LVM stripes vs. distributed. And the Oracle Database.

[Andre Braganca]

Hi, two issues:

1) I'm having some difficulty in understanding the differences between the striped policy (lvcreate -i -I) and the extent based distributed policy (lvcreate -D) when creating a logical volume.

If I have 4 disks in a VG, then I can stripe a LV across just 2 disks, but I can only distribute it across all the disks. Right ?

But if I have a 2 disk VG with a LV with 2 stripes of size 16384K, what is the difference from a distributed LV, if the PE is 16MB ?


2) I'm going to use an EVA3000 as storage for an Itanium rx2600 with HP-UX 11.23 and a 9.2 Oracle Database. In order to take advantage of all the redundant paths that link the EVA to the HBAs in the server it was recommended to create a VG with 2 LUNs. Each seen by one of the 2 controllers, due to the "Secure Path" sw. Then to create distributed LV so that the 2 controllers are used when accessing a single LV.

The HP-UX is factory installed, with standard LVM settings (no Online JFS). The Oracle DB_BLOCK_SIZE is 8K. The db_file_multiblock_read_count=8. Shouldn't I try to use a stripe size of 64K instead of an 16MB extent ? I suppose it's very different.

Regards,
Andre'

[Steven E. Protter]

1) I don't believe in putting oracle down on striped volumes. It is a significant performance drag and not recommended by oracle for OLTP Online Real Time Transaction Processing databases.

2)The alternate paths will provide you reliablilty. If you lose a fiber card under this sceneroi with PVLINKS your database will stay up and you can swap out the bad card when its convenient.

A stripe size of 64K will probably not help your performance much unless you have very wide tables rows. Larger block size does help with large read operations, say reading the entire DB.

There are a lot of back threads on this, including one that I started. Its worth going back and reading the trials of otheres.

On the search engine.

The EVA3000 handles striping anyway, is there consideration to changing the defaults on that?

SEP

[Jean-Luc Oudart]

We have a few servers attached to XP128.
On one NClass we use for oracle the disks as RAID0+1 on the SAN and on lvm we stripe the lv so at each strip we change controler. (this is when you add the disks to the vg).

We benchmark the app. and found some improvement with these settings.

another thread
http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=193104

I suppose if you could benchmark your application ...

regards,
Jean-Luc

[Alzhy]

Andre'

For "Cache-Centric" Arrays like the big EMC's and HItachi's (HP XP Line and Sun Storedge 99xx or HDS Lightning/Thunder..) --- You will get very good performance if you stripe accross LUNs (no matter what its RAID protection is on the array itself) and prefreably accross different controllers and channels..

For "Controller-Centric" Arrays - like the HP StorageWorks line (EVA, MA, HSG80's.. etc..).. NEVER stripe as it will just overrun your controllers.. Just use whole LUNs or concat the LUNs..
paying particular attention that your DBA spread the datafiles accross however many filesystems (OFA still is good practice)..

HTH.

[Andre Braganca]

No feedback on issue 1 ?

About issue 2:

The Itanium has 2 HBAs, attached to 2 SAN switches. Each switch is connected to 2 HSV controllers. Attached is the spmgr display showing the 2 LUNs.

Regards,
Andre

[Alzhy]

Andre,

Your using an EVA which is a no-brainer to use and setup.. Use the LUNs as is as it is presented to you.. period. DO NOT STRIPE on the LVM side.. you will just hose up your HSV controllers...

[James A. Donovan]

To try to answer your first question...

The whole point of the -D option is to provide a mechanism for mirroring logical volumes, while at the same time providing a rudimentary level of disk striping. This is as close as LVM can get to supporting RAID1+0. In theory, it's a nice idea, in practice I don't think it's that great. The only way you could get close to true RAID1+0 performance would be if you chose the smallest possible PE size, and then that would tend to limit how large your individual disks could be.

As for the second question, I think it's been answered quite well already....

[Mike Dunphy_1]

Distributed striping is used when you need to mirror since you cannot mirror and stripe at the same time. The smallest extent is 1MB. Whereas striped policy is for no mirroing at the software level that is. If you do not need
to software mirror do not use distributed policy

we have seen significant performance by striping across multiple controllers on a VA .. just make sure you have the performance path