I'm after a little bit of advice on the shared_buffers setting (I have read the various docs on/linked from the performance tuning wiki page, some very helpful stuff there so thanks to those people).
I am setting up a 64bit Linux server running Postgresql 8.3, the server has 64gigs of memory and Postgres is the only major application running on it. (This server is to go alongside some existing 8.3 servers, we will look at 8.4/9 migration later)
I'm basically wondering how the postgresql cache (ie shared_buffers) and the OS page_cache interact. The general advice seems to be to assign 1/4 of RAM to shared buffers.
I don't have a good knowledge of the internals but I'm wondering if this will effectively mean that roughly the same amount of RAM being used for the OS page cache will be used for redundantly caching something the Postgres is caching as well?
IE when Postgres reads something from disk it will go into both the OS page cache and the Postgresql shared_buffers and the OS page cache copy is unlikely to be useful for anything.
If that is the case what are the downsides to having less overlap between the caches, IE heavily favouring one or the other, such as allocating shared_buffers to a much larger percentage (such as 90-95% of expected 'free' memory).
Pg apparently does not have an option of using direct IO with reads which some other databases do (the O_DIRECT mode). Therefore, double-buffering with read operations seems unavoidable. Counterintuitively, it may be a good idea to just rely on OS buffering and keep shared_buffers rather small, say, 512MB.