Originally posted by: adeal

Hello Everyone,

I'm looking to locate a list of best practices for DMX3 (EMC Symmetrix) disk layout on AIX 5L. So far, I've located Nigel's website (http://www.ibm.com/developerworks/aix/library/au-analyze_aix/#resources), and the three Redbooks he mentioned (SG24-4810-01, SG24-5511, SG24-6039) In addition, I've also found the DS4800 Redbook (SG246363).

Unfortunately, I've found little information specifically referencing "cross-vendor" configurations (i.e. AIX 5L to Symmetrix). I'm hopeful that maybe I have been looking in the wrong locations, and that there is truly a large amount of data for this setup.

I'm looking for some understanding towards questions like whether one should use similar size LUNs (metas) for a particular filesystem (say, four 200GB LUNs for a 800GB volume group), which is what Nigel and the IBM Redbooks suggest, or one large meta (EMC's suggestion)? If anyone has any detail or real-world testing on this that they can refer me to, it would be appreciated.

This database Progress Database is not Oracle or DB2, so unfortunately more "common" strategies involving AIO or DIO / CIO will not benefit this configuration.

Thanks in Advance. -Adam
#AIX-Forum

Originally posted by: CRM

There is no best practise in my opinion for disk attachement for AIX, just best practise for the particular application that runs on AIX.

Your 800GB - what is this supporting?

If you have 200GB of database storage and 200GB of transaction logs, database best practise would dictate that you should split this into at least two luns so that the back end storage could be presented from different spindles.

If you have 800GB in a single filesystem but this needs tens of thousands of IOPS, then you would want to split this so that you can spread this across many LUNs and spindles.

How often will this grow, if you start with 800GB luns then the standard growth chunk is 800Gb, using 200GB chunks is much easier to expand gradually.

The fewer number of disks you have though the less you have to manage.

If you have another 200GB for scratch storage/backups then you would want this on another lun so you could present this from slower cheaper disks for example.

The main questions to ask for your 800 GB of storage would be:

1. What are the different filesystems/chunks of storage within this
2. For each chunk of storage what are the requirements in terms of IO and availability

Taking this and mapping it to the tiers available you can present storage that meets the requirements yet is also as cheap as possible.

If you have these requirements I am sure we can offer many different opinions on the optimum layout.

Oh - and rememeber that performance and tuning is an iterative process and the first layout may not be the best, you may have to jig some bits around, especially if you have storage on shared spindles with other servers.

regards

Chris
#AIX-Forum

Originally posted by: adeal

Chris, The 800GB was just a "made-up" example, but I understand your points. The biggest question I was trying to answer is around presenting the db file system itself, and whether spreading the I/Os across multiple devices (LUNs/metas) makes more sense (say 4 to 8 LUNs versus 1 large LUN).

All things being equal, if you were to obtain one large LUN having 80 disks involved, or 4 LUNs each having 20 disks each (80 total disks), would you see a performance difference in the 4 LUNs because of having more devices (hdisks)/ queues involved at an OS level?

Nigel's webpage is one of several "more recent" references that mention using LVM striping on the host (I assume this is in addition to hardware striping already on the array -- say RAID 10 or RAID 5 on the array plus RAID 0 on the host (LVM)). Other documents seem to debate that, and say that striping should be left to the disk array / hardware.

So to recap, is it best-practice to:

1. Use smaller LUNs, rather than one large LUN? (Why / Why not)
2. Use LVM striping on the host? (Why / Why not)

Hope this helps to clarify my original post. Thanks again.
#AIX-Forum

Originally posted by: CRM

Damn, and here was me thinking that the standard architect answer of "it depends" would do....

Ok - here is my 50 cents, probably going to start a heated debate...

I tend to work through any design using the following principles:

1. Keep the number of disks as low as possible, easier to manage etc, but if other factors dictate then split these up, these factors will be:

2. An HBA adapter can have 2048 IO operations outstanding, a disk can have (NB need to confirm this one) as far as I recall 512 operations outstanding. Do I have very high IO requirements and AIO where I need to tune disk/adapters

3. What are my LUN IOPS requiremetns and how do I map these onto underlying spindles. Make sure I only use my most expensive disk only where required, i.e. split down my big disks into different IOPS classes.

4. I am personally not a fan of LVM striping, if you add more space you cannot re-stripe this I understand (although see something like Oracle ASM where you can). If you look at a modern disk array you already have stripes of stripes, i.e. you can carve a LUN from stripes across a number of ARRAYS. IF you bring in a third method of striping then I have actually seen negative performance impacts. If I can keep the striping in one place (i.e. disk or operating system) then I am always happier, but that may just be me.

5. What is the array used for - a DMX3 will not be dedicated just to you, what other applications are using the same underlying spindles as you, when are they busy, do they run batch jobs overnight or during the day etc etc. Do I nees to split up a LUN into smaller LUNs to make spindle allocation easier?

6. The last thing I always bear in mind is that I will always be close but not correct with any design and until you have done some testing/benchmarking you will not know if you got it right.

regards

Chris
#AIX-Forum