On Wed, 1 Feb 2006, Alfred Perlstein wrote:

> Database caching algorithms will offer a signifigant advantage over
> a simple filesystem LRU.

You'd best go study what algorithms are being used in both locations. I
don't believe any of the OSes under discussion use a "simple filesystem
LRU" any more, and PostgreSQL's cacheing algorithm is perhaps more naive
than you think it is.

(Oh, but that does remind me, the query optimizer is not at all naive,
so you want to make sure you set effective_cache_size appropriately. I'm
not sure how many people know about that option.)

> Additionally, a larger shared segment reduces the amount of copying
> required.

True. But compare the cost of a memory copy and the cost of a disk read
or write. Now how much do you need to save?

> Additionally, a larger shared segment particularly when non-pageable,
> will protect the database's cache against errant jobs that blow
> through the kernel's cache.  Imagine the effect of untarring a large
> file suddenly.

It depends on the OS and what algorithms its using for its disk
buffering. On Solaris, for example, that effect should be very low.

Anyway, this has all been discussed to death on the PostgreSQL lists. I
suggest you check out the discussions in the FAQs, archives, and so on
if you're interested.

cjs
-- 
Curt Sampson            <cjs@cynic.net>             +81 90 7737 2974
   The power of accurate observation is commonly called cynicism
   by those who have not got it.    --George Bernard Shaw