On 12/28/12 5:11 PM, Heikki Linnakangas wrote:
>
> As it happens, I just spent a lot of time today narrowing down yet
> another report of a regression in 9.2, when running DBT-2:
> http://archives.postgresql.org/pgsql-performance/2012-11/msg00007.php.
> It looks like that is also caused by the plancache changes. DBT-2
> implements the transactions using C functions, which use SPI_execute()
> to run all the queries.
>
> It looks like the regression is caused by extra copying of the parse
> tree and plan trees. Node-copy-related functions like AllocSetAlloc and
> _copy* are high in the profile, They are also high in the 9.1 profile,
> but even more so in 9.2.
>
> I hacked together a quick&dirty patch to reduce the copying of
> single-shot plans, and was able to buy back much of the regression I was
> seeing on DBT-2. Patch attached. But of course, DBT-2 really should be
> preparing the queries once with SPI_prepare, and reusing them thereafter.
I was recently profiling an application that uses a fair amount of
PL/pgSQL with dynamic queries and also noticed AllocSetAlloc high in the
profile. I was getting suspicious now and compared 9.1 and 9.2
performance: 9.2 is consistently about 3% slower. Your patch doesn't
seem to have a measurable effect, but it might be if I ran the test for
longer.