What it is
Level-triggered epoll reports a descriptor for as long as data is buffered; edge-triggered epoll reports it once per change, so code that reads only once per wake can leave data stuck until something else arrives. When many workers wait on one listening socket, each new connection wakes all of them unless EPOLLEXCLUSIVE or SO_REUSEPORT is used. io_uring replaces most system calls with two shared rings, batches submissions into one io_uring_enter, can avoid even that with SQPOLL, and can overflow its completion ring if completions are reaped too slowly. This lab steps through each case and counts the calls.
When to use it
Use it to explain why edge-triggered code must read until EAGAIN, to see how many spurious wake-ups a shared listening socket causes, or to understand io_uring batching, SQPOLL and CQ overflow before reading a strace.
Gotchas & limits
These are small models, not the kernel: EPOLLEXCLUSIVE is modelled as waking exactly one waiter, SO_REUSEPORT uses a stand-in hash, and io_uring ticks are rounds of work, not time. The syscall counts of epoll were checked against a real kernel; the ring model is a teaching model.