Unfortunately, DigitalKitty's benchmark suffers from a couple of major flaws. Primarily, the subs under test are doing prints. That means that timing is going to be swamped by I/O. The second thing is that the hash is so small, that in the ops performed by each sub, the ones we're trying to test (each and keys) occur very few times, relative to the overhead of calling the sub, etc.

So I offer the following benchmark, which eliminates both of those sources of error.

use Benchmark;

my @words = do { local @ARGV = ( 'mondo_word_list.txt' ); <> };

my %w;
@w{@words} = @words;
@words = keys %w;

print scalar(keys %w), " words\n";

timethese( 10,
{
    foreach_loop => \&foreach_loop,
    foreach_loop_novar => \&foreach_loop_novar,
    while_each_loop => \&while_each_loop,
    array => \&array,
});

sub foreach_loop
{
    for my $key ( keys %w ) { $a = $key; $b = $w{$key}; }
}

sub foreach_loop_novar
{
    for ( keys %w ) { $a = $_; $b = $w{$_}; }
}

sub while_each_loop
{
    while( my( $key, $val ) = each %w ) { $a = $key; $b = $val; }
}

sub array
{
    for ( @words ) { $a = $_; $b = $_; }
}
[download]

Output: (slightly edited)

311142 words

      foreach_loop:  7 wallclock secs ( 6.22 usr 
foreach_loop_novar:  6 wallclock secs ( 6.17 usr
   while_each_loop:  5 wallclock secs ( 5.13 usr 
             array:  1 wallclock secs ( 1.08 usr
[download]

As you can see, for large hashes, while each wins over for keys. And you also gain a little by using the default iterator on the for loop.