Re: choosing the right benchmark code

[Matt Sergeant <ma...@sergeant.org>]

On Thu, 17 Feb 2000, Doug MacEachern wrote:
> On Wed, 16 Feb 2000, Stas Bekman wrote:
> 
> > I've discovered a weird thing while working the perlformance section of
> > the guide. I wanted to show that methods are a little bit slower than
> > functions. I wrote the benchmark to prove this, but the results I've
> > received were reversed to my expectations -- the methods were faster than
> > the functions. 
> 
> for something like this, less code == better test, CGI.pm does lots of
> things under the covers.
> 
> all you need is this:
> 
> package Foo;
> 
> sub bar { }
> 
> use strict;
> use Benchmark;
> 
> timethese(50_000, {
> 		 method => sub {
> 		     Foo->bar();
> 		 },
> 		 function => sub {
> 		     Foo::bar('Foo');
> 		 },
> 		});
> 
> Benchmark: timing 50000 iterations of function, method...
>   function:  1 wallclock secs ( 0.43 usr +  0.00 sys =  0.43 CPU)
>     method:  2 wallclock secs ( 0.89 usr +  0.00 sys =  0.89 CPU)

It's also worth noting that this is _very_ much dependant on your perl
version. Later versions of perl do much better method caching - I believe
in the current devel version of perl method calls are only very slightly
slower - rather than 2 times slower.

-- 
<Matt/>

Details: FastNet Software Ltd - XML, Perl, Databases.
Tagline: High Performance Web Solutions
Web Sites: http://come.to/fastnet http://sergeant.org
Available for Consultancy, Contracts and Training.

yanf@nadym.ru

[YANF WSCB <YA...@nadym.ru>]

unsubscribe

Re: choosing the right benchmark code

[Doug MacEachern <do...@pobox.com>]

On Thu, 17 Feb 2000, Matt Sergeant wrote:
> 
> It's also worth noting that this is _very_ much dependant on your perl
> version. Later versions of perl do much better method caching - I believe
> in the current devel version of perl method calls are only very slightly
> slower - rather than 2 times slower.

only a little better though, newer Perls make Foo->method() a bit faster
(some constant folding magic), but not Foo->$method().  and the 
improvement does not address the @ISA lookup that still happens in either
case.

Re: choosing the right benchmark code

[Autarch <au...@urth.org>]

On Thu, 17 Feb 2000, Matt Sergeant wrote:

> It's also worth noting that this is _very_ much dependant on your perl
> version. Later versions of perl do much better method caching - I believe
> in the current devel version of perl method calls are only very slightly
> slower - rather than 2 times slower.

At the risk of getting horrible completely off-topic:

9:46 ~ $ perl5.00503 -w  t.pl
Benchmark: timing 1000000 iterations of function, method...
  function:  3 wallclock secs ( 3.87 usr +  0.00 sys =  3.87 CPU)
    method:  6 wallclock secs ( 5.70 usr +  0.00 sys =  5.70 CPU)

9:47 ~ $ perl5.00503 -w t.pl
Benchmark: timing 1000000 iterations of function, method...
  function:  5 wallclock secs ( 3.85 usr +  0.00 sys =  3.85 CPU)
    method:  6 wallclock secs ( 5.71 usr +  0.00 sys =  5.71 CPU)

9:47 ~ $ perl5.00503 -w t.pl
Benchmark: timing 1000000 iterations of function, method...
  function:  3 wallclock secs ( 3.86 usr +  0.00 sys =  3.86 CPU)
    method:  6 wallclock secs ( 5.70 usr +  0.00 sys =  5.70 CPU)


9:50 ~ $ /usr/local/perl5.5.65/bin/perl5.5.650 -w t.pl
Benchmark: timing 1000000 iterations of function, method...
  function:  2 wallclock secs ( 1.67 usr +  0.01 sys =  1.68 CPU) @
595238.10/s (n=1000000)
    method:  5 wallclock secs ( 5.62 usr +  0.00 sys =  5.62 CPU) @
177935.94/s (n=1000000)

9:51 ~ $ /usr/local/perl5.5.65/bin/perl5.5.650 -w t.pl
Benchmark: timing 1000000 iterations of function, method...
  function:  0 wallclock secs ( 1.67 usr +  0.00 sys =  1.67 CPU) @
598802.40/s (n=1000000)
    method:  5 wallclock secs ( 5.62 usr +  0.00 sys =  5.62 CPU) @
177935.94/s (n=1000000)

9:51 ~ $ /usr/local/perl5.5.65/bin/perl5.5.650 -w t.pl
Benchmark: timing 1000000 iterations of function, method...
  function:  2 wallclock secs ( 1.68 usr +  0.00 sys =  1.68 CPU) @
595238.10/s (n=1000000)
    method:  6 wallclock secs ( 5.61 usr +  0.00 sys =  5.61 CPU) @
178253.12/s (n=1000000)



Well, judging from this function calls have somehow been sped up by two
by a factor of 2 but methods are untouched.  I wonder if this test is
somehow unrepresentative (I also wonder why I got 0 wallclock seconds for
a test that took 1.68 CPU).


-dave

/*==================
www.urth.org
We await the New Sun
==================*/