I didn't understand Benchmark. It's pity of me. I should have passed function as reference and didn't need timethese(). With 2 million test and revised my poor code with your advice, the difference becomes smaller but they didn't reverse.

use strict;
use warnings;
use Benchmark qw/cmpthese timethese/;

my %h;
$|=1;

sub with_keys{
    foreach my $k (keys %h){
        #no proc
    }
    return;
}
sub with_each{
    while( my($k,$v)=each %h){
        #no proc
    }
    return;
}

for my $max ( qw(10000 100000 1000000 2000000) ){
    print "\ncase $max\n";
    %h=map { $_ => 1 } (1 .. $max);
    print "hash ready ... count=" . (scalar keys %h) ."\n";
    print scalar gmtime, "\n";
    print "-" x 40 ."\n";
    cmpthese(
        -20, #20 cpu time
        {
            'test_keys'=> \&with_keys,
            'test_each'=> \&with_each,
        }
    );
    print "-" x 40 ."\n";
    print scalar gmtime, "\n";
}

__DATA__
case 10000
hash ready ... count=10000
Fri Nov 11 13:27:23 2011
----------------------------------------
           Rate test_each test_keys
test_each 109/s        --      -61%
test_keys 281/s      158%        --
----------------------------------------
Fri Nov 11 13:28:12 2011

case 100000
hash ready ... count=100000
Fri Nov 11 13:28:12 2011
----------------------------------------
            Rate test_each test_keys
test_each 6.10/s        --      -32%
test_keys 8.93/s       46%        --
----------------------------------------
Fri Nov 11 13:29:04 2011

case 1000000
hash ready ... count=1000000
Fri Nov 11 13:29:09 2011
----------------------------------------
          s/iter test_each test_keys
test_each   1.88        --      -24%
test_keys   1.42       32%        --
----------------------------------------
Fri Nov 11 13:30:05 2011

case 2000000
hash ready ... count=2000000
Fri Nov 11 13:31:05 2011
----------------------------------------
          s/iter test_each test_keys
test_each   3.89        --      -24%
test_keys   2.96       31%        --
----------------------------------------
Fri Nov 11 13:32:43 2011
[download]

Thanks for your reply.