The problem is that Devel::Size itself uses a substantial amount of memory. Internally, it builds a hash of all the SVs (scalars) it encounters whilst traversing the stucture being sized. It does has to do this in order to detect circular references that would otherwise cause it never to return.

For example, without circular reference detection, the following code would never return:

use Devel::Size qw[ total_size ];

my $scalar;
$scalar = \$scalar;

print total_size( $scalar );
[download]

Without Devel::Size, your 5e6 key hash with 33 char keys and undef values will consume a little less that 600MB.