Thread: Bitmap index scan preread using posix_fadvise (Was: There's random access and then there's random access)

Here's the WIP patch for doing prereading when doing bitmap index scans.

I was performance testing it as I was developing it here:

http://archives.postgresql.org/pgsql-hackers/2007-12/msg00395.php

Note that this only kicks in for bitmap index scans which are kind of tricky
to generate. I used the attached function to generate them in the post above.

Also note I wouldn't expect to see much benefit unless you're on a raid array,
even a small one. But if you are on a raid array then the benefit should be
immediately obvious or else posix_fadvise just isn't working for you. I would
be interested in hearing on which OSes it does or doesn't work.

*If* this is the approach we want to take rather than restructure the buffer
manager to avoid taking two trips by marking the buffer i/o-in-progress and
saving the pinned buffer in the bitmap heap scan then this is more or less in
final form. Aside from some autoconf tests and the documentation for the GUC I
think it's all in there.



--
  Gregory Stark
  EnterpriseDB          http://www.enterprisedb.com
  Ask me about EnterpriseDB's RemoteDBA services!

"Gregory Stark" <stark@enterprisedb.com> writes:

> Aside from some autoconf tests and the documentation for the GUC I
> think it's all in there.

I'm sorry, it seems I accidentally grabbed an old tree to generate this patch.
I'll send along a better more recent version. Argh.

--
  Gregory Stark
  EnterpriseDB          http://www.enterprisedb.com
  Ask me about EnterpriseDB's PostGIS support!

Attached is the correct patch, sorry for the confusion.

If anyone's interested in testing it you can do so without the randomarray.c
file by using Pavel Stehule's solution for generating arrays:

postgres=# create table h as (select (random()*1000000)::integer as r, repeat('x',512)::text as t from
generate_series(1,1000000));
SELECT

postgres=# create index hri on h(r);
CREATE INDEX

postgres=# analyze h;
ANALYZE

postgres=# \timing
Timing is on.

postgres=# set preread_pages = 0;   explain analyze select * from h where r = any (array(select
(1+random()*1000000)::integerfrom generate_series(1,100))); 
postgres=# set preread_pages = 100; explain analyze select * from h where r = any (array(select
(1+random()*1000000)::integerfrom generate_series(1,100))); 
postgres=# set preread_pages = 0;   explain analyze select * from h where r = any (array(select
(1+random()*1000000)::integerfrom generate_series(1,100))); 
postgres=# set preread_pages = 100; explain analyze select * from h where r = any (array(select
(1+random()*1000000)::integerfrom generate_series(1,100))); 
postgres=# set preread_pages = 0;   explain analyze select * from h where r = any (array(select
(1+random()*1000000)::integerfrom generate_series(1,100))); 



--
  Gregory Stark
  EnterpriseDB          http://www.enterprisedb.com
  Ask me about EnterpriseDB's 24x7 Postgres support!

Nice!

- Luke


On 1/30/08 9:22 AM, "Gregory Stark" <stark@enterprisedb.com> wrote:

>
> Here's the WIP patch for doing prereading when doing bitmap index scans.
>
> I was performance testing it as I was developing it here:
>
> http://archives.postgresql.org/pgsql-hackers/2007-12/msg00395.php
>
> Note that this only kicks in for bitmap index scans which are kind of tricky
> to generate. I used the attached function to generate them in the post above.
>
> Also note I wouldn't expect to see much benefit unless you're on a raid array,
> even a small one. But if you are on a raid array then the benefit should be
> immediately obvious or else posix_fadvise just isn't working for you. I would
> be interested in hearing on which OSes it does or doesn't work.
>
> *If* this is the approach we want to take rather than restructure the buffer
> manager to avoid taking two trips by marking the buffer i/o-in-progress and
> saving the pinned buffer in the bitmap heap scan then this is more or less in
> final form. Aside from some autoconf tests and the documentation for the GUC I
> think it's all in there.
>
>