robcron <rocron@pssd.com> writes:
> Our IT group took over an app that we have running using postgres and it has
> been on version 8.2.11 since we acquired it. It is time to get current, so
> I have created instances of our production database that mirror exact
> hardware for our existing implementation on version 8.2.11 (running Fedora
> Core 8 - wow I know) and also version 9.1.6 on Fedora 17. I am able to
> mimic the production 8.2 environment exactly without any of the load of
> production and the same for the new 9.1 environment so there is no
> perverting of numbers based on load that I can't control
> Machines are Cloud based images running 4 (dual Core) Processors, with 15GB
> of memory... AMAZON m1.Xlarge boxes - 64 bit OS.
Hm ... Amazon cloud is not exactly known for providing rock-stable
performance environment, but anyway the first thing I would have guessed
at, seeing that the plans are basically the same, was a non-C locale
setting. Another thing to check is whether the new machine has higher
timing overhead --- is the speed difference the same when you just run
the query, rather than EXPLAIN ANALYZE'ing it? (If not,
contrib/pg_test_timing from 9.2 or later might yield useful data.)
regards, tom lane