2010/11/12 Tom Lane <tgl@sss.pgh.pa.us>:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Fri, Nov 12, 2010 at 4:15 AM, Cédric Villemain
>> <cedric.villemain.debian@gmail.com> wrote:
>>>> I wondering if we could do something with a formula like 3 *
>>>> amount_of_data_to_read / (3 * amount_of_data_to_read +
>>>> effective_cache_size) = percentage NOT cached. That is, if we're
>>>> reading an amount of data equal to effective_cache_size, we assume 25%
>>>> caching, and plot a smooth curve through that point. In the examples
>>>> above, we would assume that a 150MB read is 87% cached, a 1GB read is
>>>> 50% cached, and a 3GB read is 25% cached.
>
>>> But isn't it already the behavior of effective_cache_size usage ?
>
>> No.
>
> I think his point is that we already have a proven formula
> (Mackert-Lohmann) and shouldn't be inventing a new one out of thin air.
> The problem is to figure out what numbers to apply the M-L formula to.
>
> I've been thinking that we ought to try to use it in the context of the
> query as a whole rather than for individual table scans; the current
> usage already has some of that flavor but we haven't taken it to the
> logical conclusion.
>
>> The ideal of trying to know what is actually in cache strikes me as an
>> almost certain non-starter.
>
> Agreed on that point. Plan stability would go out the window.
Point is not to now the current cache, but like for ANALYZE on a
regular basis (probably something around number of page read/hit) run
a cache_analyze which report stats like ANALYZE do, and may be
adjusted per table like auto_analyze is.
--
Cédric Villemain 2ndQuadrant
http://2ndQuadrant.fr/ PostgreSQL : Expertise, Formation et Support