Re: Raw device type for data files

"Matthew M. Lih" <matt.lih_at_trw.com> writes:

>I attended a seminar by Oracle on performance tuning, where they
>said that there are few circumstances where you would want to use
>raw devices. Apparently nowadays the speed increase over the file
>system is 10%-15% (the figure they gave me). Weighed against the
>difficulty of managing Oracle on raw devices, it sounded like
>you would have to have your back against the wall before going to
>raw devices.

Piffle! (polite for b*sh*t)

I've been working for one of the leading providers of database servers for over 5 years and have been using Oracle for 10. Assuming you are using a volume manager such as IBM's LVM, Sequent's SVM, or any of the various ports of the Veritas VxVM product almost every argument favors raw databases. I have a two page table comparing raw and cooked and the only point where cooked databases have an edge is in the number of backup options available. But even there raw databases have an edge on the restore side of the equation since copying database files from tape to raw volumes will complete sooner than restoring to filesystems.

If the customer does not have a volume manager then I tend to lean towards cooked databases.

I've been a consultant for the past couple of years specializing in system archecture design and performance evaluation and tuning. I often visit customers after Oracle consultants have been there. I have come to the conclusion that Oracle is recommending cooked databases because they require less involvement by the system administrator. In my mind this is actually an argument for raw as it increases the, often dismayingly low, amount of communication between the DBA and SysAdmin.

And regarding the performance differential...I've actually seen filesystem based databases outperform raw databases. But in each case it was due to pathological conditions (such as the use of rpt on a v7 database and creating indexes with grotesquely low selectivity). On the platforms I have tuning experience on the difference can range from 5% to 50%. But in my mind disk throughput increases are one of the least important reasons for chosing raw databases.

-- 
Kurtis D. Rader, Sr. Technical Consultant       voice: 503/578-3714
Sequent Computer Systems                          fax: 503/578-5453
15450 SW Koll Parkway, M/S WIL1-541              UUCP: ...uunet!sequent!krader
Beaverton, OR  97006-6063                    Internet: krader_at_sequent.com