[gclist] "This Old Space", with apologies to Will Clinger

[Nick Barnes]

I note, in response to Henry Baker's analysis, that GC effort is not
necessarily proportional to the amount of marking, and that this is
especially not the case in a "partitioning" GC (i.e. one which does
not condemn the whole heap at once), such as a generational GC (in
which, for instance, only one generation is condemned at once). In
such a setting, finding pointers into the condemned set may be more
costly than marking, and considering "mark/cons ratios" may not be
revealing.

For instance, in a simple generational GC with many equal-sized
generations, with a "directional mutator" (i.e. one for which the
directional hypothesis is true), the cost of a collection of
generation N is approximately proportional to N (because it includes
scanning generations 0 to N-1).

[Incidentally, this statement illustrates why any genuinely useful
model of mutator behaviour needs to take account of directionality: it
is not true for a non-directional mutator, because collecting any
generation involves scanning all generations.]

Nick B