performance of script to write very long lines of random chars

[gry]

Dear pythonistas,
   I am writing a tiny utility to produce a file consisting of a
specified number of lines of a given length of random ascii
characters.  I am hoping to find a more time and memory efficient way,
that is still fairly simple clear, and _pythonic_.

I would like to have something that I can use at both extremes of
data:

   32M chars per line * 100 lines
or
   5 chars per line * 1e8 lines.

E.g., the output of bigrand.py for 10 characters, 2 lines might be:

gw2+M/5t&.
S[[db/l?Vx

I'm using python 2.7.0 on linux.  I need to use only out-of-the box
modules, since this has to work on a bunch of different computers.
At this point I'm especially concerned with the case of a few very
long lines, since that seems to use a lot of memory, and take a long
time.
Characters are a slight subset of the printable ascii's, specified in
the examples below.  My first naive try was:

from sys import stdout
import random
nchars = 32000000
rows = 10
avail_chrs =
'0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!"#$%&
\'()*+,-./:;<=>?@[\\]^_`{}'

def make_varchar(nchars):
    return (''.join([random.choice(avail_chrs) for i in
range(nchars)]))

for l in range(rows):
    stdout.write(make_varchar(nchars))
    stdout.write('\n')

This version used around 1.2GB resident/1.2GB virtual of memory for
3min 38sec.


My second try uses much less RAM, but more CPU time, and seems rather,
umm, un-pythonic (the array module always seems a little un
pythonic...)

from sys import stdout
from array import array
import random
nchars = 32000000
rows = 10
avail_chrs =
'0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!"#$%&
\'()*+,-./:;<=>?@[\\]^_`{}'
a = array('c', 'X' * nchars)

for l in range(rows):
    for i in xrange(nchars):
        a[i] = random.choice(avail_chrs)
    a.tofile(stdout)
    stdout.write('\n')

This version using array took 4 min, 29 sec, using 34MB resident/110
virtual. So, much smaller than the first attempt, but a bit slower.
Can someone suggest a better code?  And help me understand the
performance issues here?

-- George