At 12:38 PM 4/13/2007, Steve wrote:
>>Really?
>>
>>Wow!
>>
>>Common wisdom in the past has been that values above a couple of hundred
>>MB will degrade performance. Have you done any benchmarks on 8.2.x that
>>show that you get an improvement from this, or did you just take the
>>"too much of a good thing is wonderful" approach?
>
> Not to be rude, but there's more common wisdom on this
> particular subject than anything else in postgres I'd say ;) I
> think I recently read someone else on this list who's
> laundry-listed the recommended memory values that are out there
> these days and pretty much it ranges from what you've just said to
> "half of system memory".
>
> I've tried many memory layouts, and in my own experience
> with this huge DB, more -does- appear to be better but marginally
> so; more memory alone won't fix a speed problem. It may be a
> function of how much reading/writing is done to the DB and if fsync
> is used or not if that makes any sense :) Seems there's no "silver
> bullet" to the shared_memory question. Or if there is, nobody can
> agree on it ;)
One of the reasons for the wide variance in suggested values for pg
memory use is that pg 7.x and pg 8.x are =very= different beasts.
If you break the advice into pg 7.x and pg 8.x categories, you find
that there is far less variation in the suggestions.
Bottom line: pg 7.x could not take advantage of larger sums of memory
anywhere near as well as pg 8.x can.
Cheers,
Ron