2. Even for 2011, 1k rps with an average latency of 100ms seems laughably bad. I have to be missing something here.
Also, note ratio of clients to server is 100:1. So 100ms latency almost has to be the maximally experienced: that single core node cannot spend more than 1ms per request on average, or it will fall behind.
The classic problem with these benchmarks are that they fall to what Gil Tene calls coordinated omission. I can park some of the clients for a while and process the other clients far quicker. It improves the req/s and if I'm just coming back to the parked requests within, say, 100ms, everything will look fine. But there is a lot of variance on response times.
Perhaps that has something to do with it
Don't ever choose python for performance. Its definitely fast enough for most things, but its just not the tool to leverage if you actually need performance.
Its wonderful if you just want to get shit done though, there are just so many good libraries around that you can use to get your project finished without committing several times the effort it takes in other languages