Laurent Sansonetti
lsans****@apple*****
Fri Jun 8 16:24:11 JST 2007
I committed a working experiment as r1821!
It automatically defines the method directly in C by generating Libffi
closures.
$ cat bench.rb
require 'benchmark'
require 'osx/foundation'
def job
ary = OSX::NSMutableArray.alloc.init
str = OSX::NSString.stringWithString('foo')
ary.addObject(str)
num = OSX::NSNumber.numberWithInt(42)
ary.addObject(num)
count = ary.count
ary.removeAllObjects
end
puts Benchmark.measure { 1000.times { job } }
puts Benchmark.measure { 10_000.times { job } }
puts Benchmark.measure { 100_000.times { job } }
*** Old RubyCocoa ***
$ ruby bench.rb
0.170000 0.000000 0.170000 ( 0.168513)
1.640000 0.010000 1.650000 ( 1.686002)
16.450000 0.080000 16.530000 ( 16.764010)
$ ruby bench.rb
0.170000 0.000000 0.170000 ( 0.173943)
1.640000 0.010000 1.650000 ( 1.679036)
16.510000 0.100000 16.610000 ( 16.931537)
*** New RubyCocoa ***
$ ruby bench.rb
0.070000 0.000000 0.070000 ( 0.066736)
0.650000 0.000000 0.650000 ( 0.656822)
6.490000 0.040000 6.530000 ( 6.598461)
$ ruby bench.rb
0.070000 0.000000 0.070000 ( 0.067404)
0.650000 0.000000 0.650000 ( 0.671378)
6.520000 0.040000 6.560000 ( 6.705908)
In average ~3 times faster!
All tests and samples are working on my environment. I will verify on
other environments later.
If you find something wrong please tell me.
Laurent
On Jun 6, 2007, at 11:56 AM, Laurent Sansonetti wrote:
> After a quick investigation, it seems that #define_method is the
> culprit. It's really slow when used with Proc. Using a regular eval to
> add the method seems to be better.
>
> $ cat bench2.rb
> require 'benchmark'
> class Foo
> def method_missing(*a); foo; end
> def foo; end
> end
> Foo.class_eval { define_method(:foo2) { 0 } }
> Foo.class_eval("def foo3; end")
> n = 1_000_000
> Benchmark.bm(7) do |x|
> o = Foo.new
> x.report('method_missing') { n.times { o.missing } }
> x.report('define_method') { n.times { o.foo2 } }
> x.report('eval') { n.times { o.foo3 } }
> x.report('regular') { n.times { o.foo } }
> end
>
> $ ruby bench2.rb
> user system total real
> method_missing 1.030000 0.010000 1.040000 ( 1.081370)
> define_method 0.730000 0.010000 0.740000 ( 0.773041)
> eval 0.280000 0.000000 0.280000 ( 0.303355)
> regular 0.280000 0.000000 0.280000 ( 0.296050)
>
> Laurent
>
> On Jun 5, 2007, at 5:47 PM, Laurent Sansonetti wrote:
>
>> Hi,
>>
>> This afternoon I just tried a small experiment in the RubyCocoa
>> core. You know that every time you send a message from Ruby to
>> Objective-C, it goes through #method_missing, which is supposedly
>> slow.
>>
>> I experimented creating the missing method just after
>> #method_missing, so that #method_missing would be called only once
>> (the other messages would be "direct"). I thought it would be a good
>> performance gain.
>>
>> However, from a quick test, it doesn't seem to be better (it's even
>> sometimes worse).
>>
>> Any idea? I attached the patch and the bench test to this message.
>>
>> Laurent
>>
>> <
>> define_missing
>> .diff><bench.rb>_______________________________________________
>> Rubycocoa-devel mailing list
>> Rubyc****@lists*****
>> http://lists.sourceforge.jp/mailman/listinfo/rubycocoa-devel
>
> _______________________________________________
> Rubycocoa-devel mailing list
> Rubyc****@lists*****
> http://lists.sourceforge.jp/mailman/listinfo/rubycocoa-devel