You called it right earlier. Two layers of subroutine call (your's and the one Benchmark adds) wrapped around such trivial amounts of code serves to swamp the code you're trying to test.

Get rid of one layer of subroutine call and add an internal multiplier and your results will be both more consistant and a truer reflection of the actual cost of what you are testing. At least they are on my system:

use strict;
use warnings;
use Benchmark qw( cmpthese );
use constant DEBUG => 0;

our $N ||= -3;
our $debug = 0;
our $b = 100;

cmpthese $N, {
    constant   => q{
        for( 1 .. 1000 ){ my $a = 10 * $b; $a = 10 * $b if DEBUG; }
    },
    noconstant => q{ for( 1 .. 1000 ){ my $a = 10 * $b; } },
    var => q{
        our $debug;
        for( 1.. 1000 ){  my $a = 10 * $b; $a = 10 * $b if $debug; }
    },
};

__END__
C:\test>junk6 -N=-1
             Rate        var   constant noconstant
var        2222/s         --       -19%       -19%
constant   2731/s        23%         --        -1%
noconstant 2745/s        24%         1%         --

C:\test>junk6 -N=-3
             Rate        var   constant noconstant
var        2133/s         --       -20%       -25%
constant   2662/s        25%         --        -6%
noconstant 2831/s        33%         6%         --

C:\test>junk6 -N=-10
             Rate        var   constant noconstant
var        2172/s         --       -18%       -21%
constant   2659/s        22%         --        -3%
noconstant 2746/s        26%         3%         --

C:\test>junk6 -N=-20
             Rate        var   constant noconstant
var        2171/s         --       -18%       -21%
constant   2632/s        21%         --        -5%
noconstant 2759/s        27%         5%         --
[download]