Further perf investigation shows that a big delta on a raw throughput test has to to do with how libevent deals with write events in the bufferevent. bufferevent_sock does not use edge triggered behavior, so it deletes and then re-adds the write event on every write. I have a proof of concept that does not use bufferevent and the perf improvement is 20-25%. Along with another change that I have going to swap out the header map implementation it gets it up to around 40% improvement. That is tracked in: #120