I'm not convinced reading one line at a time is that bad, as perl and the system use buffered input behind the scenes. That's why they call it "buffered IO", anyway.

I've done the following test on two computers:

#!/usr/local/bin/perl -w
use Time::HiRes 'time';
$file = 'mshtmlc.h';
my @t = time;
open IN, '<', $file or die "Can't read file: $!";
while(<IN>) {
    # nada
}
close IN;
push @t, time;
open IN, '<', $file or die "Can't read file: $!";
while(read IN, $_, 1024) {
    # nada
}
close IN;
push @t, time;

printf <<'--', $t[1]-$t[0], $t[2]-$t[1];
line by line: %.3f s
1k blocks:    %.3f s
--
[download]

The file, an incude file coming with lcc.exe, is a text file of 1.67MB and close to 28000 lines. The results of this test:

desktop PC, 600MHz, Win98, 6GB disk:

line by line: 0.490 s
1k blocks:    0.060 s

XP laptop of 2.4GHz, 30GB 2in disk:

line by line: 0.026 s
1k blocks:    0.200 s

As you can see, I even get conflicting results.