hello, this might be more a linux question, but maybe you have experienced this too. I spotted that problem when working with mod_perl and Apache::SizeLimit.

For reproducing I made a simple perl one-liner (well, more like 13 lines) that has a big string (100,000,000) which uses about 200MB RAM.
if I fork this script, on a linux kernel 2.4 all memory is shared, so still 200MB in usage for both. only if I change the string in the child, memory is copied (400MB). I think this is called copy-on-write (correct me if I'm wrong).
now, with a 2.6 kernel on amd64 (debian both systems), almost no memory is shared (like 1MB or so). here's the test script:

perl -we'my $cmd = qq{top -b -n 1 |grep };
print "$$ start\n";print qx{$cmd $$};
my $s="";my $x = "_" x $ARGV[0];
print "string\n";print qx{$cmd $$};
if (my $pid = fork) {
  print "parent $$ forked $pid\n";print qx{$cmd $$};
  $s = "<".$s.">";
}
else {
  print "child $$\n";print qx{$cmd $$};
  $s = "(".$s.")";sleep 3; print "child slept 3\n";
}
print "$$ end\n";print qx{$cmd $$};' 100000000
[download]

perl 5.6.1 and 5.8.4 on kernel 2.6 both show that effect, while 2.4er kernels all share the whole memory (somebody else tested the above script on several machines).

so, question: if this was a bug I would have probably found it with google, but I didn't find anything helpful. what can I do about it, has anyone else experienced this? I guess a pre-forking Apache/mod_perl is pretty common, so there must be others who have the same problem...

update: s/200/400/g and s/100/200/g for correct sizes