Hi

Nice improvement!

> Issue #4962 has been updated by Eric Hodel.
>
> Status changed from Open to Assigned
>
> I have made three runs of `make benchmark` using the following revisions =
of ruby:
>
> ruby 1.9.2p180 (2011-02-18 revision 30909) [x86_64-darwin10.8.0]
>
> ruby 1.9.3dev (2011-07-05 trunk 32413) [x86_64-darwin10.8.0]
>
> The benchmark bm_vm_thread_mutex3.rb was disabled as it presented an an e=
xtreme outlier for 1.9.2p180.

Ah, yes. This is the reason why I rewrote GVL. We should ignore it.


> I took the total time it took all benchmarks to run.
>
> The first run is with the ruby checkout
>
> 1.9.2: =A0204.890 206.312 209.319
> 1.9.3: =A0210.793 215.815 214.773
> diff: =A0 5.903 =A0 9.503 =A0 5.454
>
> (For diff, smaller is better)
>
> The second run is with --disable-gems for 1.9.3. =A0I modified RUNRUBY in=
 Makefile:
>
> RUNRUBY =3D $(MINIRUBY) $(srcdir)/tool/runruby.rb --extout=3D$(EXTOUT) $(=
RUNRUBYOPT) -- --disable-gems

I recommend to change $(MINIRUBY) to ./ruby if your 1.9.2 is not
miniruby. It help to avoid
see unrelated benchmark difference. :)


>
> 1.9.2: =A0215.472 206.452 205.110
> 1.9.3: =A0201.837 194.694 191.747
> diff: =A0 -13.635 -11.758 -13.363
>
> The third run is with my changes to delay work in rubygems.rb:
>
> 1.9.2: =A0208.982 211.249 208.637
> 1.9.3: =A0198.714 201.984 198.293
> diff: =A0 -10.268 -9.265 =A0-10.344
>
> Here are the average differences from 1.9.2-p180:
>
> stock ruby trunk: =A0 =A0 =A0 6.953
> --disable-gems: -12.919
> rubygems patches: =A0 =A0 =A0 -9.959
>
> Is the slowdown of 2.96 seconds between --disable-gems and my fixes acros=
s all benchmarks acceptable?
>
> Should I look for additional improvements?

Great!

Can you please tell us a result of vm3_gc and io_file_read? They have
most big degressions
and I'm worry about it.

Anyway, personally I think it is acceptable and no more improvemnt
because usually people only
compare 1.9.2 and 1.9.3 and don't compare individual patches in 1.9.3 chang=
es.

Endoh-san, What do you think?