[SunHELP] I/O size tuning, Solaris 8, Oracle 8, EMC

[listmail]

A while back a question was posted to the sunhelp mailing list that I 
find interesting. Does anyone have any feedback on this topic?

We run Oracle Applications E-Buisness suite 11.5.8 on a couple of 
Solaris 8 boxes and a EMC Clariion FC4700. We currently have a few 4 
spindle - raid 1+0 luns that have a 128k stripe size. We chose to split 
the data and indexes in this setup.

Going forward into a new design, we are currently investing into to a 
larger 16 spindle Raid 1+0 setup and combing all the index and 
tablespace data files together. One question that has come up is whether 
or not we should change to a larger stripe size. We currently have 
maxphys and vol_maxio set at their defaults, but yet I have seen a few 
documents out there that keep referencing this 1mb stripe. Any time that 
someone may have to provide some feedback related to this topic would be 
greatly appreciated. Im only asking for opinions as I want to consider 
as much as possible before making a final descision. Unfortunately I do 
not have a test enviroment within the FC4700.

It would appear to me that staying at the 128k IO size, but yet 
implementing the 1mb stripe size could also be an option.





~~~Previous Discussion

SunHELP] I/O size tuning : VxVM/EMC Symmetrix/FC-AL/Solaris/Oracle
DAUBIGNE Sebastien - BOR ( <removed )
Thu Feb 6 08:40:32 CST 2003
Previous message: [SunHELP] Staroffice 6.0 dumps core after repair, 
loops(?) forever
Next message: [SunHELP] xhost (cont.): automating access control
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Our customer has a DSS (OLAP) running on a E10K/Solaris 2.6/VxVM3.2/VxFS3.4
/ Oracle RDBMS 8.1.7, with 4 FC-AL JNI cards to an EMC Symmetrix 8730 array.
The EMC disks are internally stripped (RAID 0+1), with meta-volumes of 4 HVE
(stripe-unit=960Kb, full-stripe = 4x960Kb).

After reading many VxVM/Oracle/Sun/JNI documents (which concludes that 1 Mb
I/O is optimal for sequential I/O on modern disks because 1 I/O of 1 Mb is
faster than 8 I/O of 128Kb due to average positioning time), we were
convinced that sequential I/O should issue 1 Mb I/O to the Symmetrix to get
optimal performances, so we told our customer to tune the following Oracle /
kernel parameters :

- Solaris maxphys to 1 Mb
- VxVM vol_maxio to 1 Mb
- Oracle db_multiblock_read_count (for seq I/O) to 1 Mb

Hence, Oracle sequential I/O (full-scan) are issued with a size of 1 Mb
instead of 128 Kb (due to maxphys default), which is what Oracle/Sun/VxVM
tell to do in many tuning papers.

But now, EMC Professionnal Services made some analysis on the Symmetrix
(with EMC Worlload Analyzer) and said that *average* I/O size are too high
(200 Kb with 800 Kb peaks, which is less than 1 Mb because it is an average
with mixed sequential (1Mb) and random (16 Kb) Oracle I/O), because it leads
to large Front End (FA) queues (30 to 75 ms), with busy devices and many
locks in the array cache memory.
So EMC advice is to remove kernel parameters tuning to reduce I/O to the 128
Kb Solaris default.

We don't agree with EMC and think that large queues are an expected
consequence of large I/O size which is the best size for seq I/O. Reducing
I/O size will reduce average wait queue BUT will increase number of I/O,
with poor sequential throughput.

I feel like EMC's conclusions are based on "old school" analysis (old disks,
small I/O) and/or OLTP load.

Could someone share any similar experiences, and give advice ?