Performance Design for OLTP Oracle

[Michael Denney]

Hi all,

Currently involved in a performance consideration and possible re-design of an hpux-oracle database. My considerations are primarily for the os and storage. Here are the requirements so far:

oracle 9i 2tb oltp.
hpux 11.11.
emc symmetrix(moving to dmx3000).
High frequency of small IO.

My thoughts so far on storage performance tuning:

Cooked filesystems for archive and redo filesystems.
Raw filesystems for data and indexes.
Configure new DMX as 1+0 raid in the disk array.
Create 8gb hypers, each hyper spread(extent size?) across as many disks as possible.
Pool the 8gb hypers into 8 member meta-luns, and stripe again at the hyper level during meta-lun creation. So each meta-lun will be a 64gb presentable disk to the operating system.
Each 64gb disk will be a single logical volume-filesystem.

Or should we go with many small hypers presented to the OS?(a bunch of 8gb disks instead of 64gb) Possibly do LVM striping on top of the disk array back end striping?


Thanks for all your input! Any links would be appreciated.

Regards,
Michael

[Christine Hartman]

definitely stripe your metas and then use LVM or VVM to stripe your LVs. This sounds like a bad idea...the old school of thought never stripe at the OS if you are striping at the hardware level...however with EMC's technology you CAN do both...and you will get better performance out of it too. We did this here (although our environment is smaller only 500GB database). We also split out the temp/system table spaces from the rest of the data. They also did mirror copies of redo and archive logs...we attached clariion disk to handle the mirror copy of those logs.

For the archive logs and redo logs filesystem we presented 11GB hypers and striped at the OS using a smaller stripe size. For the larger data files we used striped meta vols and used larger stripe sizes at the OS.

This configuration, although a little more complex did offer better performance....and we really needed it during month end processing!

Hope this helps

[Jov]

Hi Michael,

Striping at the H/W level makes sense, but also stripping at the OS level will make things more complex and might not give the same return depending on its executed.

The thing to watch out of when striping at both the HW (SAN) and OS, is to not have overlapping stips hitting the same disk.

If your SAN will not likely change much in the future, then plan carefully to ensure the disk I/O is evenly distributed then it might make sense, given the affort/complexity verses return in performance gained.


Jov

[RAC_1]

In case of ORACLE I prefer SAME (Stripe All and Mirror Everything)

Also, for archive and redo file systems, if you have OnlineJFS, you can mimic raw file system behaviour with mincahe=direct,convosync=direct mount options.

Oracle does it's own buffering, then whay do it OS level again??

Hope this helps.