"Anand Kumar, Karthik" <Karthik.AnandKumar@classmates.com> wrote:
> We finally made some headway on this - we noticed messages like
> the below
> in /var/log/messages whenever the issue happened:
>
> Mar 26 07:39:58 site-db01b kernel: postmaster: page allocation failure.
> Anyone have any idea why memory was so fragmented, and what
> causes memory to be defragged? Is it something postgres does? Are
> there any kernel specific settings that control it?
While I agree with other replies that you would benefit from
transaction-based connection pooling, there may be another issue at
play here. I can't be sure from evidence so far, but this might be
related to something I've been looking at related to NUMA memory
and how the OS buffers and PostgreSQL shared_buffers interact with
it. Most of the available benefit can be realized without any
change to the PostgreSQL code by using the cpuset features of the
OS. If you want to investigate this, the first thing would be to
get a look at the shape of things. Please post the output of this:
numactl --hardware
The above just reports on the hardware -- it doesn't change
anything.
--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company