Best Disk Setup for Oracle 8.1.7

[Scott Matthesen]

I have 6 18.2 GB drives in an SC-10 disk array. They are setup on 2 separate buses. We would like to mirror these, however, mirroring and striping appear to be one or the other. Just wondering what the best performance approach is to this (still mirroring)? Also, once a best approach is offered, how do I accomplish it?

Thanks in Advance,
Scott Matthesen

[Alan Riggs]

MIrroring and striping are an "either/or" proposition. Striping will almost always give superior write performance, while mirroring can give superior read peformance. since reads tend to more effective in cache (writes always have to hit disk eventually) striping is the better performance bet. BUT, pure striping with no parity leaves you vulnerable to catastrophic data loss in teh event of disk failure, not an option I would choose for a produciton systems.

You can do extent striping, which stripes physical extents across disks and be used with mirroring. The performance benefits of this are significantly reduced, though, since a strip size of 4MB (or larger) reduces the benefit of striping to block writes that span 2 extents. (In general, there might also bee some collateral beenfit in balancing load across spindles for highly active lvols.)

A more significant concern that I see, though, is simply having enough spindles to isolate the required oracle components. If you mirror for data protection, you are down to three spidles. this pretty much requires a dedicated spindle each for archive logs, data files, and index. And you have to double up redo with one of those. That leaves no room for extent striping, even, since to do so you would have to have data/index/archive sharing a physical device, which will almost certainly negate any performance gain from the extent stripe.

[Steven Sim Kok Leong]

Hi,

For my environment, for the index and database Oracle filesystems, I have XP-level RAID 1 mirroring and OS-level striping of these filesystems with a stripe size of 64kb.

One point to note is that it is that because redo-logs are only sequentially accessed, striping the Oracle redo-log filesystems are definitely not encouraged. I have redo-logs mirrored at the Oracle level rather than the hardware or OS level.

As Alan has pointed out, for performance reasons, try to spread your I/O across as many disk spindles possible.

For availability reasons, make sure that your redo-logs and their mirrors reside on different disks across different disk controllers, so as to eliminate any potential single-point of failure (SPOF).

Hope this helps. Regards.

Steven Sim Kok Leong
Brainbench MVP for Unix Admin
http://www.brainbench.com