Thanks to all who contrbuted to this thread. I have put together a quick benchmark. Unless I made an error in my code, map is the clear winner on performance grounds, as well as confusing any newbie that has to look at my code grounds ;-)

               Rate  or delete        do      copy     slice       map
or delete 1620746/s         --       -1%      -50%      -71%      -78%
       do 1642036/s         1%        --      -49%      -70%      -78%
     copy 3215434/s        98%       96%        --      -42%      -56%
    slice 5555556/s       243%      238%       73%        --      -24%
      map 7299270/s       350%      345%      127%       31%        --
[download]

#!/usr/bin/perl 
use strict;
use warnings;
use Benchmark qw(cmpthese);

# create a hash with a few values deviating
my $t_count = 3;
my %hosts  = map { $_ => $t_count - ( rand 50 > 49 ? 1 : 0 ) } (1 .. 1
+000);

my $count = 10_000_000;
cmpthese($count, {
    '       do' => ' do  {delete $hosts{$_} unless $hosts{$_} == $t_co
+unt} for keys %hosts      ',
    '      map' => ' map {delete $hosts{$_} unless $hosts{$_} == $t_co
+unt} keys %hosts          ',
    'or delete' => ' $hosts{$_} == $t_count or delete $hosts{$_} for k
+eys %hosts                ',
    '    slice' => ' delete @hosts{ grep { $hosts{$_} != $t_count } ke
+ys %hosts }               ',
    '     copy' => ' %hosts = map { $hosts{$_} == $t_count ? ($_, $hos
+ts{$_}) : ()} keys %hosts ',
});
[download]

Cheers,
R.