On 05.11.2012 16:32, Claudio Freire wrote:
> On Sun, Nov 4, 2012 at 7:23 PM, Dong Ye<yed@vmware.com> wrote:
>>> You are using prepared statements, this makes me think that this
>>> regression might be due to support for parameter specific plans for
>>> prepared statements. [1] Can you run the test on both versions without
>>> prepared statements and see if the regressions remains.
>>
>> Without prepare statement, we got 48837.33 avg notpm with 9.1.6 and 43264.54 avg notpm with 9.2.1.
>> notps over time shows the slowdown of 9.2.1 is evident during the entire course of the run.
>> Their profiles are posted on http://pgsql.privatepaste.com/b770f72967 (9.1.6) and
http://pgsql.privatepaste.com/6fa8b7f174(9.2.1).
>
> You know... it does look as if 9.2.1 is generating a lot more pressure
> into the memory allocator (AllocSetAlloc notably higher).
Did you check the access plans of the queries? 9.2 planner might choose
a slightly worse plan. Or perhaps index-only scans are hurting
performance with the DBT-2 queries.
- Heikki
