Re: Benchmarking large scores 1.3.4 vs 1.3.5

[Mikael Djurfeldt <mdj at mdj dot nada dot kth dot se>]

Jan Nieuwenhuizen <janneke@gnu.org> writes:

> In short, for small scores results are comparable, but for large scores
> (no swapping, though), 1.3.5 is much memory-frienlier (about 1/5 less
> memory), whereas the processing time shoots up by a factor 3.5.

> This probably is a LilyPond thing, but it would be good if someone
> would benchmark with large applications.

Recent 1.3.5's have a new policy for triggering heap allocation and
determining the sizes of heap segments.  It fixes a number of problems
in the old GC:

1. Couldn't handle two separate pools of cells.  (Would run wild,
   allocating memory for certain programs.)

2. Would sometimes allocate far too much heap so that the Guile
   process became gigantic.

3. Would sometimes allocate too little so that too much time was spent
   in GC.

4. Would often trigger heap allocation several times in a row.  (The
   new scheme predicts how large the segments needs to be in order not
   to need further allocation.)

The new GC scheme also is prepared for POSIX threading.

It was introduced in March this year.