On Tue, 15 Jan 2019 at 13:17, Tom Lane <t...@sss.pgh.pa.us> wrote: > > David Rowley <david.row...@2ndquadrant.com> writes: > > I also did a quick benchmark of v6 and found the slowdown to be > > smaller after the change made in build_simple_rel() > > Thanks for confirming. I was not very sure that was worth the extra > few bytes of code space, but if you see a difference too, then it's > probably worthwhile.
It occurred to me that a common case where you'll hit the new code is INSERT INTO ... VALUES. I thought I'd better test this, so I carefully designed the following table so it would have as little INSERT overhead as possible. create table t(); With fsync=off and a truncate between each pgbench run. insert.sql = insert into t default values; Unpatched: $ pgbench -n -f insert.sql -T 60 postgres tps = 27986.757396 (excluding connections establishing) tps = 28220.905728 (excluding connections establishing) tps = 28234.331176 (excluding connections establishing) tps = 28254.392421 (excluding connections establishing) tps = 28691.946948 (excluding connections establishing) Patched: tps = 28426.183388 (excluding connections establishing) tps = 28464.517261 (excluding connections establishing) tps = 28505.178616 (excluding connections establishing) tps = 28414.275662 (excluding connections establishing) tps = 28648.103349 (excluding connections establishing) The patch seems to average out slightly faster on those runs, but the variance is around the noise level. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
