Oracle in XP512 and Extend Based Striping

[George Nikoloudis_1]

Hi,

I would like to ask the below:

We are having an Oracle Database of 500GB stored on an XP 512 with 4 RAID Groups in RAID0+1 configuration.

Each RAID Group is sliced on 34 GB, i.e. there are 7 slices per RAID Group.

Each PVG is one slice and each PVG on FIlesystem.

From the Host (Superdome32 , 12 CPUS, 24 GB Memory) there are 4 FCs (2Gb each), 2
primary and 2 alternative

We are facing IO bottleneck problems and we are thinking to use Extend Based Striping filesystem option for maximum distribution of Data.

In the new configuration we are going to create a PVG with 4 slices of RAID Group and 4 file systems with extend based striping.
Also we are going to reorganise the datafiles in Oracle in order to get a blanced IO.

I have the following questions:

1. Is this solution will provide the expected upgrade the system IO (e.g. 25% better IO)?

2. What slices should be used? e.g. two from primary and two from mirror or seperate the mirror and primary?

3. What are the key indicators from system statistics where I can use in order to identify performance problems in IO?

Thanks for the support

[Sanjay Kumar Suri]

I have following to say (or have doubts):

1. Do you mean slice same as LDEV?
2. Which emulation type you have used at XP level?
3. System seems to be quite powerful. Connection from host to XP512 are also adequate with 1+1 redundancy.
4. Are you planning to change the emulation in the new plan?
5. What is the basis of reoganization of data files?
6. How your redo log files/archives files distributed on the XP512? What emulation, disk RPM etc? Do the location of redo log files/archives shared with other datafiles?
7. What is Cache memory/shared memory on XP512?
8. What all applications are runnning simultaneously on the server.
9. Are the files systems created on LDEVS are spread on all ACPs (bit complex)?
10. Are you using RAID1 or RAID5?

Sorry for asking so many question.

sks

[George Nikoloudis_1]

Hi,
some answers

1. Yes
2. OPEN-E
3. -
4. What is the plan?
5,6. redo and archives are seperated from other datafiles. Generaly we tried to follow theOracle directives
7.Maximum Available, as the system is delveed by default
8. There are other application running on the XP but not in the same Raid Groups
9. No we have only one ACP
10.RAID1

[G. Vrijhoeven]

Hi George,

We use extend based striping on our XP1024, but we never used it without so i can not say if it will improve performance ( We think so off- cource) What we do is use both our pri and alternate controllers to excess the extends to use more paths and more switches.

But on a XP you do not communicate direct with disks, but with cache. If that cache is not big enough and you must access the disks to retrieve/write data to i do not think changing striping will help much ( It could help but I think it will not differ more than 15% extra I/O speed)

I assume s.o. checked the Oracle side/ application side. In that department it could be possible to gain more speed.

HTH,

Gideon

[SS_6]

Hi,
Before suspecticng disks layout or XP512 please make sure it is a bottleneck because of disks and XP512. I have seen bottelnecks remains same because the problem was in the programmes not in the XP's.It looks that you have good hardware resources. Use autolun for load balancing and XP prformance manager tool at analyse further.

[Duncan Edmonstone]

George,

Have you read this:

http://otn.oracle.com/deploy/availability/pdf/SAME_HP_WP_112002.pdf

I'm not sure why you would use extent based striping, as you have no requirement to mirror in LVM (the XP is doing that for you) and this restricts you to a minimum stripe size of 1 extent (probably 4MB depending on how your VGs are set up). Better to just use normal striping which gets a better granularity of stripe size.

HTH

Duncan

[Jean-Luc Oudart]

George,

As connected to XP512 where you already have mirroring and striping, I'm not to sure about "extent based striping".

We have Oracle databases onto XP128 (RAID 0+1), we use lvm striping (across 8 LDEVs, alternate controlers) with a stripe size of 64K.
We did benchmark it and difference was less than 5% for our application. Before we had some Qlen on disk. This just help to remove it !

We did not have the chance to change the stripe size, but as Qlen pb was removed I don't think we could get anything more.

I would suggest you tune the Oracle design / application. Run perfstat reports to find out the bottleneck(s) in the application.

Whatever you hear, hardware upgrade is the very last solution/option .
Start with the database/application.

you may to check these threads :
http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=193104

http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=507450

http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=234347

Regards,
Jean-Luc

[Jean-Luc Oudart]

George,

you may read this performance cookbook.
cf. attachment

Regards,
Jean-Luc