I stand corrected.

my %set;
my $how_many = 10_000_000;
for( my $i = 0; $i < $how_many; $i++ ) {
    $set{$i} = undef;
}

print 'mem usage: ', my_mem(), "\n";

sub my_mem {
    my ($proc_info)
        = grep { $_->[2] == $$ } map { [ split ] } `ps l | tail -n +2`
+;
    return $proc_info->[6];
}
__END__
mem usage: 1323312
[download]

Changing $set{$i}=undef to $set{$i}=1:

mem usage: 1402268
[download]

Using undef instead of 1, you save about 78M on ten million items—about a 6% difference. Or you could look at it as 8 bytes per item, unless I did my math wrong (which becomes more and more probable as time t approaches lunch).

When $set{$i}=10 (no chance of using sv_yes):

mem usage: 1402268
[download]

When $set{$i}='' (empty string):

mem usage: 1872484
[download]

And finally, Set::Light:

use Set::Light;

my $set = Set::Light->new();
my $how_many = 10_000_000;
for( my $i = 0; $i < $how_many; $i++ ) {
    $set->insert( $i );
}
__END__
mem usage: 1127960
[download]

It beats them all! It beats the undef case by 195M. Note, however, that most of my tests ran in 20–25 seconds. The Set::Light test took much much longer—almost four minutes. I'm pretty sure all that time is spent in destruction, because the test reports its results fairly quickly and then takes a long time to exit.