I recall a while ago discovering that async I/O latency -- the time it takes for a write() to wake up a thread watching the peer fd with libevent -- was surprisingly high compared to the equivalent synchronous I/O where the receiver is woken up from a blocking read().

I do not recall whether this was Linux or Mac, but I *think* it was limited to Mac. In any case, this might have a significant performance impact on whatever platform has to deal with it, because libevent is the basis for MessagePumpForIO on all POSIX platforms and thus the basis for I/O thread scheduling and IPC.

We should take time to measure this and possibly do something about it.