model-writtenrun 5849 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Geoff Blake proposed using ISB to delay ARM spinlock loops in #1, aiming to provide behavior analogous to x86 PAUSE despite negligible initial pgbench gains.
- Tom Lane requested evidence across ARM implementations in #2, and Geoff supplied Graviton2 measurements showing essentially unchanged pgbench performance and approximately 3% improvement in HammerDB in #3.
- Tom introduced a separate test_shm_mq contention harness in #5 because PostgreSQL's ordinary workloads had shed much of their spinlock contention; his Apple M1 results showed little benefit from the proposed delay.
- Andres Freund questioned whether the harness actually made all workers contend for one lock in #6 and #8; Tom explained its shared setup-lock reuse in #7 and #9, and Andres acknowledged his misreading in #10.
- Geoff tested the harness on a 64-core Graviton2 machine and reported roughly 10% lower average latency at 63 workers, with overlapping variability, in #11; additional measurements in #12 supported benefits mainly at high worker counts.
- Tom committed the implementation from #1 as a82a5eee31, adding an ARM64-only guard and explanatory comments; his announcement in #14 cited marginal M1 Pro gains, no measurable Raspberry Pi regression, and missing ISB support on pre-v7 ARM.
- The thread ended with Geoff's thanks in #15; the ARM64 implementation had landed, while the experimental contention harness from #5 was not included in the commit and no further work was requested.