I already have some reasonably good data about the size of the complied perl code (not the native executable) and how much of it is ours and how much comes from CPAN, etc.; using Devel::Size & Devel::Size::Report. With this analysis it is clear that most of this is our own code. And I really am talking about the size of the compiled code, not data.

Actually, changing the build options did make a big difference. I think excluding threading was the big win. Altogether those options, coupled with a switch to 5.10 from 5.8, saved about 30%. But even as it stands, the in-memory image include a lot of overhead just for the filenames and module names: about one instance of each of those strings for each executable line of code in the source files.