Own worker threads: a process-wide reserve of std::thread workers instead of a private Rayon pool (#15) by shinaoka · Pull Request #16 · tensor4all/threadteam-rs

benches: wait scenarios (short sections, gaps, pool held vs fresh, width, foreign load) with Rayon and libgomp references

shinaoka

and

claude

committed

a3a9320

waits: staged like the OpenMP runtimes (bounded poll spin, yields, then sleep) for members, joins and barriers; scheduling pressure throttles (faer#107)

shinaoka

and

claude

committed

904501e

waits: libgomp-like default budget (131072 polls, ~3 ms), pool throttle only on repeated preemptions; wait counters and an interleaved A/B script

shinaoka

and

claude

committed

c0c024b

waits: a preemption throttles the thread that saw it, not the pool; fresh-pool phases in the bench

shinaoka

and

claude

committed

2773c36

waits: back to the pool threshold (c0c024b) as the default, per-thread hold as a knob; libgomp held team and load in the A/B runs

shinaoka

and

claude

committed

d430347

waits: spin 300000 polls (GOMP_SPINCOUNT) by default; a knob to turn preemption detection off for the libgomp-like comparison

shinaoka

and

claude

committed

e76036c

waits: libgomp's waits by default (no preemption detection, spin 300000); detection and the pool throttle only during a new pool's warm-up (200 ms)

shinaoka

and

claude

committed

1f2145b

waits: libgomp's waits only (no preemption detection, no warm-up, no pressure knobs)

shinaoka

and

claude

committed

625a593

own worker threads: a process-wide reserve of std::thread workers leased per Pool, a native batch queue and broadcast list, cooperative waits (#15)

shinaoka

and

claude

committed

0e259b5

fix(own-threads): post-review deadlocks, cursor, affinity and docs

shinaoka

and

claude

committed

d6fe3f6

bench-ab: parse params that hold commas; keep the Rayon reference rows apart

shinaoka

and

claude

committed

78b59a1

alloc_free: a failing window reports each allocation's size, thread and backtrace

shinaoka

and

claude

committed

64684c4

fix(sched): `enter` hands its closure over without a Mutex

shinaoka

and

claude

committed

a11768f

Navigation Menu

Repository navigation

Own worker threads: a process-wide reserve of std::thread workers instead of a private Rayon pool (#15) - #16

Own worker threads: a process-wide reserve of std::thread workers instead of a private Rayon pool (#15)#16
shinaoka merged 13 commits into
mainfrom
feat/own-threads

Commits on Oct 29, 2026

benches: wait scenarios (short sections, gaps, pool held vs fresh, width, foreign load) with Rayon and libgomp references

waits: staged like the OpenMP runtimes (bounded poll spin, yields, then sleep) for members, joins and barriers; scheduling pressure throttles (faer#107)

waits: libgomp-like default budget (131072 polls, ~3 ms), pool throttle only on repeated preemptions; wait counters and an interleaved A/B script

waits: a preemption throttles the thread that saw it, not the pool; fresh-pool phases in the bench

waits: back to the pool threshold (c0c024b) as the default, per-thread hold as a knob; libgomp held team and load in the A/B runs

waits: spin 300000 polls (GOMP_SPINCOUNT) by default; a knob to turn preemption detection off for the libgomp-like comparison

waits: libgomp's waits by default (no preemption detection, spin 300000); detection and the pool throttle only during a new pool's warm-up (200 ms)

waits: libgomp's waits only (no preemption detection, no warm-up, no pressure knobs)

own worker threads: a process-wide reserve of std::thread workers leased per Pool, a native batch queue and broadcast list, cooperative waits (#15)

fix(own-threads): post-review deadlocks, cursor, affinity and docs

bench-ab: parse params that hold commas; keep the Rayon reference rows apart

alloc_free: a failing window reports each allocation's size, thread and backtrace

fix(sched): `enter` hands its closure over without a Mutex