What is the bottom line on Oracle layout on Virtual Arrays?

From: Peter Hitchmough <oracle_arch_at_disley.plus.com>
Date: 2 Mar 2004 09:30:35 -0800
Message-ID: <58cb6300.0403020930.225e9dbf_at_posting.google.com>


What is the *truth* when it comes to configuring Oracle files on an Enterprise Virtual Array like an HP EVA3000 or EVA5000 or an EMC Symmetrix frame.

As I Understand It -

A Virtual Array is a storage appliance with a large number of disks
(say over 50), dual-path controller access and, significantly, a very
large disk cache. The cache is supported by front-end and back-end
(disk-side) processors and a range of algorithms that aim to keep all
read and write disk access via the cache.

And this is not RAID. It's a definite storage evolution. The VA presents the virtual disks to hosts connected via switched fabric. The virtual disks have the apparent behaviour of different RAID levels, e.g. 0, 1 and RAID-5. HP calls these settings Vraid0, VRaid1 and VRaid5. How they are actually implemented is left to the VA.

In a search for sanity, I have read opinion, history ancient and modern, experts' comment and practitioners' experiences. This seems to be a hot topic but it seems that the following need to be verified or de-bunked:

  1. Given the cache-centred architecture of disk arrays, how much effort should be made separating redo logs from data. I suspect that using VRaid-1 over VRaid-5 is more about data redundancy than performance. Where RAID-1 was always better than RAID-5 for log files, because of the pattern of sequential log writes. Is there any sense in creating a separate VRaid-1 disk group for logfiles?
  2. I/O balancing should be achieved by spreading the virtual disks between the controllers on creation, so that each controller is working optimally. The issue seems to shift from hot/cold disks to hot/cold controllers. Can we influence the disk layout?
  3. Is it true that the array handles physical layout totally, rebalancing and migrating data according to usage patterns? If so, the best practice may be to choose appropriate stripe size and to choose an extent size to exploit the many disks in a disk group. The aim would be to feed the cache efficiently at all times.
  4. Do I hang my DBA boots up then and use Vraid-5 throughout? What are the probable recommendations for rollback/redo? archive logs? control files?
  5. A fear born of experience here with SAN is a delinquent controller. What can I do to minimise the impact of a controller failure on the database? A one-disk failure? A two-disk failure?
  6. What is the suitability of Virtual Array technology for DW? OLTP? large operational datastore with heavy IUD activity?

My plan is to review experiences and try to test some options before my next EVA is built. It looks like the book on storage and Oracle has been rewritten yet again.

Many thanks,
Peter Hitchmough

"Why are there marks against these errors in the Error Messages manual?"
"I'm trying to collect the set." Received on Tue Mar 02 2004 - 18:30:35 CET

Original text of this message