Re: Iterating on IndexTuple attributes and nbtree page-level dynamic prefix truncation

On Thu, 17 Jun 2021 at 17:14, Matthias van de Meent
<boekewurm+postgres(at)gmail(dot)com> wrote:
>
> I'll try to
> benchmark the patches in this patchset (both 0001, and 0001+0002) in
> the upcoming weekend.

Somewhat delayed, benchmark results are attached. These are based on 7
iterations of the attached benchmark script ('scratch.sql'), with the
latest 'UK Price Paid' dataset. Again, the index_test table is an
unlogged copy of the land_registry_price_paid_uk table, with one
additional trailing always_null column.

Results for 0001 are quite good in the target area of multi-column
indexes in which attcacheoff cannot be used (2-4% for insertion
workloads, 4-12% for reindex workloads), but regresses slightly for
the single unique column insertion test, and are quite a bit worse for
both insert and reindex cases for the attcacheoff-enabled multi-column
index (4% and 18% respectively (!)).

With 0001+0002, further improvements are made in the target area (now
4-7% for the various insertion workloads, 5-14% for reindex). The
regression in the insert- and reindex-workload in attcacheoff-enabled
multi-column indexes is still substantial, but slightly less bad (down
to a 2% and 15% degradation respectively).

Evidently, this needs improvements in the (likely common)
attcacheoff-enabled multi-column case; as I don't think we can
reasonably commit a 10+% regression. I'll work on that this weekend.

Kind regards,

Matthias van de Meent

Benchmarks were all performed on WSL2 running Debian 10, on an AMD
5950X, with shared_buffers = 15GB (which should fit the dataset three
times), enable_indexscan = off, autovacuum disabled, and parallel
workers disabled on the tables, so that the results should be about as
stable as it gets.