Ok, good news first. The recent drop of fixes from Hotspot seem to
have fixed the problems in methodHandleWalk.cpp. Hooray!
The bad news is that performance seems to have *severely* degrade
since previous builds.
Here's JRuby running something simple, the fib benchmark:
~/projects/jruby ➔ jruby -X+C --server bench/bench_fib_recursive.rb 10 35
9227465
3.540000 0.000000 3.540000 ( 3.489000)
9227465
2.767000 0.000000 2.767000 ( 2.767000)
9227465
2.773000 0.000000 2.773000 ( 2.773000)
Those numbers were in the 0.7s range as recently as this past weekend.
Other benchmarks seem similarly crippled, posting as much as 10x worse
performance than a build from this weekend.
A sampled profile doesn't help much...it just claims almost all time
is spent in the fib method itself:
Compiled + native Method
94.3% 1317 + 0 bench.bench_fib_recursive.method__0$RUBY$fib_ruby
0.1% 1 + 0 RicochetBlob
94.4% 1318 + 0 Total compiled
Now I am running an x86_64 build, since that's what Stephen's
update.sh builds now. I remember Christian mentioning that the
ricochet frame stuff had only landed for i386. Is that the source of
this perf degradation?
I'm going to try to twiddle Stephen's script to do an i386 build now.
- Charlie
_______________________________________________
mlvm-dev mailing list
mlvm-dev@openjdk.java.net
http://mail.openjdk.java.net/mailman/listinfo/mlvm-dev
