On Fri, Jan 11, 2013 at 8:19 AM, Heikki Linnakangas
<hlinnakangas@vmware.com> wrote:
On 10.01.2013 21:48, Matheus de Oliveira wrote:
I have made a small modification to keep the plans, and it got from
33957.768ms to 43782.376ms.
C and SPI_execute_with_args: 58567.708 ms
C and SPI_(prepare/keepplan/execute_plan): 43782.376 ms
C and heap_insert: 33957.768 ms
So switching to prepared plans helped quite a lot, but it's still slower than direct heap_inserts.
Humm... You are right, I misread what it before, sorry. The 33957.768ms was with heap_insert.
One thing that caught my eye:
CREATE OR REPLACE FUNCTION partition_insert_trigger_spi()
RETURNS trigger
LANGUAGE C
VOLATILE STRICT
AS 'partition_insert_trigger_spi','partition_insert_trigger_spi'
SET DateStyle TO 'ISO';
Calling a function with SET options has a fair amount of overhead, to set/restore the GUC on every invocation. That should be avoided in a performance critical function like this.
I (stupidly) used SPI_getvalue [1] and expected it to always return as YYYY-MM-DD, but them I remembered it would do that only with DateStyle=ISO.
But the truth is that I couldn't see any overhead, because the function was without that on my first tests, and after that I saw no difference on the tests. I think I should use SPI_getbinvalue instead, but I don't know how to parse the result to get year and month, any help on that?
[1]
https://github.com/matheusoliveira/pg_partitioning_tests/blob/master/src/spi/partition_insert_trigger_spi.c#L103Regards,
--
Matheus de Oliveira
Analista de Banco de Dados
Dextra Sistemas - MPS.Br nível F!
www.dextra.com.br/postgres
