On Thu, Dec 8, 2016 at 9:07 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> You could do that, but first I would code up the simplest, cleanest
> algorithm you can think of and see if it even shows up in a 'perf'
> profile. Microbenchmarking is probably overkill here unless a problem
> is visible on macrobenchmarks.
This is what I would go for! The current code is doing a simple thing:
select the Nth element using qsort() after scanning each WAL sender's
values. And I think that Sawada-san got it right. Even running on my
laptop a pgbench run with 10 sync standbys using a data set that fits
into memory, SyncRepGetOldestSyncRecPtr gets at most 0.04% of overhead
using perf top on a non-assert, non-debug build. Hash tables and
allocations get a far larger share. Using the patch,
SyncRepGetSyncRecPtr is at the same level with a quorum set of 10
nodes. Let's kick the ball for now. An extra patch could make things
better later on if that's worth it.
--
Michael