Gregory Smith <gregsmithpgsql@gmail.com> writes:
> Pushing SELECT statements at socket speeds with prepared statements is a
> synthetic benchmark that normally demos big pgbench numbers. My benchmark
> farm moved to Ubuntu 23.04/kernel 6.2.0-20 last month, and that test is
> badly broken on the system PG15 at larger core counts, with as much as an
> 85% drop from expectations.
> ... I could use some confirmation of where this happens from
> other tester's hardware and Linux kernels.
FWIW, I can't reproduce any such effect with PG HEAD on RHEL 8.8
(4.18.0 kernel) or Fedora 37 (6.2.14 kernel). Admittedly this is
with less-beefy hardware than you're using, but your graphs say
this should be obvious with even a dozen clients, and I don't
see that. I'm getting results like
$ pgbench -S -T 10 -c 16 -j 16 -M prepared pgbench
tps = 472503.628370 (without initial connection time)
$ pgbench -S -T 10 -c 16 -j 16 pgbench
tps = 297844.301319 (without initial connection time)
which is about what I'd expect.
Could it be that the breakage is Ubuntu-specific? Seems unlikely,
but ...
regards, tom lane