Oracle OFA compliance with XP256 disk array

[Aidan Currie]

We have installed an N-class HP cluster with an XP256 storage array shared between the cluster nodes. Raid 1 is being used. We will run Oracle .1.7 server, but NOT Oracle Parallel Server.

As an Oracle DBA about to install Oracle Server 8.1.7, I normally follow the Oracle OFA standards by placing the various components of an Oracle instance ( data, indexes, rollback segments, for example ) on separate disks, according to the OFA architecture.
However, the HP engineers are telling me that, because of the 4 giga byte cache and internal load balancing, I do not need to perform this exercise, but rather consider the XP256 as a black box.

I ( and the HP engineers ) have been unable to find any white papers, giving more technical detail on this argument.

Can anyone share their experiences with configuration in a similar HP / Oracle environment ...?

best regards
Aidan Currie

[Dave Wherry]

Aidan,
I have to side with the HP engineers on this one. I've actually had this discussion/debate with our DBA. It does contradict all you were taught over the years about setting up disks for a database. Moving from a JBOD or small array world to a large array, there is a disconnect you have to make. There;s a new way of thinking.

All writes go to cache and are immediately acknowledged to the host as complete. There is no waiting on the disk.

You also say you are setting up Raid-1. I assume that is because Raid-1 is known to outperform Raid-5 in writes. Once again it all goes to cache. No difference. On the other hand Raid-5 outperforms Raid-1 on reads and gives you more usable space. All my disks are Raid-5 and I/O has been great.

I have a white paper from Hitachi and Oracle discussing Raid-5 outperforming Raid-1 for OLTP. I'll send it to you if you post your email address.

[Matt Walls]

I disagree completely. You SHOULD follow through with the exercise of splitting up files, and follow the standards.

While it is true you can consider the XP a "black box", you should also take advantage of it.
I have migrated our Oracle DB (which is 750 GB) from an EMC3700 to an XP256, and now we have migrated over to the XP512.

Put your hottest files on Raid 1. while people here may disagree, we ran extensive tests on the XP, and raid1 still performs better, although not by much.
i can go into a lengthly discussion if you wish, however the bottom line is to follow the standards. the unix admin and the dba need to work very close on this, as well as having a very good internal map of the xp.

Matt

[Victor BERRIDGE]

I also have done some time ago some performance test...and I agree with Dave,
the OFA was done at a time when to gain speed you had to balance the load acros multiple interfaces and disks in order to get the greatest I/O bandwidth...
the split is achieved at LUN / formatting of the arrays so THAT you do not have to try to balance between disks or you may fall in the case of having always the same disks writing...
to increase performance you can always use more controllers and do some statical load balancing, this will give far better perfoemance than standard RAID1, but as said previously these bays are RAID5 optimized...

Best regards
Victor

[Victor BERRIDGE]

I also have done some time ago some performance test...and I agree with Dave,
the OFA was done at a time when to gain speed you had to balance the load acros multiple interfaces and disks in order to get the greatest I/O bandwidth...
the split is achieved at LUN / formatting of the arrays so THAT you do not have to try to balance between disks or you may fall in the case of having always the same disks writing...
to increase performance you can always use more controllers and do some statical load balancing, this will give far better perfoemance than standard RAID1, but as said previously these bays are RAID5 optimized...

Best regards
Victor

[Alexander M. Ermes]

Hi there.
I agrre with Dace as well. We have different kind of array running ( EMC is standard, but we also have HP ). The main story is always the same. Go as you are told by Oracle and ditribute your databases on the disks working close together with the Unix admin.
It may take a while, but it is definitely worth it. We have 21 databases running with no performance problems depending on I/O because of careful planning. The time you put in now, you will save later.
Rgds
Alexander M. Ermes

[UNIX Admin Group_1]

It will probably pay off in the end to at least put your control files on seperate LV's (or filesystems if not raw). Our Oracle DBA's started getting lazy with the idea that the XP did everything for them as far as performance goes. They put all 3 control files in the same filesystem. It got corrupt - pull out the tapes!! If he had followed part of OFA, at least they would have been in seperate file systems, whether the were on seperate physical disks or not.