model-writtenrun 8250 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro proposed preallocating DSM space inside main shared memory to reduce allocation overhead and enable huge pages, retaining operating-system allocation when the region fills; the patch also fixed DSA initialization for reused, nonzero memory #1.
- Robert Haas supported optional preallocation while preserving the ability to allocate more memory dynamically #2, and his preference for verb-object naming #4 led Thomas to use preallocate_dynamic_shared_memory and add documentation #5.
- Thomas reported substantially better parallel hash join scaling on a 72-thread machine #6, after which Andres Freund asked how much improvement came from huge pages versus allocation and page-table overhead #7.
- Thomas reported roughly 1.1× improvement without huge pages and attributed most of the remaining improvement to huge pages; he posted a revision named min_dynamic_shared_memory with a size_t-safe maximum, leaving hash-chunk and NUMA improvements for later work #8.
- Thomas committed #8 as 84b1c63ad4, adding changes to control and ordinary segment handle generation alongside documentation and comment edits.
- Thomas confirmed the feature had landed and left possible GUC renaming open; he proposed a future slot-and-generation handle design to simplify control flow and avoid linear scans under an exclusive lock, without attaching that further work #9.