256 MiB capacity · callback iterations 0 · ordinary write · publish MAX

Fixed test procedure

10 million messages per run, 50,000 warmup messages, callback iterations 0. Input is prepared before write timing, and every payload word varies.

Fixed round limit: 214,748,364 B, equal to 80% of the 256 MiB layer-2 budget. All queues use identical round boundaries; the final round ends at the actual message count.

Only accumulated producer write time is timed. End-of-round handoff, tail publication and callback draining occur outside timing; every message must still be consumed. TwoBufQueue copy and callback threads run independently.

Ordinary write reads back the first byte after the complete copy; Session does not add this readback.

Plateau rules for combined selection

Use common configurations measured at all four lengths. Writer candidates are three neighboring block sizes growing by powers of two; native SPSC candidates are three neighboring spin or wait settings.

The four-length composite scores of three neighboring configurations differ by at most 10%; the coefficient of variation of composite scores grouped by run number is at most 20%. Individual results, variation and neighborhood changes for every length are retained.

Every configuration passes complete validation. Three throughput and three latency runs succeed, with zero write failures, zero Writer backpressure and resolvable latency. Reference coefficients of variation remain capped at 20% / 25% / 40% / 60%.

All metrics use one configuration. Search-stage results are not mixed with earlier independent rechecks. This is the best combined plateau within the measured common configurations, not proof of a global optimum.

Host and build

macOS-14.4.1-arm64-arm-64bit · arm64 · 8 logical CPUs.

Clang 18.1.8 · Release · -O3 -DNDEBUG · no sanitizer.

Producer user_interactive QoS, without hard affinity; copy and callback threads are unpinned. Timer: mach_absolute_time; 1,000 writes per timing group.

Additional measurements use the original executable with identical SHA-256. Upstream headers, complete payloads and copy strategies are unchanged. AC power was confirmed beforehand.

Data and auditability

The original batch contains 1,284 complete validations and 9,708 measurements, with 0 failed runs. The additional ReaderWriterQueue neighborhood contains 24 complete validations and 144 measurements, with 0 failures.

The three reports share the same native SPSC controls. Copy strategies are fixed as disclosed in the table; they and compiler-generated code also affect results.

Earlier per-metric selections are backed up. All four metrics here use common configurations; reference values were fixed before additional measurements. Detailed run settings record actual mode and random_seed values.