Completely ignorable exploration codebase.
This is a benchmark for various approaches to having shared proxy backend connections in a multi-threaded memcached proxy. I've been experimenting with various libevent approaches to see how threads scale and what the perf output looks like. It's pretty bleak.
I've also been learning io_uring and figuing out how to plug it in here.
It's not as straightforward as I hoped:
- In this proxy case we want to stuff requests sequentially down a backend, note them in a queue, and read them back at any point.
- Due to async nature of
io_uring, this means we can get a read before a write is confirmed. Or writes are done out of order, etc. - Which means we have to confirm writes and then schedule reads, vs simply waiting on a persistent epoll read hook. In the libevent code this is mitigated by holding a lock around the backend object during the write syscall.
I've also tried handing writes off to the event threads but this scales
poorly. My next attempt at io_uring will be to batch the write syscall
across different backends simultaneously.
Mixing uring and libevent isn't always easy either. Also the timers aren't super fast.