Re: RAID 0+1 and recommended stripe size

Hi,

Since we are talking about striping here, I'd like to point out one more issue, which can affect multiblock IO especially on raw devices.

For example, if you do have 64k stripe *unit* size, let say stripe width is 4*64=256k and you have a raw device header let say 8k in the beginning of datafile, then 256k multiblock IO actually has to read from 1st spindle twice.
56k from first, 64k from second, third and fourth and 8k from 1st again.

The normal file systems aren't impacted that much, if read ahead buffering is used on OS level and FTS normally gets to immediately next blocks in file.
Unixes should be ok with read buffering, I dont know about later Oracle versions on Windows, at least Oracle 8.0 on NT4 didn't use any read buffering on OS level.

Thus, multiblock IO aligning to stripe boundaries might be useful during large DB implementations.

I think no read ahead buffering can be used with CFS either, because the data in OS buffer could be changed by some instances without notifying other ones.

Tanel.

> That's why I suggested playing around. I've seen 1m stripes work very
> well with Solaris where the OS max io was set to 1m (which btw required
> some special tweaking of the file system config to get it to work - a
> default ufs/vxfs file system only allowed 256k). But as well, because
> full scans then looked so attractive to the optimizer we had to tweak
> the optimizer_... parms to balance things a bit...
>
> We ended up with whizz-bang IO performance, but this is quite different
> to the advocates of SAME which claims that "you simply set everything to
> 1m and voila!"
>
> hth
> connor
>
> --
> =========================
> Connor McDonald
> http://www.oracledba.co.uk
>
> "Some days you're the pigeon, some days you're the statue"
Received on Fri Mar 21 2003 - 14:24:39 CST