Hi all,
At Ola's request I've been trying to figure out why RDoc performance
is as bad as it is. Well, I don't see any cheap wins there, but here
are some observations:
1. RDoc is doing a ton of low-level byte pushing. Consider
RDoc::RubyParser#get_tk at rdoc/parsers/parse_rb.rb:1465 (which is
where RDoc spends the bulk of its CPU time at the parsing stage). Many
of these low-level operations are several times slower in JRuby, until
you turn on JIT compiler.
2. Many methods invoked under get_tk are not JIT-compiled because of
optional or "rest" (*args) arguments.
3. At this level of granularity, profiling is not accurate enough to
pinpoint an actual bottleneck, even if there is one. I don't, by the
way, think that there is - whichever way I profiled it, CPU time is
spread across multiple very low-level methods.
One thing that does take much time is DefaultMethod.prepareOptOrRestArgs().
Consider this code:
--------------------------------
def foo(*args) end
start = Time.now
1000000.times { foo }
finish = Time.now
puts "Elapsed time: #{finish - start}"
--------------------------------
It prints out the following:
Ruby
Elapsed time: 0.460686
JRuby
Elapsed time: 2.536
JRuby with DefaultMethod#prepareOptOrRestArgs() stubbed out:
Elapsed time: 0.865
JRuby with (*args) removed from script (enabling JIT compilation of foo)
Elapsed time: 0.632
I will continue playing with this stuff and post my findings in this
thread. Any ideas/feedback are appreciated.
--
Alex Verkhovsky
---------------------------------------------------------------------
To unsubscribe from this list please visit:
http://xircles.codehaus.org/manage_email
