JuliaLang / julia / 1558

78%

Build Type

push

buildkite

Committed by web-flow

Commit Message

scheduler: avoid O(nthreads) wake-storm on every `@spawn` (#61826)

Fixes #61820
Fixes #50425

Linux - Ryzen 9 5950X
<img width="1560" height="720" alt="image"
src="https://github.com/user-attachments/assets/e3f667fb-6fed-46ed-837f-dfd1b8dd925a"
/>

Linux - Ryzen Threadripper PRO 7995WX 96-Cores
<img width="1560" height="720" alt="image"
src="https://github.com/user-attachments/assets/359e75b8-1a7e-4596-99a7-e7e5878f1b4d"
/>


Windows - i7-8700
<img width="1560" height="720" alt="image"
src="https://github.com/user-attachments/assets/1f9f4671-3e9a-45a7-b196-04dd779ed2f3"
/>



macOS - M2 Pro 6 p cores
<img width="1560" height="720" alt="image"
src="https://github.com/user-attachments/assets/0a5f7ea3-d443-4b23-a42b-b4ab3ee9940b"
/>







Developed with Claude:

---

`schedule` for a non-sticky task previously broadcast a wake to every
thread via
`jl_wakeup_thread(-1)`, performing a per-thread lock/signal/unlock under
`wakeup_thread`'s loop. Per-insert cost was linear in `jl_n_threads`,
and on
systems where the producer can be preempted (e.g. SMT + oversubscribed
thread
count on Windows/Linux) every iteration hit the kernel park/unpark path,
producing the >100x slowdown reported in #61820.

Add `jl_wakeup_threadpool(tpid)`, which wakes at most one sleeping
thread in the
target pool, with a round-robin start hint to spread wake load. Workers
re-check
the queue before sleeping (the existing store-buffering dance), so
bursty inserts
naturally wake additional consumers across the per-insert calls without
a
broadcast.

Restricting wakes to the task's own threadpool is also a correctness
improvement,
since `Partr.multiq_deletemin` only ever returns tasks from the caller's
pool --
waking out-of-pool threads was pure overhead.

The round-robin start hint is sharded across 64 cache-padded stripes
indexed by
the producing thread's tid. A single global atomic counter became the
dominant
cost of `@spawn` at high producer counts on multi-die parts (e.g. Ryzen
5950X... (continued)

Coverage Stats

2 of 2 new or added lines in 1 file covered. (100.0%)

482 existing lines in 17 files now uncovered.

66004 of 84824 relevant lines covered (77.81%)

22990385.8 hits per line