Hi,
On 2020-01-16 18:01:53 +0100, Tomas Vondra wrote:
> Plus we don't have all that many slabs (like 2)
FWIW, I have a local patch that adds additional ones, for the relcache
and catcache, that's how I noticed the leak. Because a test pgbench
absolutely *tanked* in performance.
Just for giggles. With leak:
pgbench -n -M prepared -P1 -c20 -j20 -T6000 -S
progress: 1.0 s, 81689.4 tps, lat 0.242 ms stddev 0.087
progress: 2.0 s, 51228.5 tps, lat 0.390 ms stddev 0.107
progress: 3.0 s, 42297.4 tps, lat 0.473 ms stddev 0.141
progress: 4.0 s, 34885.9 tps, lat 0.573 ms stddev 0.171
progress: 5.0 s, 31211.2 tps, lat 0.640 ms stddev 0.182
progress: 6.0 s, 27307.9 tps, lat 0.732 ms stddev 0.216
progress: 7.0 s, 25698.9 tps, lat 0.778 ms stddev 0.228
without:
pgbench -n -M prepared -P1 -c20 -j20 -T6000 -S
progress: 1.0 s, 144119.1 tps, lat 0.137 ms stddev 0.047
progress: 2.0 s, 148092.8 tps, lat 0.135 ms stddev 0.039
progress: 3.0 s, 148757.0 tps, lat 0.134 ms stddev 0.032
progress: 4.0 s, 148553.7 tps, lat 0.134 ms stddev 0.038
I do find the size of the impact quite impressive. It's all due to the
TopMemoryContext's AllocSetCheck() taking longer and longer.
> and it's only really used in debug builds anyway. So I'm not all that
> woried about this wasting a couple extra kB of memory.
IDK, making memory usage look different makes optimizing it harder. Not
a hard rule, obviously, but ...
Greetings,
Andres Freund
