[shinaoka]

Design study for #22. Theory only, nothing implemented. It adds one file against develop: docs/design/hybrid-scheduling.md.

Maintainer direction (2026-10-30): 「threadteamを独自スレッドで拡張していくとしよう．OpenMPの機能をメインにするが，Rayonの柔軟性も取り込めないか？理論的に検討させよう」. Addendum: 「threadteamはOpenMP的にして，work stealingはRayonにますという手もあるね．例えば，tenferroのCPUEngineが両方のプールを持ってもいいわけだし．」

Recommendation

One runtime, team first, with scoped stealing.

• Team sections stay as they are: reserved, co-scheduled, with barriers and stable numbers.
• Tasks from join, scope and prepared graphs are stolen only inside the creating token's domain. Pool scope is opt-in, as with OpenMP 6.0's threadset.
• Stealing is child stealing. Continuation stealing needs compiler support or a cactus stack.
• Waits. A waiter runs only descendants of what it waits for (libomp's task scheduling constraint, oneTBB isolate). Tasks carry no team authority, by type (TaskExec). A plain barrier runs nothing; a draining barrier is opt-in.
• Budget. Stealing creates no authority. Finished siblings' workers can help the rest inside the domain. A nested team claims idle domain workers or is refused, and never draws from the reserve.
• Dependent tasks. Prepared graphs (taskgraph-like) with priority by remaining critical path.
• Determinism. Reproducible reductions over fixed chunks.

The two-pool option (an engine owns a threadteam pool and a Rayon pool) is evaluated item by item:

• concurrent use oversubscribes, so phase exclusion is required;
• idle CPU of the inactive side: Rayon's 33 search rounds against spin_polls;
• stranded shares: the P/m loss of serial inner kernels;
• placement;
• the ownership rule rewrite, which is a maintainer decision;
• Parallelism::Pool;
• the OpenMP and TBB precedent (KMP_COMPOSABILITY, libirml IPC, MKL_THREADING_LAYER).

It is kept as the baseline every native stage has to beat. The note ends with a comparison table and six deciding measurements.

Coverage

• Sections 3 to 10 answer questions 1 to 8 of Design study: OpenMP-style teams with Rayon-style flexibility on threadteam's own workers (theory only) #22 in order.
• Section 11 covers the two-pool items (1) to (6).
• Section 12 is the comparison table and section 13 the deciding measurements.
• Section 10 holds the invariants I1 to I10 with proof sketches, the TLA+ properties and the staging table (stages 0 to 5, each with its kernels and measurement).

Sources

• Citations. Every citation was opened and checked against the source text.
• Corrections to the brief. Fibril is SPAA 2016. KMP_COMPOSABILITY, libirml and --ipc come from the SciPy 2018 paper. oneTBB has no src/irml. No OMP_* variable controls deterministic reductions. Kokkos tasking was removed in 5.0.
• Not verified. The claims that could not be checked are listed at the end of the note.
• Method. pi-council in plan mode (GPT-6-astra chair, DeepSeek, GLM), in two meetings. A first single meeting stopped on an invalid chair decision.

Docs only. Base develop. The monitor merges.

🤖 Generated with Claude Code