Re: RAID Questions

Gregory P. McLear (gmclear_at_macbo.win.net) wrote:
: Hello Netters.
 

: I have several questions regarding the layout and design of Oracle7
: on a machine using RAID. My company recently purchased an NCR
: 3555 quad Pentium server to act as host for our manufacturing
: system. This machine has five 1GB drives controlled by a
: hardware RAID controller configured for RAID 5. In laying down the
: database I've made several assumptions but before we go production
: on the new machine, I thought it wise to check with the mighty
: Oraphites as to whether their humble disciple was all wet or not.

Ah well. You could have bought a Data General Aviion and saved some money.......
(there goes my bonus)

: Assumption 1:
 

: Forget all about designing for balanced I/O, ie. distributing
: tablespaces across separate drives, because RAID 5 guarantees the
: load is distributed.

Pretty much true. It would be a rare coincidence that the data you query happens to be separated by the blocking factor you used to create your RAID-5 stripe (do you have control over that with the NCR ??). IF you can control the blocking factor (i.e. sector size of each stripe element) bring it into line with your db_block_size, and make sure when you create your filesystems you align them to the stripe in the RAID stripe.

: Assumption 2:
 

: Forget about clustering and hashing for the same reasons. All this
: would do is fatten the dictionary and create additional overhead
: in manipulating dictionary objects.

Absolutely. Also, dont have multiple db_writers, this will only confuse the disk i/o balancing.

: Assuming that designing for portabilty to other hosts is not a
: concern, am I safe to make the two previous assumptions. I
: realize that if the database ever needed to be moved to a host
: that didn't support RAID, redesign for optimal performance would be
: necessary.
 

: Last but not least, a question. The NCR RAID controller supports
: raw slices. On average, with all things being equal, how great of
: a performance increase can one expect to see by using raw slices.
: I need to weigh this against the administrative hassles of using
: raw slices.

Raw disks are a benefit. You dont have UNIX buffering up the same database pages that are in the SGA, so you save on memory usage. Also, you avoid the layer in the kernel that handles page to disk mapping, since Oracle does it for itself in raw disk mode. Allover, you should get an improvement no less than 25% (we do).

: Thanking all in advance,
: Greg McLear
: Mitchel & Scott Machine Co.
 

: sans .signature

Jon Pope					|jpope_at_balrog.europe.dg.com
Data General UK Ltd				|jon_pope_at_DGE.ceo.dg.com
UK RDB Consultant, Oracle Kernel Specialist	|
------------------------------------------------------------------------------