It turns out that I haven't been the only person surprised by the TCP incast pattern, including folks who don't use Erlang at all (e.g. Hadoop users). I was happy to write about it, to help spread the word. Even if I'd stumbled upon Kyle Brandt's posting http://blog.serverfault.com/2011/06/27/per-second-measuremen... ("Per Second Measurements Don’t Cut It"), I don't think I'd've realized the full consequences of packets dropped by microbursts of traffic when there's "obviously plenty of bandwidth available".