Re: Lua 5.1 (work4) now available

["Adam D. Moss" <adam@...>]

Roberto Ierusalimschy wrote:
> That is strange, but not too strange. With the previous GC, mem figures
> would grow from a minimum (just after a collection) to twice that
> minimum, when a new collection was triggered. With an incremental
> collector, we do not have that flutuation. But it seems that the maximum
> memory used by both collectors would be more or less the same.

I agree that it sounds like it should be,to be.

> > Attached is a patch which tunes the new default GC tuneables
> > in accordance with my experiments in perf/peakmem tradeoff.
>
> Do you have any explanation why those numbers are better?

The patch just implements the constants from my tunings with
the last few 5.1works which I've sent the speed/mem numbers
to this list for in the past.  The constants reflect the
speed/mem sweet-spot which was reasonably visible (though not
utterly conspicuous) in those results, then I gave them another
spin with my apps to check that they were still a similar
improvement over 5.1w4 (which they were).  There's no great
logic to the numbers, just a lot of repeated testing with
different tuning.

Again the tuning carries the warning that what's a good tuning
for me might not be a good tuning for others, which is why it's
interesting to hear people try other benchmarks with it.

Regards,
--Adam
--
Adam D. Moss   . ,,^^   adam@gimp.org   http://www.foxbox.org/   co:3
"... so make sure your illness is early in the financial year."