> to handle the same number of requests/sec you need much more workers, which has a lot of overhead

When doing multiprocessing, it's advisable to fork() instead of creating fresh processes (or launching them in containers etc). As long as you preload the necessary libraries/modules in the parent process before the fork, children will reference the same memory in a copy-on-write fashion. It keeps overhead extremely low.

This [1] article is for the Unicorn (ruby) webserver but the way it explains forking, signals, and pipes for inter-process control it's absolutely illuminating. These techniques are applicable to any unix-based applications, including php daemons.

[1] https://thorstenball.com/blog/2014/11/20/unicorn-unix-magic-...