On 2016-04-12 19:42:11 +0530, Amit Kapila wrote:
> Yes, it seems generally it is a good idea, but not sure if it is a complete
> fix for variation in performance we are seeing when we change shared memory
> structures.
I didn't suspect it would be. More whether it'd be beneficial
performance wise. FWIW, I haven't seen the variations you're observing
on any machine so far.
I think at high concurrency levels we're quite likely to interact with
the exact strategy used for the last-level/l3 cache. pgprocno,
allPgXact, BufferDescs are all arrays with a regular stride that we
access across several numa nodes, at a very high rate. At some point
that makes very likely that cache conflicts occur in set associative
caches.
> Andres suggested me on IM to take performance data on x86 m/c
> by padding PGXACT and the data for the same is as below:
>
> median of 3, 5-min runs
Thanks for running these.
I presume these were *without* pg_prewarming the contents? It'd be
interesting to do the same with prewarming; aligning these structures
should be unrelated to the separate issue of bufferdesc order having a
rather massive performance effect.
Greetings,
Andres Freund