DS8000 disk array and host base striping

[JOE L. TAYLOR JR]

I'm in the process of attaching our rp7420, oracel database server and a rp5470 a front-end server to a DS8000 disk array. In the IBM documentation it recommends host base striping with a stripe size between 4mb and 64mb. If I choose a 4mb stripe do I need to create my file system with a block size of 4mb as recommended in other posts where it is recommended to match FS block size with stripe size. Is 4mb stripe size to large?

[Zinky]

It depends but generally Oracle is happy with 8K block size. However other savvy DBAs prefer multiple block sizes which started to be possible I think with Oracle 9i.

Our stripe sizes on our 4 or 8 way stripes range anywhere between 32K (OLTP) to 128K (DSS). 64K stripe sizes are for our general purpose DB instances.

4MB stripe sizes seems particularly large to me but if thay is IBM's best practice then...

[Jeff Schussele]

Hi Joe,

Well if you use host-based striping then the smallest stripe available would be the smallest PE size available which is 1 MB.
And this is achieved using extent-based striping.
Anything smaller "must" be array-based.

Rgds,
Jeff

[Zinky]

Not if you are using VxVM.

vxassist -g ds800dg make oravol1 layout=striped ncols=8 stwidth=64

(If you've the proper ASL, it will intelligently pick the lUNS for ya.)

Or if you are using non-extent based striping with LVM..(of course I cannot say for sure since we ditched LVM before I could play around with striping on thet VM)

[Jeff Schussele]

Nelson's correct of course.
I was referring to mirrored host-based striping.
If the array will be a RAID of some sort then host-based mirroring would not be required & host-based striping can size from 4 KB to 32768 KB in powers of 2.
So what you want is doable.

Rgds,
Jeff

[JOE L. TAYLOR JR]

We would be using LVM to configure the disk. The lv and file system that I was going to mimic in this test was striped for 3 disk with a stripe size of 128. The file system block size is 8k. My big concern was the huge difference between a 128k stripe size and a 4mb stripe size. And then if I did make the stripe size 4mb would I need to create the block size for the file system at 4mb as well. Again big gap between 8k and 4mb, so I wasn't sure if this was a valid configuration. Thanks for your ideas on this.

[Jeff Schussele · Accepted solution]

Hi (again) Joe,

I wouldn't be overly concerned about the block size. I never found a compelling reason to deviate from the default 8 KB size.
It only affects the buffer cache read ahead performance. And only when reads will be *hugely* contigous & little randomness would it pay to increase it.

My 2 cents,
Jeff

[Jeff Schussele]

I should also add that another reason you don't really need to adjust block size is that there are kernel parameters that can be set to "tune" the read ahead size - such as for 11i:

vxfs_max_ra_kbytes
vxfs_ra_per_disk

Rgds,
Jeff

[JOE L. TAYLOR JR]

Will leave block size at 8k. Will contact IBM regarding stripe size, 4mb seems too high.