----- Original Message -----
From: "Tom Lane" <tgl@sss.pgh.pa.us>
To: "Qing Zhao" <qzhao@quotefx.net>
Cc: <pgsql-performance@postgresql.org>
Sent: Tuesday, April 06, 2004 1:47 AM
Subject: Re: [PERFORM] possible improvement between G4 and G5
> Qing Zhao <qzhao@quotefx.net> writes:
> > We have got a G5 64-bit processor to replace an old G4 32-bit
> > processor. Given everything else equal, should we see a big
> > improvement on PG's performance?
>
> Nope. Database performance typically depends on disk performance first,
> and RAM size second.
I'm surprised by this thought. I tend to hit CPU bottlenecks more often than
I/O ones. In most applications, db I/O is a combination of buffer misses and
logging, which are both reasonably constrained. RAM size seems to me to be
the best way to improve performance, and then CPU which is needed to perform
the in-memory searching, locking, versioning, and processing, and finally
I/O (this is not the case in small I/O subsystems - if you have less than a
dozen drives, you're easily I/O bound). I/O is often the thing I tune first,
because I can do it in place without buying hardware.
Conceptually, an RDBMS converts slow random I/O into in memory processing
and sequential logging writes. If successful, it should reduce the I/O
overhead.
/Aaron
