Here is a memory and CPU optimized version of kyle (++) 's solution:

#! /usr/bin/perl
use strict;

my @lines = <DATA>;
my @matching = map {{IDX=>$_,  ID_KEY => ( $lines[$_] =~ /id:\s?(\d+)/
+ )}}
          grep $lines[$_] =~ /id:\s?\d/, 0 .. $#lines;
my @sorted_matching = sort { 
                              $a->{ID_KEY} <=> $b->{ID_KEY}
                            } @matching;

@lines[map {$_->{IDX} } @matching] = 
   @lines[map {$_->{IDX}} @sorted_matching];
print @lines;
__DATA__
#Comment
*ss id addr id: 099 bbb ccc aaa
*ss id addr id: 003 aaa bbb ccc
#Comment 2
*333 23 ss id: 002 aaa bbb ccc
*22 233333333 34432 233 44
[download]

Memory is optimized because this avoids copying the entire $line[$_] for sort purposes. Instead, it copies only the extracted sort key, and index. Since the regular-expression for extracting the sort key is performed only once per record (instead of on each compare on the sort), CPU is optimized.

I would be the first to admit that this probably classic "premature optimization" if the data size and frequency of use is small, but it was fun.