavenging by quite a bit.

We haven't done such measurements, but you might want to look at my
paper in last year's OOPSLA GC workshop (available via ftp/Mosaic).
For the programs I measured there, the card scanning time doesn't
dominate scavenging time at all; only around 10% of the scavenge time
is spent there.  What you're proposing sounds like a "sequential store
buffer" (SSB) as used by the UMass people and in some ML
implementations (I think).  The paper estimates that for Self, a SSB
wouldn't work well because the frequency of stores is too high.

-Urs