Pretty good write up - I reached the same conclusion doing something similar a few years back.

For the record, I got even better perfs with shared-memory by avoiding the mutex in `raw_sync::events::BusyEvent` - I used a blocking loop with atomics, and yielded the CPU (`yield_now` in rust) between the iterations of the loop.

Downside is that it becomes a bit less consistent. And it is a _lot_ more complex than a unix domain socket :)