RE: write performance - broad generalization
Date: Mon, 6 Apr 2009 09:47:33 -0400
Message-ID: <BAY141-DAV878691098756F2DA270BBA6840_at_phx.gbl>
Message-ID: <000f01c9b6be$3cd4f2c0$d17810ac_at_bell.corp.bce.ca>
This is good, as I didn't think I was off base as this is what I want to test. Just a single threaded operation. Pure sequential write performance on a fresh db with no other users. No data etc. If I need to validate more I can. Often times I'll just get handed a db to take over (I'm mostly on DSS systems) so I just like to run this as it's very quick to do/setup.
-----Original Message-----
From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org]
On Behalf Of Christo Kutrovsky
Sent: Friday, April 03, 2009 3:47 PM
To: greg_at_structureddata.org
Cc: ganstadba_at_hotmail.com; Oracle-L Freelists
Subject: Re: write performance - broad generalization
Michael,
I agree with Greg, you need to test multiple concurrent write to get maximum performance. Multiple controllers and multiple paths to the disk are only used when you have several IOs to be written at the same instant.
Note that this is a pure sequential write performance test. It is very different then what DBWR will write.
If you want to perform the same read test, you can use rman's "backup as copy validate" and setup multiple default channels.
Things to be carefully about:
- direct_io
- async_io
- shared metavolumes on the SAN and activity on other machines.
The times you get should be very consistent. Especially single thread. If it changes, then follow up on some of the above points.
-- -- http://www.freelists.org/webpage/oracle-lReceived on Mon Apr 06 2009 - 08:47:33 CDT