On Thu, Jun 8, 2023 at 3:09 PM Gregory Smith <gregsmithpg...@gmail.com> wrote: > 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. > Attached are full scaling graphs for all 4 combinations on this AMD 32 thread > 5950X, and an Intel i5-13600K with 20 threads and similar impact. The > regular, unprepared sockets peak speeds took a solid hit in PG15 from this > issue too. I could use some confirmation of where this happens from other > tester's hardware and Linux kernels.
Since it doesn't look like you included results on pre-23x Ubuntu, I thought I would reply with my own results using your example. I also have a 32 thread AMD 5950X but am on Ubuntu 22.10 (kernel 5.19). I did not see the regression you mention. HEAD pgbench -S -T 10 -c 32 -j 32 -M prepared -p 5432 pgbench tps = 837819.220854 (without initial connection time) pgbench -S -T 10 -c 32 -j 32 -M simple -p 5432 pgbench tps = 576845.930845 (without initial connection time) REL_15_STABLE pgbench -S -T 10 -c 32 -j 32 -M prepared -p 5432 pgbench tps = 794380.991666 (without initial connection time) pgbench -S -T 10 -c 32 -j 32 -M simple -p 5432 pgbench tps = 534358.379838 (without initial connection time) - Melanie
