On Wed, Feb 22, 2012 at 10:02 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Wed, Feb 22, 2012 at 2:23 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
>> The industry accepted description for non-sequential access is "random
>> access" whether or not the function that describes the movement is
>> entirely random. To argue otherwise is merely hairsplitting.
>
> I don't think so.
PostgreSQL already uses a parameter called "random_page_cost" to
describe non-sequential access. Perhaps that is wrong and we need a
third parameter?
> For example, a bitmap index scan contrives to speed
> things up by arranging for the table I/O to happen in ascending block
> number order, with skips, rather than in random order, as a plain
> index scan would do, and that seems to be a pretty effective
> technique. Except to the extent that it interferes with the kernel's
> ability to do readahead, it really can't be to read blocks 1, 2, 3, 4,
> and 5 than to read blocks 1, 2, 4, and 5. Not reading block 3 can't
> require more effort than reading it.
By that argument, ANALYZE never could run longer than VACUUM ANALYZE,
so you disagree with Tom and I and you can't explain Pavel's
results....
cost_bitmap_heap_scan() uses "random_page_cost" to evaluate the cost
of accessing blocks, even though the author knew the access was in
ascending block number order. Why was that?
Note that the cost_bitmap_heap_scan() cost can be > than
cost-seqscan() for certain parameter values.
--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services