Here are two tests.

If both files (two 12000 lines of rand 10e6's) are presorted with unix sort (70ms each on a PIII 500MHz), you can narrow in with an approximation method (seek, read a character from each file, compare and cut the range in half again.. not a great algorithm maybe) to see where the byte streams diverge. My Perl program took 10ms to get there, and then I stopped working on it.. Diff is faster I think (under 1ms). Perl overhead for just making those calls, see below, is under 1ms, so two sorts and a diff are about 140ms. This resembles the hash time.

`sort oddbig > oddbigsort`; # these are all backticks
`sort oddbig2 > oddbigsort2`;
@ans = `diff oddbigsort oddbigsort2`;
print "ANS: $ans[1]\n";
[download]

But those hash slices are positively gorgeous. They get my vote!

Incidentally japhy's code on the same system took 170 ms, see my implementation below. I wonder if the hash code involves something like a sort.

#!/usr/bin/perl -w

use strict;
use Benchmark;

my $t = &Benchmark::timeit(1,'

my %seen;

my $f1="oddbigsort";    my $f2="oddbigsort2";
open (A,$f1);           open (B,$f2);
my @smaller_list = <A>; my @larger_list = <B>; 

if ($#larger_list < $#smaller_list) {
 @seen{@smaller_list} = ();
 delete @seen{@larger_list};
}
else {
 @seen{@larger_list} = ();
 delete @seen{@smaller_list};
}

my @culprits = keys %seen;
foreach (@culprits) {print "$_\n";}

close (A); close (B);

');

print timestr($t),"\n";
[download]