Re: How to keep a table in memory?

Quoth tgl(at)sss(dot)pgh(dot)pa(dot)us (Tom Lane):
> Devrim =?ISO-8859-1?Q?G=DCND=DCZ?= <devrim(at)CommandPrompt(dot)com> writes:
>> So, IMHO, saying "trust your OS + PostgreSQL" is not a 100% perfect
>> approach for the people who are asking to keep their objects on RAM,
>> even though I know that there is nothing we can say right now.
>
> Well, nothing is a 100% solution. But my opinion is that people who
> think they are smarter than an LRU caching algorithm are typically
> mistaken. If the table is all that heavily used, it will stay in memory
> just fine. If it's not sufficiently heavily used to stay in memory
> according to an LRU algorithm, maybe the memory space really should be
> spent on something else.
>
> Now there are certainly cases where a standard caching algorithm falls
> down --- the main one I can think of offhand is where you would like to
> give one class of queries higher priority than another, and so memory
> space should preferentially go to tables that are needed by the first
> class. But if that's your problem, "pin these tables in memory" is
> still an awfully crude solution to the problem. I'd be inclined to
> think instead about a scheme that lets references made by
> higher-priority queries bump buffers' use-counts by more than 1,
> or some other way of making the priority considerations visible to an
> automatic cache management algorithm.

Something I found *really* interesting was that whenever we pushed any
"high traffic" systems onto PostgreSQL 8.1, I kept seeing measurable
performance improvements taking place every day for a week.

Evidently, it took that long for cache to *truly* settle down.

Given that, and given that we've gotten a couple of good steps *more*
sophisticated than mere LRU, I'm fairly willing to go pretty far down
the "trust the shared memory cache" road.

The scenario described certainly warrants doing some benchmarking; it
warrants analyzing the state of the internal buffers over a period of
time to see what is actually in them.

If, after a reasonable period of time (that includes some variations
in system load), a reasonable portion (or perhaps the entirety) of the
Essential Table has consistently resided in buffers, then that should
be pretty decent evidence that cacheing is working the way it should.
--
output = ("cbbrowne" "@" "gmail.com")
http://linuxdatabases.info/info/slony.html
A Plateau is the highest form of flattery.