AutoRaid Configuration

[Marc Bohnert]

I have a K class machine running HPUX 10.20. It is using AutoRAID. The RAID unit has one controller and the HP is connected to it via one SCSI controller.

I have my filesystems configured such that they are under one volume group (vg00). I am running an Oracle database and even though the data files for the data and indexes are on different logical volumes the RAID automatically decides how to stripe them across the disks.

I want to add another RAID controller (to be able to take advantage of wirte caching and another I/O path) and another HP SCSI controller.

How do I maximize performance given this hardware configuration? Do I need to layout my filesystems into multiple volume groups or do I not change anything and let AutoRAID do the configuring? I want to insure that I get the highest level of throughput.

Thanks in advance,

Marc

[A. Clay Stephenson]

Add a second controller to the AutoRAID itself
and a 2nd HSC SCSI controller to the K-box.
I prefer to leave vg00 strictly for OS stuff.
For best performance you should use the OnlineJFS product.

I would split the oracle data into two distinct groups. 1) Redo logs and archive logs
and 2) Tables and indexes

Estimate the total size required for the redo logs and archive logs and then estimate the total space required for everything else. This total size plus a bit more will be the required size of a new volume group.

Create two identically sized LUN's (e.g LUN 1 & LUN 2) on the autoraid which together are the desired size.

Create a new volume group (e.g vg02) with
LUN1 primary path using AutoRAID controller X (alt Y) and LUN2 primary Path using controller Y (alt X).

Create a logical volume (e.g. oradata) for the tables and indexes which is stripped across both LUNS probably using 64K strip size.
Create a logical volume (e.g. oradata2) for the redo logs and archive logs which is also stripped across both LUNS. The idea here is to
fully utilize both external SCSI buses attached to the AutoRAID. Both of these logical
volumes should have vxfs filesystems created on them.

The final performance trick is to take advantage of some of the OnlineJFS mount options. For the table and indexes filesystem
use the mount options -o delaylog,nodatainlog,convosync=direct,mincache=direct. The convosync and mincache options bypass the UNIX buffer cache and give you all the performance of raw disks. For the other filesystem with the redo and archive logs use
just the -o delaylog,nodatainlog options as these perform better with UNIX buffer cache.

The final performance tip is to leave as much AutoRAID space as possible as unconfigured. This maximizes RAID 0/1 space. In any event try not to exceed about 70% as configured LUN's
and less (55-60%) is better still.