On Sat, 3 Oct 2009, Gerhard Wiesinger wrote:
I wouldn't read 128k blocks all the time. I would do the following:
When e.g. B0, B127, B256 should be read I would read in 8k random block I/O.
When B1, B2, B3, B4, B5, B7, B8, B9, B10 are needed I would make 2 requests
with the largest possible blocksize:
1.) B1-B5: 5*8k=40k
2.) B7-B10: 4*8k=32k
I see what you mean now. This is impossible in the current buffer manager
implementation because blocks are requested one at a time, and there are
few situations where you can predict which are going to be wanted next.
The hash index and sequential scan are two that were possible to predict
in that way.
The fadvise patches already committed didn't change the way blocks were
read in, they just used knowledge about what was coming next to advise the
OS. That's quite a bit different from actually asking for things in
larger chunks and only using what you need.
Implementing larger chunking reads or similar asynchronous batch I/O is a
big project, because you'd have to rearchitect the whole way buffers are
managed in the database to do it right. Greg Stark's earliest proof of
concept prototype for async I/O included a Solaris implementation that
used the AIO library. It wasn't feasible to actually use that underlying
implemention in the database in the end though, because the model AIO uses
expects you'll fire off a bunch of I/O and then retrieve blocks as they
come in. That's really not easy to align with the model for how blocks
are read into shared_buffers right now. He had some ideas for that and
I've thought briefly about the problem, but it would be a major overhaul
to some scary to touch database internals to pull off.
Given that the OS and/or RAID implementations tend to do what we want in a
lot of these cases, where smarter/chunkier read-ahead is what we you need,
the payback on accelerating those cases hasn't been perceived as that
great. There is a major win for the hash index reads, which Solaris
systems can't take advantage of, so somebody who uses those heavily on
that OS might be motivated enough produce improvements for that use case.
Once the buffer cache at large understood how to handle batching async
reads, Solaris AIO would be possible, fancier stuff with Linux AIO would
be possible, and the type of chunking reads you're suggesting would be
too. But none of that is happening without some major rearchitecting
first.
Unfortunately there aren't that many people with the right knowledge and
motivation to start tinkering around with the buffer cache internals to
the extent that would be required to do better here, and pretty much of
them I'm aware of are hacking on projects with a much clearer payback
instead.
--
* Greg Smith gsmith@xxxxxxxxxxxxx http://www.gregsmith.com Baltimore, MD
--
Sent via pgsql-general mailing list (pgsql-general@xxxxxxxxxxxxxx)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-general
