AIX/ORACLE - Raid devices - async I/O [SUMMARY]

These were the responses I got to the following question.

> We are using AIX 3.2.5 and Oracle 7.0.16. We are currently looking at the
> I/O configuration and I had a couple of questions.
 

> 1) The Unix Gang wants to take our raid subsystem and create one filesystem
> across all five drives in the array replacing the 5 filesystems we have
> now. The only thing on this array is oracle datafiles so it sounds like
> a reasonable plan. Instead of having 5 FS's that are striped across
> 5 disks we'll have 1 FS across the same disks.
>
> All things being equal, can anybody think of a good reason not to do this?
 

>2) The Oracle instalation manual make a big todo about the async_io kernel
> extentions. And setting the Oracle initializaion parameter to make use
> of async_io does seem to produce a real performance benefit. This was
> suprising to me since we are not using raw devices and I would have
> expected the FS code to already be asynchronous.
>
> Are not writes to the JFS already async and if so how does the async_io
> option increase performance of I/O?

Eric Keen <e_keen_at_access.digex.net>

1. It doesn't matter who does the striping - you know your data.
2. On AIX I've measured improvements over multiple db writers, but on the latest installs it forces async io anyway.

   One thing you should look at is altering the method Oracle reads from    disk - change it from fixed block to variable (AIX and SUN support it,    but some UNIX platforms don't) via:

   use_readv = true

cbarr_at_world.std.com (Chris Barr)

1. One problem is that disk MTBF is a lot higher. Think about it - if you have a 1% chance of being hit by a car when crossing a 1-lane highway, what's the chance when crossing a 5-lane one? Once the FS is striped across 5 disks, the potential for failure is greater. Of course, your RAID mirroring negates this - if it's the right level of RAID. It should be a big win for you to have a much bigger file system space.
2. I just talked with someone today today who'd been at IOUW and mentioned that Oracle people were strongly recommending async_io writes & reads be turned on in Sun Solaris, to take advantage of new enhancements, perhaps by Sun, not Oracle. But the funny thing was that some Oracle people said async-io was set on by default, some said it was off.

   His comment was that it closely approaches raw-file-io performance.    And, of course, without the admin headaches.

S.Cohen_at_ma30.bull.com (Steve Cohen)

1. This might not be as clever as it sounds. Unless the RAID device has a BIG cache, your five separate disks will now logically act like one disk. (I assume they are going to give you RAID 5).

   If you are doing transaction processing, and your disks are busy now,    you will very likely see a monstrous decrease in the amount of work you    can do on those drives.

   We at Bull sell our own RAID which can have up to a 64 megabyte cache    for this reason.

-- 
Regards, 

Lee E. Parsons                  		
Systems Oracle DBA	 			lparsons_at_world.std.com