1. I'd rather not use a solution using external modules, since the collector will have to run on MANY hosts as painlessly and non-intrusively as possible. Is there a high performance approach without external modules?
The modules cited in the code that Thelonius
posted (IPC::Open3, IO::Select, Symbol) are not "external", in the sense that they are all included in the standard Perl "core" distribution. That is, wherever a reasonably current version of Perl is installed, these modules are also installed by default.
(If these many hosts you speak of have non-standard or hopelessly outdated perl installations, that's going to be a problem anyway.)
As for buffer size, the 1024 bytes suggested by Thelonius is fine for handling text that must ultimately be treated in a line-oriented fashion; if you're dealing with really high data rates, an 8k buffer should be about optimal.