model-writtenrun 9159 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tony Reix opened the thread seeking System V shared memory to obtain larger pages and improve AIX performance; Thomas Munro and Robert Haas distinguished dynamic segments from the main shared-memory region, and Robert suggested disabling USE_ANONYMOUS_SHMEM for the latter #1 #2 #3 #5.
- Andres Freund pointed to his earlier configurable shared-memory implementation from thread #31129, while Thomas investigated AIX's explicit large-page flags and required privileges #8 #9.
- Tony submitted an AIX-specific compile-time patch that selected System V memory and added large-page flags, reporting up to 37% improvement; he subsequently clarified that the measured gain came from switching mmap with 4KB pages to System V with 64KB pages under otherwise identical conditions, with 16MB pages still untested #11 #13.
- Thomas recommended respecting huge_pages and making the shared-memory implementation configurable, and Tony agreed but needed implementation help #12 #14 #15.
- Thomas supplied a two-patch replacement that revived Andres's shared_memory_type GUC and separately adapted Tony's AIX work to huge_pages settings and fallback behavior; he targeted PostgreSQL 12 rather than backpatching the feature #16.
- Tom Lane opposed the GUC's documentation and compatibility burden, favoring compile-time control or automatic selection; Robert Haas defended explicit configuration because operating-system allocation limits and performance trade-offs made automatic selection problematic, while Andres supported adding the setting #17 #18 #19 #20 #22.
- Thomas's FreeBSD experiments found a measurable high-concurrency benefit from System V memory with kern.ipc.shm_use_phys enabled, and he proposed retaining existing defaults while improving documentation and awaiting AIX testing of the second patch #23 #24.
- Thomas posted the revised GUC patch with documentation cross-references and again requested AIX huge-page testing; Tony's initial build attempt encountered an apparently unrelated documentation-tool crash #27 #28.
- Thomas committed #27 as f1bebef60e, with small documentation and commit-message edits, including clarification that FreeBSD and NetBSD memory-pinning advice concerns System V memory; the AIX huge-page component remained pending #30.
- After landing, Thomas identified that Linux accepted but ignored huge_pages=on with System V memory and proposed supporting SHM_HUGETLB to provide a testable counterpart to AIX; Tony's subsequent AIX results confirmed selectable mmap and System V mappings but still showed huge_pages=on failing #31 #32.
- After a seven-month gap, Thomas rebased the remaining AIX patch at Alvaro Herrera's request, explaining that it still needed AIX testing and review and might require allocation-size rounding; Linux System V huge-page support also remained a future proposal #33 #34.
- Alvaro warned that the untested remainder could be returned with feedback, then offered to move it to the November Commitfest when Tony resumed work; the last message proposed restoring S_IRUSR and S_IWUSR flags and described ongoing build and test-configuration difficulties, leaving AIX huge-page support uncommitted and awaiting validation #35 #36 #37 #38 #39.