On 2011-11-30 08:49, Charles Oliver Nutter wrote: > Seconds after I email...A DISCOVERY... > > It seems like invokedynamic is causing headaches for coro. Here's the > numbers with coro fibers and JRuby's invokedynamic support turned > *off*: > > 100 fibers / 1000 passes: 0.076000 0.000000 0.076000 ( 0.077000) > 100 fibers / 1000 passes: 0.076000 0.000000 0.076000 ( 0.076000) > 100 fibers / 1000 passes: 0.105000 0.000000 0.105000 ( 0.105000) > 100 fibers / 1000 passes: 0.070000 0.000000 0.070000 ( 0.070000) > > Perhaps the fix you (Lukas) made to get coro + indy working together > has impacted performance? Hm, maybe... the fix was really just a tiny tiny bugfix, so that shouldn't have caused any performance regressions, although, of course, I can't say for sure. But maybe something in invokedynamic has changed so that it's impacted by coro? I can reproduce it and I'll have to look into this. Has your usage of invokedynamic changed a lot since the last "perfect" performance numbers with invokedynamic?
- Lukas > Awesome, awesome, awesome to see the above perf numbers...regardless! > > - Charlie > > On Wed, Nov 30, 2011 at 1:46 AM, Charles Oliver Nutter > <head...@headius.com> wrote: >> Ok, good news and not as good news! >> >> The good news is that coro seems to be working in the latest >> openjdk-osx-build, and it definitely improves JRuby's coroutine >> performance! >> >> For running bench_fiber_ring with 100 fibers passing 1000 messages, >> here's numbers for the threaded impl: >> >> 100 fibers / 1000 passes: 1.070000 0.000000 1.070000 ( 1.071000) >> 100 fibers / 1000 passes: 1.070000 0.000000 1.070000 ( 1.070000) >> 100 fibers / 1000 passes: 1.092000 0.000000 1.092000 ( 1.092000) >> 100 fibers / 1000 passes: 1.077000 0.000000 1.077000 ( 1.077000) >> >> And with the coro impl: >> >> 100 fibers / 1000 passes: 0.215000 0.000000 0.215000 ( 0.215000) >> 100 fibers / 1000 passes: 0.217000 0.000000 0.217000 ( 0.217000) >> 100 fibers / 1000 passes: 0.212000 0.000000 0.212000 ( 0.212000) >> 100 fibers / 1000 passes: 0.216000 0.000000 0.216000 ( 0.215000) >> >> Hooray! >> >> Now for the not-as-good news... >> >> Here's Ruby 1.9.3 on the same benchmark: >> >> 100 fibers / 1000 passes: 0.160000 0.000000 0.160000 ( 0.155562) >> 100 fibers / 1000 passes: 0.150000 0.000000 0.150000 ( 0.156581) >> 100 fibers / 1000 passes: 0.160000 0.000000 0.160000 ( 0.155351) >> 100 fibers / 1000 passes: 0.150000 0.000000 0.150000 ( 0.156776) >> >> Now even getting close to 1.9.3 is really awesome, but I'm wondering >> if either I'm doing something wrong (maybe broke something in the >> coro-based fiber impl?) or if something regressed in coro, because >> Lukas's blog post showed JRuby + coro performing significantly >> *better* than C Ruby. >> >> This is also a bit tricky to profile...since call stacks are jumping >> around a bit :) A dumb sampled profile doesn't show much other than >> Ruby code being hit heavily...which I'd expect. >> >> Lukas: Are you able to reproduce these numbers with JRuby master and >> bench/bench_fiber_ring.rb? Here's the command line I'm using: >> >> jruby --1.9 -Xfiber.coroutines=true bench/bench_fiber_ring.rb 20 100 1000 >> >> - Charlie > _______________________________________________ > mlvm-dev mailing list > mlvm-dev@openjdk.java.net > http://mail.openjdk.java.net/mailman/listinfo/mlvm-dev _______________________________________________ mlvm-dev mailing list mlvm-dev@openjdk.java.net http://mail.openjdk.java.net/mailman/listinfo/mlvm-dev
