Re: Why?

Michael Austin <maustin_at_firstdbasource.com> wrote in message news:<ZV_Dc.4564$145.4190_at_newssvr23.news.prodigy.com>...
> Hans Forbrich wrote:
> > Alexander Skwar wrote:
> >
> >
> >>>No database should run RAID5.
> >>
> >>Why?
> >
> >
> > See http://www.baarf.com/ for a genreal discussion
> >
> > /Hans
>
>
> Do you know why they call if BAARF? Because after you read it you
> will... barf. I have still not found anyone who has been able to -
> through os-level stats - quantify this supposed "write" penalty on real
> computers using SAN technology and latest storage technology. If you
> are talking about that (AMD/INTEL) server with n-internal IDE's then
> maybe. You can show all the math "proving" a point, but as you and I
> well know, what is theoretical and what is real are quite different from
> "server" to "server". 90% of all database applications out there would
> never see using RAID5 as a problem or bottleneck.
>
> Michael Austin.

your "90%" figure - is that in your opinion due to the fact that the app code needs tuning/re-writing, that memory is cheap enough that most data will be found in the buffer cache, or due to the symptom that the (major) bottleneck resides in the middle tier, leaving the backend underworked?

well, its quite possible that the FCHBA(s) might in fact be the bottleneck in the case of a SAN, particularly if mosly small IOs are taking place.

is the app code performing full table scans, index fast full scans with a db_block_size * dbfmbrc = 1 MB, or is it performing controlfile writes, redo log writes, index range scans where the typical IO size is 512 to 8192 bytes?

yes, upgrading from 1 Gb to 2 Gb might provide the headroom, or using multiple adapters so as to not have a single point of failure, provided that multipathing os supported would provide more (aggregate) bandwidth.

Anjo Kolk has presented on this exact subject, including 2 conferences this year, RMOUG and Hotsos Perf Symposium in Dallas last March. He wrote the oraperf site and is working for Veritas. Look for some of his papers at the usual sites, hotsos, orapub and possibly Veritas.

Pd Received on Tue Jun 29 2004 - 01:46:50 CDT