AIX

Hello,

It's been a while since I've done AIX..., but I'm planning
a new TSM on AIX disk-only backup solution.

I'm planning to make an AIX volume group out of 40 luns of 1 TB.
I'm planning to create one big file system on here.
The purpose for this is to use this as a device class FILE for a TSM disk-only backup solution.

My question is,
would it make any difference, performance-wise, if I split up this capacity over multiple file systems, or even multiple VG's ?

right now this is the plan:
---------------------------------
-> one filesystem on there /FILECLASS/ -> 40 TB
(all backup data will be written to the same file system and VG)

would this be better?:
----------------------------
-> two filesystems on here: /FILECLASS1/ -> 20 TB
/FILECLASS2/ -> 20 TB
(when you specify multiple directories for a device class file in TSM, it spreads its volumes in
a balanced way over the directories that you specify,
so in this case the backups would be landing on 2 different filesystems in the same VG)

would this be even better ?:
------------------------------------
-> one FILESYSTEM on here: /FILECLASS1/
-> one FILESYSTEM on here: /FILECLASS2/
(when you specify multiple directories for a device class file in TSM, it spreads it volumes in
a balanced way over the directories that you specify,
so in this case the backups would be landing on 2 different VG's)

would this make a difference for performance or might I just as well put everything on one file system in one VG ?

Also is there some kind of optimal lun size for this (from an AIX OS perspective) ?


some extra info:

(without DIO since the TSM device class FILE works in a sequential way).

(EMC Ax4 with 2TB SATA2 disks)

any hints/thoughts would be appreciated

There will be no difference if you have 1 or 10 filesystems - and you should put all of your luns into the same VG to make sure you have the max buffers. As these are on a per disk basis, this will in any case require a lot of file system buffer tuning. I think I would rather use 100 GB luns and corresponding more adapters if I really want more performance. Apart from that, you can choose a smaller pp unit, what will distribute your content more evenly across the disks.
Make sure that your filesystem is spread across the luns with maximum distribution. If you can, use more than one pair of hba's. And I would mount your filesystem(s) with rbrw to save memory and with noatime as well - dio would be JFS, and cio would saturate the disks and make no sense anyways.
Regards
zxmaus

Thanks alot for your reply zxmaus,
very helpful information, much appreciated!

I think we will go ahead as planned in the classical way,
however...
someone has also explained to me the possibility of META-LUNS on this EMC disksystem:

We have 3 RAID6 arrays on the storage system.
Instead of creating 40 1 TB luns,
We could also make 1 huge lun per array,
so we would have 3 huge luns,
and then concatenate these 3 luns together into a META-LUN.
The storage system then stripes over all the disks automatically this way.

This way, one giant disk of 40 TB would be presented to AIX,
and a VG could be created, out of 1 PV.

So all the striping would be handled by the storage system itself and
there is no more parallelism in AIX then.

However,
I have never heard of anybody working like this,
and the thought of having one big 40 TB lun is pretty scary.
Does anybody have any experience with this?


I guess we could also make a combination of both:
We could also make 500 GB luns on the 3 arrays,
and concatenate each time 3 500 GB luns into 1.5 TB meta luns for example,
presenting about 26 1.5 TB disks to AIX in this way.
this way we would be using both parallellism mechanisms:
- on the AIX OS with the VG
- on the storage box with META-LUNs

I'm starting to consider the combination of both mechanisms
as described above.

Any thoughts on this?

well if you write to ONE lun - its slower than if you write to 10 luns in parallel as AIX simply doesnt care about what is behind the adapters.

Real life experience about a year ago - attaching my 124 luns - 33 GB each - from one to 3 pairs of hba's and changing the 4 DB filesystems from minimum to maximum distribution across the disks - improved our EOD batch (oracle) from 3.5 hrs to 14 min.

Regards
zxmaus

Thanks again for your quick reply zxmaus.

What exactly do you mean with ?
something with the PP size?
could you clarify this a bit?

as for the HBAs, we are really limited here as all HBAs on the server are in use...
* 2 for a VTL
* 2 for another disk system
* 2 for tape

so we'll have to share HBA's with one of the above (no other choice...),
it will be the 2 HBA's from the other disk system that will be shared with the new disk system.

We could also share the HBAs that are currently used for tape devices with our new disk system, but this seems like a much worse idea.
Though the device class FILE, at TSM level,works in a sequential way,
it will be writing in the background into files on an AIX JFS2 file system,
so sharing the tape HBAs seemed like a no go to me (though the FILE device class is sequential I/O at application level (TSM)) my head hurts..

Possibly we will try both (sharing HBAs with disk system first, afterwards share with the tape HBAs and see what gives best performance).

what I meant with maximum distribution is a chlv -e x logicalvolume - basically you make sure that the first physical partition goes onto disk1, the next one onto disk 2 ... if you have 10 disks, than the 11th physical partition goes again onto disk1 and the 12th onto disk2 ... in opposite to minimum distribution where the first x physical partitions go onto disk1 until its full and so on ...
Hope that clarifies it now a little more

Kind regards
zxmaus