Re: checkpoints/bgwriter tuning verification

On Thu, Oct 29, 2009 at 12:46 PM, Vick Khera <vivek(at)khera(dot)org> wrote:
> On my primary DB I'm observing random slowness which just doesn't make
> sense to me.  The I/O system can easily do 40MB/sec writes, but I'm
> only seeing a sustained 5MB/sec, even as the application is stalling
> waiting on the DB.

Just one point on top of everything else you'll hear. 40 MB/sec
sequential throughput does not equal 40MB/sec random PLUS checkpoint
throughput. Random access is gonna lower that 40MB/sec way down real
fast.

First step to speed things up is putting pg_xlog on its own disk(s).
Since pg_xlog is mostly sequentially access, it's much faster when
it's not competing with the rest of the db. THEN if you need faster
disks you can buy them and throw them at path/data/base/. However,
often just splitting things out like this will make a world of
difference on a write heavy database.