SUMMARY:DU-Oracle-RAID-Speed

From: Wayne Sweatt <sweatt_at_dps.state.nm.us>
Date: Wed, 06 Jan 1999 15:01:56 -0700

Here were the good responses:

( May I make a suggestion? Those of you who are "experts" in a given area concerning Digital Unix
or any of it's common applications and storage devices, PLEASE! Do not respond to any of my questions if you want to flex your knowledge (or simile thereof) as an attempt to ridicule those who are
asking questions to gain knowledge. I have received responses recently which of this nature, and
in fact, one of the responses for this query was one, including a reference to Oracle Server 7.3.3
as "incredibly old". This buffoon doesn't realize that 7.3.3 is the latest version of Oracle 7 for DU, and as
my Oracle migration expert/DBA remarked: "anyone who would dare run ANYTHING from Oracle
that ends in ".0" must be nuts or spending someone else's money.." Anyway...

Here are the nice responses:


#1

The common knowledge is that RAID-5 is not suited for some types of data files / segment types. Especially you should keep redo-log, rollback & temporary data well *off* RAID-5 as this will seriously hamper performance. Having several raidsets will of course help, but you would be better off keeping redo, rollback & temporary on RAID 0/1/0+1, on a separate controller preferrably. Other datatypes, such as index and table, can live on RAID-5 without much trouble, but this will depend on the type of application - OLTP or OLAP. Intensive OLTP will suffer on RAID-5 because of the checksum calculation scheme, but you'll take the edge off by moving redo etc. off to other types of RAID as I said.
There is a lot of written material on this subject - check out O'Reillys optimizing/tuning guide for a head start on Oracle tuning.
Good luck.

--
Thomas Strandenaes
University of Tromsoe
NORWAY
#2
>How can we organize our disks so that we can optimize disk access
>and reliability?
The RAID tradeoffs can be represented by a triangle where you can chose among performance, availability and cost.  If cost isn't a significant issue, use host based mirroring and striping across as many I/O adapters as you can manage.  If more convience is desired at a slight cost in availability use controller based mirroring and host based striping.  If the I/O load is read-mostly, RAID-5 is a fair price trade-off for better cost.  As the I/O load has more writes, RAID-5 becomes a poor choice. 
If the performance of the page/swap space is going matter, get more memory.
Benchmark the difference between segregating the operating system, log files, table spaces, etc by controller or by striping off them across multiple controllers.
>Can we use the raw disk/Async I/O function with the HW RAID controllers?
The operating system shouldn't know the difference between a simple disk behind a controller and a RAID-5.  It may want to in order to allow for state transition times after member failures, but doesn't need to.
>Is Ultra SCSI the way to go?
You'll be hard pressed to find current generation SCSI disks that don't support Ultra speeds.  For adapters and controllers, take advantage of the extra performance they may offer.
Alan Nabeth
#3
Wayne,
Just a couple of thoughts, as I am pushed for time at the moment, I will be glad to carry on this discussion at length with you next year ;-)
I would stay away from RAW disk, unless you are planning to run Oracle Parallel Server, which requires it...
I would suggest AdVfs filsystems and RAID 5, depending on your controller configuration, you mentioned 3, but fail to describe HSZ KZPSA Storage works involved or local disk, although at 90 GB I have to assume some external disks...
Randall R. Cable
***********************
Wayne Sweatt
Principal Software Analyst
Litton / PRC
505.827.9288
***********************

Received on Wed Jan 06 1999 - 22:02:29 NZDT