[gclist] Name that hypothesis
Thu, 05 Dec 1996 17:41:46 +0000
> Henry Baker wrote, in excerpts from two messages:
> > If the 'null hypothesis' is that the likelihood of
> > dying is completely independent of how long you have lived already,
> > then the 'generational hypothesis' is one that says that young objects
> > die even more often than would be expected by this null hypothesis.
> > Now second order effects -- e.g., cache locality -- might allow you
> > to take advantage of even the 'null hypothesis', but to win better than
> > this you will need the 'generational hypothesis'.
> > I didn't mean to imply that the curve had to have discontinuities, but
> > that it had to have a different shape/slope in some places than the
> > curve for the 'null hypothesis'.
> The claims Henry is making here are commonly believed, but
> they are not true unless you take a very narrow view of
> generational garbage collection. I have designed, simulated,
> and analyzed a generational collector that derives a clear
> first order advantage over non-generational collectors for
> the radioactive decay model Henry proposed in SIGPLAN Notices,
> April 1993. This advantage can be demonstrated with a
> perfectly uniform memory system.
> Henry's claims are probably true of all generational collectors
> that are being used for production work in 1996. I just want to
> warn that Henry is talking about current practice, not theory.
Many generational collectors currently out there will still win for
the radioactive decay model, because collecting an old generation
requires scanning all younger generations but collecting a young
generation only requires scanning the remembered set in the older
generations. But IMO this is not what the "generational hypothesis"
addresses; this is the subject of the "directional hypothesis" (or
whatever we decide to call it).