Andrew Sullivan <andrew@libertyrms.info> writes:
> On our 16 G Solaris (Ultra SPARC) boxes, we found that using a gig
> for shared buffers was actually worse than a slightly lower amount,
> under Sol 7. The filesystem buffering is too good, so even though
> the system call to the "filesystem" (which turns out to be just to
> memory, because of the buffer) has a measurable cost, the
> implementation of the shared-buffer handling is bad enough that it
> costs _more_ to manage large buffers. Smaller buffers seem not to
> face the difficulty. I haven't a clue why.
Well, part of the reason is that a lot of the data in shared_buffers
has to be effectively duplicated in the kernel's I/O caches, because
it's frequently accessed. So while I'd think the cost of fetching a
page from the buffer pool is lower than from the OS' cache, increasing
the size of the Postgres buffer pool effectively decreases the total
amount of RAM available for caching.
Cheers,
Neil
--
Neil Conway <neilc@samurai.com> || PGP Key ID: DB3C29FC