Raid configuration for database

[Kevin Brown]

From Documentation I have gotten from Sun, if you expect the system to be doing
more than 20% writes then they wouldn't recommend raid5.

> I think you need to put more time into investigating why you have had such
> trouble with RAID 5. I've run several systems with RAID 5 configurations,
> all were highly reliable and blazing fast. Add one or two hot spare drives,
> and 100% reliability is nearly guaranteed.

> Symbio-Tech makes a telephone switch for pre-paid calling.
> 
> A critical component in the product is an Oracle database.  The database is
> heavily biased toward OLTP.  It has to execute transactions very fast, and
> it can never go down.  No customer is large enough to effectively use
> tape-backup.  For the most part they cannot be relied upon to replace hard
> drives on a regular schedule.
> 
> We have had terrible experiences with RAID-5.  It is slow and has a
> distressing tendency to die catastrophically.
> 
> We have been using RAID-1.
> 
> We are currently building what for us is a large system.  The Oracle
> database is getting 7 drives.  If we follow tradition there will be three
> logical drives, each with two physical drives mirrored to each other.  Drive
> number 7 is hot backup.
> 
> Lets forget the hot backup for this problem.  (Also ignore the fact that we
> should use RAID-1+0.)
> 
> With the default RAID-1 configuration we have 6 drives: scd0, 1, 2, 3, 4,
> and 5.  We pair them getting logical drives mirror01, 23, and 45.
> 
> Alternatively we could partition each drive.  If we use every possible
> combination, ignoring order (that is mirror01 == mirror10), and never
> mapping a drive to itself, then each drive has 6 partitions.  Furthermore,
> each partition is shared by two disks. There are 15 combinations.
> 
> 01 | 12 | 23 | 34 | 45
> 02 | 13 | 24 | 35
> 03 | 14 | 25
> 04 | 15 |
> 05 |
> 
> -------------------------------
> 
> Assume that the simple RAID-1 of 01 | 23 | 56 is the null hypothesis.
> 
> 1) Is the 15 partition model more or less reliable than the null case?
> 
> 2) Is the 15 partition model generally faster or slower than the null case?
> 
> 3) More specifically, is the 15 partition model likely to generate more or
> less disk contention when used by a database than the null case?