I just saw that and decided to time the mechanism I suggested. At 1/3rd of a second for the first million, it even surprised me :)

#! perl -slw
use strict;
use Devel::Timer;

my $T = new Devel::Timer;

sub firstNprimes {
    my $n = shift;
    open my $primes, '<:raw', 'primes.all' or die $!;
    my @primes = split ' ', do{ local $/ = \( $n * 10 ); <$primes> };
    close $primes;
    return \@primes;
}

$T->mark( '201' );
my $ref1 = firstNprimes( 201 );

$T->mark( '100,000' );
my $ref2 = firstNprimes( 100_000 );

$T->mark( '1,000,000' );
my $ref3 = firstNprimes( 1_000_000 );

$T->report;
__END__
C:\Perl\test\data>..\junk4



Devel::Timer Report -- Total time: 0.3176 secs

Interval  Time    Percent

----------------------------------------------

02 -> 03  0.3169  99.78%  100,000 -> 1,000,000

01 -> 02  0.0007   0.21%  201 -> 100,000

00 -> 01  0.0000   0.01%  INIT -> 201
[download]

That's the ascii version, and it does slow down quite quickly due memory allocation as you go larger; 2 million takes 8 seconds.

I might try the binary file version later.