On Wed, May 7, 2014 at 11:50 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>> Doesn't match my experience. Even with the current buffer manager
>> there's usually enough locality to keep important pages in s_b for a
>> meaningful time. I *have* seen workloads that should have fit into
>> memory not fit because of double buffering.
>
> Same here.
I think that it depends on whether or not you're thinking about the
worst case. Most people are not going to be in the category you
describe here. Plenty of people in the Postgres community run with
very large shared_buffers settings, on non i/o bound workloads, and
report good results - often massive, quickly apparent improvements.
I'm mostly concerned with obsoleting the 8GB hard ceiling rule here.
It probably doesn't matter whether and by how much one factor is worse
than the other, though. I found the section "5.2 Temporal Control:
Buffering" in the following paper, that speaks about the subject quite
interesting: http://db.cs.berkeley.edu/papers/fntdb07-architecture.pdf
--
Peter Geoghegan