On Wed, May 18, 2016 at 6:25 AM, Andres Freund <andres@anarazel.de> wrote:
> currently we IIRC use linearly sorted datums for the search in
> individual btree nodes. Not surprisingly that's often one of the
> dominant entries in profiles. We could probably improve upon that by
> using an order more optimized for efficient binary search.
Did you ever try running a pgbench SELECT benchmark, having modified
things such that all PKs are on columns that are not of type
int4/int8, but rather are of type numeric? It's an interesting
experiment, that I've been meaning to re-run on a big box.
Obviously this will be slower than an equivalent plain pgbench SELECT,
but the difference may be smaller than you expect.
--
Peter Geoghegan