Would be good to get a throughput baseline before any future optimization experiments.
Conversely, avoid going overboard on optimizations until there's a baseline...
There is now a UDP client/UDP upstream test. In terms of seeing what overheads the service itself has, this is probably the most useful one to exercise. Could conceivably add another for TCP or HTTP upstream later, but that would likely just be bottlenecked on the client socket.
Assuming the numbers are right, it looks like we're getting about 12 kqps as of fbaed2a2, so now we've got a baseline before trying out some refactoring. Might also be able to hook up pprof or something while the benchmark is running as well.