On Wed, Apr 20, 2016 at 8:50 AM, Kevin Grittner <kgrittn@gmail.com> wrote:
> In case anyone notices some code left at the bottom of bufmgr.h
> related to inline functions, that was left on purpose, because I am
> pretty sure that the fix for the performance regression observed
> when the "snapshot too old" feature is disabled will involve making
> at least part of TestForOldSnapshot() an inline function -- so it
> seemed dumb to rip that out now only to put it back again right
> away.
I pushed something along those lines. I didn't want to inline the
whole function because IsCatalogRelation() and
RelationIsAccessibleInLogicalDecoding() seemed kinda big to inline
and require rel.h to be included; so bringing them into bufmgr.h
would have spread that around too far. Putting the quick tests in
an inline function which calls a non-inlined _impl function seemed
like the best compromise.
My connectivity problems to our big NUMA machines have not yet
been resolved, so I didn't have a better test case for this than
200 read-only clients at saturation on my single-socket i7, which
was only a 2.2% to 2.3% regression -- so I encourage anyone who was
able to create something more significant with
old_snapshot_threshold = -1 to try with the latest and report the
impact for your environment. I'm not sure whether any more is
needed here.
--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company