Artificial benchmarking has proved to be a wise path to go down indeed. It has uncovered some issues where I was leaking a little bit (either PerlIO objects or the perl stack in general, hard to tell which), that weren't apparent in my (rather strenuous I thought) real-world testing. An update to 0.14 is coming sometime Sunday that moves some of the leaky XS code regarding converting socket FDs into perl io objects back in perl where at least it works correctly, and a change in the handling of socket closing is pending too, as my original understanding of the whole orderly tcp shutdown issue was wrong (it turns out to be an very application-protocol-specific thing, so I'll leave that to the module users if they need it).