| Oracle FAQ | Your Portal to the Oracle Knowledge Grid | |
 |
 | |||
| Oracle FAQ | Your Portal to the Oracle Knowledge Grid | |
|
| |||
Home -> Community -> Mailing Lists -> Oracle-L -> Re: Solid State Disks for Databases
Hi all,
I am a little bit late here but never too late :)
I used to work for the SW company making the SW for the biggest telecom companies on the world. As part of having billions transactions per day in some of them, happily storing all mobile and other network events into Oracle for crazy batch processing we were put into the position to test SSD.
I personally tested SSD and will say that (as Cary
Millsap pointed out) I needed first needed to make the
proper test case where disk I/O is really the
bottleneck and SW is doing the best, the db design is
the best and so on and where SSD is going to be a big
gain in % of overall response time.
I should also test stupid apps without tuning and so
on, but that was stupid for me to buy SSD before doing
proper good tuning.
The first suspect was lg writing and famous log file sync. That was very natural problem to us because we used bulk inserts/DML to do a lot of transactions per second having in mind that 64 or 108 CPU's are our limiting HW factor dependant on which customer prefer which OS.
So I end it up having crazy fast SSD, but could not gain too much because of many scenarious failing to make big overall improvement using SSD.
One thing to say when I removed I/O bottleneck it
looks as Oracle becoming very messy database where the
bottleneck (as a few people said) moved to the CPU.
I had crazy latching problems without crazy
books/notes/knowledge to help me (apart from Steve
Adams book and site, ..).
I needed even to experiment with spin_count to get SSD
be a little bit faster then normal raw device when I
hit redo latches problems and I could not tweak
anymore.
Also the big thing with SSD was that they never think
about integrating all of that into the proper huge HW
architecture.
Who will need SSD for the small inexpensive SW anyway,
what will be ROI for that :)
I believe SSD have a usage somewhere for non-tuned applications, using bad things everywhere and disk I/O for log file sync, disk reads, ... is very bad so it is much "cheaper" (I will say easier because it is hard to find good person to do tuning for less) to pay for SSD then to tune, make their applications better.
Regards,
Zoran
-- http://www.freelists.org/webpage/oracle-lReceived on Wed Oct 05 2005 - 14:37:40 CDT
 |
 |