big project ported

[Jeff Sturm]

Tom Tromey wrote:
> Jeff> Some GC profiling code would also be a great addition, to
> Jeff> measure the frequency and duration of GC, among other
> Jeff> things... I don't know if such a thing exists for the boehm
> Jeff> collector.
> 
> Offhand I don't know if the collector has this info.  Newer versions
> might have more, though.  (Upgrading is on my to-do list, still.)
> Specific suggestions for what to have would be good; we could make PRs
> for them so they wouldn't get lost.

I'd like to know:

 - how often the collector runs
 - how long (on average) the collector runs
 - how many objects (total) are allocated
 - average allocation size

...among other things.  the first two are loosely reported by the
-verbosegc option of the JDK.  I don't know if there are appropriate
hooks in the Boehm collector offhand.  The last two measurements can be
done in libgcj proper.

At this moment I strongly suspect weaknesses in the GC.  I have a JDBC
driver working under gcj now, so I can run some benchmarks against large
datasets.  For small runs (<1000 rows), gcj outperforms the IBM JDK by
about 21%.  For larger runs (>250000 rows), gcj is far slower than the
JDK.

Did you say you and Hans were exchanging notes about the boehm-gc
interface in libgcj?  Anything interesting?

-- 
Jeff Sturm
jsturm@sigma6.com