Is pre-linking worth it?

I've got similar results from some of my embedded work too. prelinking saves about 3-5% of system ram in my cases because it prevents paging in of the relocation and symbol tables, and it doesn't dirty as many pages in the shared libraries.

20% doesn't seem that surprising to me, glibc for instance has 50k you have to load just to do symbol resolution. You'd only need to fault in 250k of text from glibc to get to 20% overhead. Even my desktop has only faulted in 224kb of glibc.