The loop span through its body regardless of whether there was anything
to do -- ~10% of a desktop core for an empty shard, ~70% of a small VPS
core, and a process that never idles is exactly what burstable vCPU
plans throttle.
The loop now blocks in NetState.WaitForCompletion whenever every queue
it drains is empty, waking on the next timer tick or the moment work
arrives. Receive completions, new connections, and cross-thread
LoopContext.Post (via IORingGroup 1.0.10's sticky Wake) are all in the
wait set, so sleeping adds no latency to any of them; only timer-driven
logic sees wheel lag, bounded by server.eventLoopIdleWaitMs (default
2ms, 0 = never sleep).
Measured on a real world of 190k items / 33k mobiles: 10.4% of a core
to 0.8-1.0%, with peak tick lag unchanged. Spin mode independently
gained 7x the iterations per core from the ring's AcceptEx rework.
Sleeping also gives the GC natural pause points, which the old spin
loop denied it -- memory no longer climbs until a save forces a
collection.
A sleep is bounded by the next wheel turn, so a correctly honoured
sleep can never miss a deadline; the only way sleeping harms the wheel
is the host returning the wait late. That overshoot is measured on
every sleep, and an escalating backoff (server.lateWakeThreshold)
suspends sleeping when it persists -- server work like saves or heavy
commands cannot trip it by construction. Hosts without high-resolution
waits are detected once at startup and spin instead. The admin gump
shows the verdict instead of the now-meaningless CPS figure, which is
removed.
Time accounting for diagnosis is compiled out of normal builds: build
with -p:EventLoopProfiling=true to enable EventLoopProfiler (per-phase
wall time, sleep overshoot, GC pauses, stolen-time residual, ~15min
ring buffer) and the [LoopStats command with CSV dump. See
dev-docs/debugging-event-loop.md for the diagnosis funnel.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
6aedbbe2ef