for what i's worth...

#!/usr/local/bin/perl
use strict;
use warnings;
use Benchmark qw[ cmpthese timethese ];

package Benchmark;

my @keys = ('s', 'm' x 10, 'l' x 100, 'h' x 1_000, 'r' x 100_000);

sub test {
    my ($iters, $key) = @_;
    my %hash = ();
    # regardless of which key we use, build all the keys so the time
    # needed to clone a big string doesn't overshadow the time to clon
+e
    # a small string.
    for (my $i = 0; $i < $iters; $i++) {
    $hash { (map { $_.$i } @keys)[$key] } = undef;
    }
}

my $many = 100;

cmpthese(5*$many, {
    'short' => sub { test($many, 0) },
    'medium' => sub { test($many, 1) },
    'long' => sub { test($many, 2) },
    'huge' => sub { test($many, 3) },
    'ridiculous' => sub { test($many, 4) },
});

__END__

laptop:~> monk.pl
Benchmark: timing 500 iterations of huge, long, medium, ridiculous, sh
+ort...

laptop:~> monk.pl
Benchmark: timing 500 iterations of huge, long, medium, ridiculous, sh
+ort...
      huge: 31 wallclock secs (31.57 usr +  0.00 sys = 31.57 CPU) @ 15
+.84/s (n=500)
      long: 32 wallclock secs (30.39 usr +  0.03 sys = 30.42 CPU) @ 16
+.44/s (n=500)
    medium: 33 wallclock secs (30.03 usr +  0.01 sys = 30.04 CPU) @ 16
+.64/s (n=500)
ridiculous: 187 wallclock secs (159.85 usr + 18.47 sys = 178.32 CPU) @
+  2.80/s (n=500)
     short: 30 wallclock secs (30.09 usr +  0.02 sys = 30.11 CPU) @ 16
+.61/s (n=500)
             Rate ridiculous       huge       long      short     medi
+um
ridiculous 2.80/s         --       -82%       -83%       -83%       -8
+3%
huge       15.8/s       465%         --        -4%        -5%        -
+5%
long       16.4/s       486%         4%         --        -1%        -
+1%
short      16.6/s       492%         5%         1%         --        -
+0%
medium     16.6/s       494%         5%         1%         0%         
+--
[download]

which seems to indicate that the key size can affect things, but only if you're dealing with seriously huge keys.

Note:

• zentara's benchmark doesn't take into account the fact that building up a big string takes longer then building up a short string .. hence i tried to make sure the only different between the tests was which key was added to the hash.
• sleepingsquirrel's benchmark reuses two keys over and over (which does test the time spent hashing the keys, and is fair about the time spent building the string) so he doens't consider the added cost of storing lots of big keys. some people don't consider memory utilization a valid thing to consider when doing a Benchmark like this, but if it takes longer to execute access data because the data structure is really huge, that seems like a fair thing to consider.