On Wed, May 4, 2016 at 1:25 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> If somebody had even hinted that such a problem might exist, Kevin
> probably would have fixed it before commit, but nobody did. As soon
> as it was raised, Kevin started working on it. That's about all you
> can ask, I think.
Right; I have not been ignoring the issue -- but I prioritized it
below fixing correctness issues and performance issues when the
feature is off. Since there are no known issues in either of those
areas remaining once I push the patch I posted earlier today, I'm
taking a close look at the three proposals from three different
people about ways to address it (along with any other ideas that
come to mind while working through those). Fortunately, the access
problems to the EDB big NUMA machines have now been solved (by
tweaking firewall settings), so I should have some sort of
meaningful benchmarks of those three alternatives and anything else
the comes to mind within a few days. (Until now I have been asking
others to do some runs, which doesn't gather the data nearly as
quickly as actually having access.)
Amit's proof-of-concept patch is very small and safe and yielded a
3x to 4x performance improvement with the old_snapshot_threshold =
1 on a big NUMA machine with concurrency in the 32 to 64 client
range. I don't know whether we can do better before beta1, but it
is something. I'll be happy to try any other suggestions.
--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company