Re: RAID Level 5 and RedoLog-Files

From: Mark Rosenbaum <mjr_at_netcom.com>
Date: 1997/01/26
Message-ID: <mjrE4MDL9.C8F@netcom.com>#1/1

In article <5c8q3f$ah1_at_nntp.interaccess.com>, Michael Yakus <yakus_at_interaccess.com> wrote:
>"Helmut Hahn" <hhahn_at_tmt.de> wrote:
>
>>We plan to set up an Database for a universal research center on NT4.0 with
>>RAID 5.
 
>>The question is how such a RAID should be configuered. Must RedoLogFiles be
>>holded twice or more, or is it sure enough to reley on the RAID system and
>>setting up only one.
 

>>Or should the RAID be configuered to act like two seperate devices and set
>>up two LogFiles and having the chance to divide up the Tablespaces on these
>>devices for performance issues?
>
>Since RAID-5 suffers a large write performance hit, it would
>probably be better to implement Oracle mirrored redo logs to
>ordinary disks than to use RAID-5 disks since these are updated
>so frequently.
>

There seem to be some confusion on what RAID is so let me try to explain

RAID Redundant Array of Inexpensive/Independent Drives

It started in Berkeley and they still have some paper on the web. It was an attempt to increase performance and reliability for small and inexpensive drives of the time. The RAID levels something like this

RAID 0	Striping, typically block interleaved across drives. Increases
	performance but decreases reliability because any drive that
	fails brings the whole stripe down. Back up is also a pain
	since you now back up N drives as a single unit.

RAID 1	Mirroring, the data is on a pair (or more) of disk. Increases
	reliability, and performance on reads but not writes.

RAID 2&3 Bit interleaved and not used.

RAID 4	RAID 4 is striped with parity. Good reliability, Good read
	performance poor small block write performance.  An example would
	be 4 data and one parity drives. The 4 data drives would be
	like a RAID 0 and the parity drive typically XORs the data
	on the 4 drives together. This is effective for data recovery
	because a XOR b XOR a = b. So, if you have a parity drive
	that has all of the data XORed together and a data drive fails
	just XOR the other drives with the parity drive and the missing
	data is recoverd. This design hits the parity drive VERY hard
	so RAID 5 was developed.

RAID 5	Striped and parity with parity rotating across all
	drives. The problem with RAID 4 & 5 is that the parity drive
	must be updated when data on only a single drive is updated.
	This requires that the old data and parity both be read 
	(2 physical I/Os) then the old data is XORed out and the
	new data is XORed in, and then both data and parity is
	written out (2 more physical I/Os) for a total of 4 
	physical I/Os for a single logical small block write.
	If the data to be written is large enough (fill all data
	drives) then parity can be calculated on the fly and
	performance is N-1 * the performance of a single drive.

Recomendations:

Use RAID 0	for performance, 
Use RAID 1	for critical data, (typicall small amounts)
Use RAID 0 & 1	for high performance critical data(typicall small amounts)
Use RAID 5	for critical historic data that does not change much
		or changes completely (large block writes).

RAID 5 can be very effectively used with partitioning in Data Wharehouses.

Hope this helps

Mark Rosenbaum			Otey-Rosenbaum & Frazier, Inc.
mjr_at_netcom.com			Consultants in High Performance and
(303) 727-7956			Scalable Computing and Applications
POB 1397			ftp://ftp.netcom.com/pub/mj/mjr/resume/

Boulder CO 80306 Received on Sun Jan 26 1997 - 00:00:00 CST