Oracle performance on VA/EVA with stripping

[Charbel Bou-eid]

Dear all,

I know this debate has taken a lot of your times, but I'm sorry to raise it again as I didn't really see any decisive answer to this dilemna.

Therefore allow me to ask the following questions on Oracle (I'm not a DBA) performance on HP VA/EVA:

1- Is it always better to create data & indexes on different VGs (i.e LUNs)?

2- If so, is it not better to create opposite primary paths to these two LUNs via pvlinks (VA) or Secure Path (EVA)?

3- Also is it advisable for data & indexes VGs to include two LUNs each with opposite primary path access to each of the LUNs within the same VG while also creating opposite primary path access for the VGs themselves. Example

VG 1 = LUN 1 & LUN 2
VG 2 = LUN 3 & LUN 4

VG 1 = /dev/dsk/cxtyd1 /dev/dsk/cx'ty'd1 /dev/dsk/cx'ty'd2 /dev/dsk/cxtyd2

VG 2 = /dev/dsk/cx'ty'd3 /dev/dsk/cxtyd3 /dev/dsk/cxtyd4 /dev/dsk/cx'ty'd4

4- If 3 is acceptable, is stripping LVs in these 2 VGs (-i 2 -I 64) will increase or not the performance.

5- Finally, I've read once that redo log LVs should not be stipped. But what about this rbs (I think it is raw back segment)?

Many thanks in advance for all your imputs.

Best regards,
Charbel

[Steven E. Protter · Accepted solution]

1) No, not when its striped.
2) Set up alternate paths with PVLINKS if two pathes actually exist to the disks. This will help with reliability, not performance.
3) I'm not sure how much of a difference it makes. ORacle doesn't like to be striped.
This item should work technically
4) Striping is the bottleneck.

5) Redo logs should be raid 1 or raid 1/0. Data should be raid 1.

I would suggest that you learn how to configure the disk array to present the disk in a different way. Either present the disks unmirrored and set up raid the way oracle recommends or configure it for raid 10 and present it that way. That of course costs you half your raw disk space.

I sense a pattern. You want to tweak striping for best performance. If thats the case, the VA unit already does that pretty well. There is no need to do anything other than alternate links for reliability. Maximiize the number of disks data/index and redo are on.

If you want to follow Oracle guidelines, here they are again. Oracles guidelines in this case are the path to good performance.

SEP

[Fred Ruffet]

You should have a look at this document from oracle, especially the "recomendation" chapter at the end.

RBS=Rollback Segments. Used to Rollback transaction when cancelling. The ycan be on RAID5 (considered as standard data file)

regards,

Fred

[Fred Ruffet]

I think I forgot something :)

[Uwe Zessin]

Charbel,

2) on the VA74x0 it is more tricky, because each controller maintains one of the redundancy groups and all embedded LUNs. You can access the LUN through the other controller, but it will internally re-route the I/O request.

On the EVA you always have two ports per controller usually going to two different fabrics. With two fibre channel adapters you can do load-balancing between both adapters and one controller's ports.

3/4) I am not sure I fully understand that, but on the EVA it makes sense to equally distribute the virtual disks over both controllers (unlike on the VA it is possible here, even if you have only one disk group), because you can then use both controller's CPU power and cache capacity.

5) VA and EVA will always stripe the data within their redundancy/disk groups anyway.