Re: Redo Log Size

"Johan Lorier" <johan.lorier_at_nl.abb.com> wrote in message news:c4ce07bf.0108290529.e8e7c2a_at_posting.google.com...
> I wonder why many small logswitches degrade performance? When there
> are more checkpoints, let's say every 15 minutes, the diskload will be
> distributed more evenly over the day in my opinion, or am I missing
> something?

Bear in mind that when you switch out of a log group, DBWR only flushes dirty buffers covered by the log file being switched away from. Yet CKPT has to update the headers of *all* datafiles, and all controlfiles.

Lots of small checkpoints are therefore grossly inefficient from the point of view of CKPT.

Of course, your point has a degree of validity: the whole idea of fast_start_io_target is to get DBWR constantly dribbling dirtied blocks back down to disk -practically doing constant checkpointing.

But it remains true that lots of small (or constant) checkpoints will result in even, but lower, performance. Better that, I suppose, than relatively high performance forever interrupted by mammoth checkpoints. But better still (in my opinion, anyway) is high performance uninterrupted by any checkpoints (at least until you manually force one at a time when no-one will care about the massive hiatus).

Hence, I suppose, Oracle's hint that you'd only use fast_start_io_target when you absolutely, positively, have to bound your Instance Recovery time for the sake of an SLA.

Regards
HJR Received on Wed Aug 29 2001 - 14:49:23 CDT