Re: Slowdown in switch to card-based collector?

[Dirk Herrmann]

On 26 Oct 2000, Mikael Djurfeldt wrote:

> > then it could just be that the heap is much larger with the newer
> > Guile -- structs became twice fatter not very long ago.
> 
> Good idea.  However, that change was introduced 2000-08-09, that is
> before both test points.

Heap size probably _has_ grown a bit, since the environments were added
after the first test point.  Further, Marius has implemented a different
approach to properties, the help system was changed etc.  I don't know if
these changes could sum up to such a big difference, though.

Another point is, if you use the new mechanism to find the stack top, the
scanned stack area will probably also be a little larger:  Up to now only
stack frames below 'main' were scanned.  With the new approach, you would
also scan stack frames belonging to the system's startup code even before
'main' is called.

Finally:  Conservative marking can lead to different gc times every time
guile has been modified a little.  For example, I regularly run the test
suite.  There are some tests that check for correct gc behaviour.  From
time to time these fail, because due to conservative gc sometimes an
object stays alive.  It's clear that this can lead to unpredictable
changes to marking time...

Best regards
Dirk