SLRU_BLCKSZ

Simon Riggs <simon@2ndquadrant.com> writes:

Code changes to separate out the symbol for SLRU_BLCKSZ from standard
BLCKSZ. Same reasons as for doing this for XLOG_BLCKSZ.

Have you done any benchmarking that shows this is a good idea?

My own thoughts on the subject have been along the line that pages for
different slru logs possibly need to be different sizes, so that eg
we could have more than 2K subtransactions per page. I'm not sure this
is worthwhile (is there any place in slru.c that depends on a fixed page
size for efficiency?) but it needs to be considered before we make
purely cosmetic changes in the area.

regards, tom lane

On Wed, 2006-05-17 at 10:05 -0400, Tom Lane wrote:

Simon Riggs <simon@2ndquadrant.com> writes:

Code changes to separate out the symbol for SLRU_BLCKSZ from standard
BLCKSZ. Same reasons as for doing this for XLOG_BLCKSZ.

Have you done any benchmarking that shows this is a good idea?

Some, with more coming.

My own thoughts on the subject have been along the line that pages for
different slru logs possibly need to be different sizes, so that eg
we could have more than 2K subtransactions per page.

That's what I thought at first.

I'm not sure this
is worthwhile (is there any place in slru.c that depends on a fixed page
size for efficiency?) but it needs to be considered before we make
purely cosmetic changes in the area.

You raised the possible impact of clog extension doing WAL logging, so
doing that less often seems sensible.

Increasing all of them could have benefit, for different reasons - so
having different block sizes for each seems like overkill - and harder
to record in the control file and utilities.

--
Simon Riggs
EnterpriseDB http://www.enterprisedb.com