On Wed, May 7, 2014 at 1:13 PM, Peter Geoghegan <pg@heroku.com> wrote:
> On Wed, May 7, 2014 at 11:04 AM, Josh Berkus <josh@agliodbs.com> wrote:
>> Unfortunately nobody has the time/resources to do the kind of testing
>> required for a new recommendation for shared_buffers.
>
> I meant to suggest that the buffer manager could be improved to the
> point that the old advice becomes obsolete. Right now, it's much
> harder to analyze shared_buffers than it should be, presumably because
> of the problems with the buffer manager. I think that if we could
> formulate better *actionable* advice around what we have right now,
> that would have already happened.
>
> We ought to be realistic about the fact that the current
> recommendations around sizing shared_buffers are nothing more than
> folk wisdom. That's the best we have right now, but that seems quite
> unsatisfactory to me.
I think the stock advice is worse then nothing because it is A. based
on obsolete assumptions and B. doesn't indicate what the tradeoffs are
or what kinds of symptoms adjusting the setting could alleviate. The
documentation should be reduced to things that are known, for example:
*) raising shared buffers does not 'give more memory to postgres for
caching' -- it can only reduce it via double paging
*) are generally somewhat faster than fault to o/s buffers
*) large s_b than working dataset size can be good configuration for
read only loads especially
*) have bad interplay with o/s in some configurations with large settings
*) shared buffers can reduce write i/o in certain workloads
*) interplay with checkpoint
*) have different mechanisms for managing contention than o/s buffers
merlin