Right. When I added default headers, the score went way down.

Probably libevent2 is the bottleneck and no way to make it faster without switching its backend (the cl-async's author is working on rewriting it with libuv now).

Out of curiosity did you try basic-binary-ipc? I would expect it would lose out to cl-async, but have never benchmarked the two against each other.

No difference even though I ran the benchmark again after removing the line.

Interesting. I was expecting at least cpu branch improvements given the lack of assert.

Also use gettimeofday() instead of clock(), because you don't know how far into the second you are at the time it's called.

(loop for i from 1 to 10 sum i)