MLDBM is bound to slow down your script somewhat, due to the nature of what it's doing for you (storing your multi-level datastructures on disk). If you read the WARNINGS section of the MLDBM doc, you'll see that it does matter how you access your MLDBM-tied hashes.

In particular, this means that instead of doing this:

for my $k (keys %{$h{something}}) {
    print $h{something}{$k}[0]{foo}{bar};  # FETCH _every_ time!
}
[download]

.. you're better off doing this, to avoid FETCH-ing the top-level value every time:

my $root = $h{something};                  # FETCH _once_
for my $k (keys %$root) {
    print $k->[0]{foo}{bar};
}
[download]

(Examples taken from MLDBM doc)

You might want to review your code for these kinds of situations.

HTH