The best approach would be to leverage the benefits of Lexical Scoping. ...a brief example...

for( 1 .. 100 ) {
    my @array;
    my $c = 0;
    while( $c++ < 500_000 ) {
        push @array, rand;
    }
    print "\@array holds ", scalar( @array ), " elements.\n";
}
[download]

On each iteration of the outer "foreach" loop, @array is declared, filled, checked for an element count, and then falls out of scope, at which time the memory is released back to Perl for the next iteration.

If you watch Perl's memory usage during the run of this script you will see that after the first iteration of the 'foreach' loop, Perl never requires any significant additional memory.

The other thing to consider is your algorithm itself. Do you need the entire file to be slurped into an array? Or can you iterate over it line by line and process each line individually? The latter will almost always be more memory efficient. And finally, even if you do slurp the entire file into an array, each time you slurp it again into the same array, the previous contents are discarded and that memory becomes available again to Perl. Nevertheless, careful use of lexical scoping solves a whole slew of potential problems, memory usage being only one of them.