Very concise and clear answer. That makes complete sense. Thanks. - E
"Andrew Mobbs" <andrewm_at_chiark.greenend.org.uk> wrote in message
news:3yB*8L37o_at_news.chiark.greenend.org.uk...
> Ethan Post <Blah_at_Blah.com> wrote:
> >Running some testing thats simple purpose is to commit as frequently as
> >possible and generate redo. Watching iostat I see redo showing 20%
> >utilization. The volume containing rollback segment maxes out at around
.8
> >MB per second and shows 100% utilization. I have moved the rollback
> >segments around and tried different sizes. What I see is the same volume
> >jump to 15 MB per second at times (i.e. copying an archive log), this can
be
> >sustained for 10-20 seconds or so. How come the very same volumes max
out
> >at .8 MB per second for rollback segments?
>
> Seeks.
>
> If the disk head has to move then that's several ms that you can't do
> I/O on that disk. As you see, when you're doing nearly no seeks (copying
> a single large file), the disk manages 15MB/s. A fast SCSI disk will
> seek in a nominal 4ms. If you're doing 200 seeks/second, the disk will
> be spending 80% of its time moving the head, and only 20% doing I/O.
>
> Redo generation is fairly nice to disks, there's lots of annoying small
> writes, but it's all fairly linear. Rollback segments are written by
> DBWR as and when it feels like, and will be far more scattered.
>
> I prefer to look at sar for disk stats. My rule of thumb for max
> performance is more than 10ms service time is poor, for redo keep it
> below 6ms.
>
> A good write-cache on the disk array can do wonders for the service time,
> just make sure it's battery backed.
>
>
> --
> Andrew Mobbs - http://www.chiark.greenend.org.uk/~andrewm/
Received on Sat Oct 06 2001 - 02:14:31 CDT
