When faced with an obviously IO-bound database, its tempting to deal with the most obvious (symptom) the IO subsystem immediately. Unfortunately, this usually treats symptoms rather the cause, is often expensive and usually futile. Here's my article on a more systematic approach:
Systematic Oracle tuning (pdf)
and here's the powerpoint from the recent NOCOUG conference.
Couldn't have been any more correct.
Although on the other hand just adding few more spindles could be the only issue sometimes. As noticed back in 2006: we tuned and tuned and tuned till the last possible sql was bound. Only to get very little positive impact.
Turns out, all our disks were upto 90% full. Although they are high capacity disks, once "a" disk goes beyond certain capacity, performance starts degrading. In our case, we had no choice but to add more spindles to relieve contention.
Never had any issues after bringing them down to about 60% full.
So open for debate?
Posted by: Sejas | November 23, 2007 at 05:06 PM