Replace the per-call std::thread::scope fan-out in run_indexed with one
process-wide rayon pool. run_indexed's callers nest — sweep (grid
points), monte_carlo (seeds), and walk_forward (window bounds), where a
walk_forward window runs an inner sweep — and each std::thread::scope
spawned its own available_parallelism() OS-thread batch, so the two
levels multiplied to ~available_parallelism()^2 threads (measured
150-240 concurrent on a 24-core box during a 42-cell campaign, ~10x
oversubscription). rayon's pool is process-wide and persistent: a nested
par_iter work-steals within the same N workers instead of spawning, so
live workers never exceed the pool size.
Determinism (C1) is preserved bit-for-bit: run_indexed collects via
IndexedParallelIterator in enumeration order (not completion order), and
the float aggregation (summarize_r / McAggregate / stitch) runs after
the ordered collect, so IEEE-754 non-associativity introduces no
variance. The existing _with_threads(1) == _with_threads(8) == public
determinism tests stay green, now driving local rayon pools of 1 and N
workers via ThreadPool::install.
Structural inversion: the public sweep/monte_carlo/walk_forward become
the core (calling run_indexed on the ambient pool through a shared
assemble_* helper); the _with_threads variants collapse into thin,
test-only wrappers (#[cfg(test)]) that run the same assembly inside a
local pool. run_one stays Fn + Sync (no + Send): run_indexed borrows it
in the map closure and the wrappers borrow it into install — passing it
by value would move it into rayon's Map and force F: Send, a bound the
public callers do not carry (hence the load-bearing
#[allow(clippy::redundant_closure)] on run_indexed).
Adds a nested-oversubscription regression test (asserts nested
run_indexed never exceeds the pool size — the pre-rayon shape ran
POOL^2) and an ignored wall-time benchmark. rayon admitted under the
amended-C16 per-case dependency policy; it stays within C1 (parallelism
across sims, never within one).
This is the shared pool only — the first of three steps toward
saturating the cores. The diagnosis found the larger loss is the idle
valleys (~42% of the box idle: the sequential cell loop and the
single-threaded bootstrap), which the pool alone does not fill. Two
follow-ups are filed and recorded on #268: the Registry write path is
not concurrency-safe (append_family races on a shared families.jsonl),
the prerequisite for parallelizing the C1-disjoint cell loop with a RAM
budget bounding concurrently-active instruments (the FileCache retains a
symbol's parsed files while any reference is live).
Verified: cargo test -p aura-engine (276 passed, the C1 determinism
tests green), the new oversubscription test green, the benchmark runs
(no overflow), clippy -D warnings clean, cargo test --workspace no
failures.
closes#268
claude
b048923f1c