Re: GC improvements

[Ludovic Courtès]

Hi,

address@hidden (Ludovic Courtès) writes:

> This interleaving of initialization and sweeping makes it pretty hard to
> track exactly where fresh cells come from.  I guess one solution might
> be to maintain a list of the uninitialized segments and pick cells
> directly from there before actually sweeping.

Thinking about it, this solution looks like the beginning of the
generational approach to GC: we'd define uninitialized segments (i.e.,
those whose FIRST_TIME field is true) as one generation, and the others
as another generation.

I think we could even have 3 segment generations:

1. The kids, i.e., the uninitialized segments;

2. The youngsters, i.e., segments recently initialized and where
   sweeping gave good results in the past;

3. The elderly, i.e., those for which sweeping has been unproductive for
   some time already.

Remarks:

* This is a very coarse approach to generational GC as it's done on a
  per-segment basis.  For this reason, it doesn't make sense to have
  more that 3 generations (due to fragmentation, each segment may
  include a variety of kids, youngsters, and elderly).  Even 3
  generations (instead of just 2) is a questionable approach.

  OTOH, even though there is fragmentation, I think it's reasonable to
  assume that some segments, those created and populated at startup
  time, will actually only contain elderly cells, whereas segments
  initialized during the life time of the program (especially for
  long-running or interactive applications) may be more fragmented and
  may be considered as "young".  Hence the intermediary generation.

  Additionally, I'm assuming that elderly segments cannot move back to
  the "youngsters" generation, which seems to be generally assumed in
  generational GC.

* This poor man's generational GC has the advantage of being quite easy
  to implement.  Basically, instead of having a single segment table as
  we have now, we'd have to maintain a set of three segment tables,
  which would not be too complicated it seems.

  In particular, it's much easier to implement than the per-cell
  generational GC, as proposed by Greg Harvey in
  http://home.thezone.net/~gharvey/guile/ggc-notes.txt .  It's also less
  intrusive: basically `scm_i_sweep_some_segments ()' is the only place
  where generations would be taken into account.  Of course, the outcome
  would certainly be better with a per-cell approach, but I'd expect the
  per-segment approach to yield a non-null benefit for a pretty low
  cost.


What do you think?  Are there flaws in this reasoning?  :-)

Thanks,
Ludovic.