Re: Oracle disk subsystem - 10KRPM vs. 15KRPM drive question

here are a few pointers. Not exhaustive by any means, but a starting point?

>
>- 8 or 10 pairs of Ultra160 drives with each pair peer connected to a
>separate SCSI channel for the main stripe sets (2 stripe sets, 64 and
>256Kb as per Jonathan Lewis's book)

sounds good. You prolly don't need as many, but it all depends on load pattern. think of it in terms of where is the load gonna be distributed to. instead of just how many MBs are stored in each pair.

>
>- A pair of mirrored Ultra160 drives for online redo, on their own
>SCSI channel

interesting point about these: even with Ultra160, the outer surface of a disk is noticeably faster than the inner in terms of aggregate physical access time. That means: partition the disks into at least 3 areas and use the outer one (smaller cylinder numbers) as your redo. You'll get max burst and aggregate write speed that way.

You're after max burst write speed. Make sure (if you use file systems) that the file system block size, db block size and write burst max size used by redo writer and any file system writer(s) does not exceed the size of the disk drive cache: any overreach may result in a drastic reduction in burst write speed.

Better yet: consider using one of those new ultra-fast memory "disk drives" for your redo logs. They are safe and redundant, when properly configured. And nothing beats their write speed.

>
>- Another pair of large, mirrored Ultra160 drives for archive logs,
>again on their separate SCSI channel

hmmmm, dunno. Personally, I'd go for non-mirrored and back them up somewhere else or on tape. Burst write speed is not important here, sustained write speed is. Alternate the use of the disks, if possible.

>
>Question: What about the relative benefit of opting for 15KRPM vs.
>10KRPM drives for this setup? Rotational speed difference gives a
>theoretical performance advantage of 50% to the faster drive at
>roughly twice the price.

Use the faster drives on the redo. If you're interested in maximizing the 50% of OLTP. The others you can compensate somewhat (again, depending on the characteristics of the load!) by using the RAID and db cache. Investigate the use of multiple buffer caches!

Cheers
Nuno Souto
nsouto_at_optushome.com.au.nospam Received on Thu Mar 07 2002 - 05:49:12 CST