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

William D Clinger will@ccs.neu.edu
Tue, 10 Dec 1996 17:15:32 +0000


Nick Barnes wrote:
> 1. Do you account for the cost of finding roots when collecting the
> older generations? Henry Baker's illustration does not.

My full paper does, but the abstract Henry cited does not.

> 2. I think the lesson we should all be learning here is that
> radioactive decay is often a very poor model indeed.

The radioactive decay model is a _horrible_ model of most
real programs, but it is an _excellent_ theoretical model
precisely because it does _not_ have any of the properties
of real programs that conventional garbage collectors are
designed to exploit.  (There are other, deeper reasons as
well, but less relevant to this thread.)

For example, Nick Barnes identified seven hypotheses that
could be exploited by a generational gc (in gclist-digest
V1#122).  Someone might ask whether generational gc can
possibly win when all of those hypotheses are false.  The
radioactive decay model tells us that it can.

So what?  So we can design garbage collectors that work
well for a broader range of programs.  It turns out that
the modifications you would make to a conventional generational
collector to make it work well with the radioactive decay
model also happen to make it work better with several kinds
of program behavior that people like Paul Wilson and Hans
Boehm have cited as the most common exceptions to the
generational hypothesis.  These modifications have very
little effect on programs that do satisfy the generational
hypothesis.

Will