On Wed, 8 Apr 2009, Tom Lane wrote:
> If the kernel can't read-ahead a seqscan by itself, it's unlikely to
> be smart enough to be helped by posix_fadvise ... or at least so I
> would think.
There's some interesting comments on this subject (and about what fadvise
DONTNEED does) in the RRD research paper about managing their buffer
cache:
http://www.usenix.org/event/lisa07/tech/full_papers/plonka/plonka_html/index.html
They suggest the Linux read-ahead is pretty aggressive by default, which
might explain why I wasn't able to replicate any speed-up with the
seqeuential scan patch on my system. (The original submission showed a
significant speedup on Linux, but was from what sounded like a somewhat
broken kernel--known buggy controller driver I seem to recall)
I suspect we may need to find a platform where the default OS readahead is
a slacker, *and* that pays attention to POSIX_FADV_SEQUENTIAL, in order to
show any improvement from that patch.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD