On Thu, Jul 8, 2021 at 10:40 PM Hannu Krosing <han...@google.com> wrote: > > Very nice results. > > I have been working on the same problem but a bit different solution - > a mix of binary search for (sub)pages and 32-bit bitmaps for > tid-in-page. > > Even with currebnt allocation heuristics (allocate 291 tids per page) > it initially allocate much less space, instead of current 291*6=1746 > bytes per page it needs to allocate 80 bytes. > > Also it can be laid out so that it is friendly to parallel SIMD > searches doing up to 8 tid lookups in parallel.
Interesting. > > That said, for allocating the tid array, the best solution is to > postpone it as much as possible and to do the initial collection into > a file, which > > 1) postpones the memory allocation to the beginning of index cleanups > > 2) lets you select the correct size and structure as you know more > about the distribution at that time > > 3) do the first heap pass in one go and then advance frozenxmin > *before* index cleanup I think we have to do index vacuuming before heap vacuuming (2nd heap pass). So do you mean that it advances relfrozenxid of pg_class before both index vacuuming and heap vacuuming? Regards, -- Masahiko Sawada EDB: https://www.enterprisedb.com/
