Re: Oracle 10g RAC performance

On Fri, 27 Apr 2007 06:15:04 -0700, Cristian Cudizio wrote:

> Oracle says you can reach 80-90% of performance increment for each node.

I work with parallel technologies for a few months now and it hasn't been my experience. As a matter of fact, my first parallel installation was Oracle 6.22 on 2 node VAX Cluster, I was working with SGI Challenge and clustered them under Oracle 7.x, I was maintaining Oracle 8i OPS and 9i RAC systems and I am maintaining 10.2 RAC databases now. As Gopal (K. Gopalakrishnan, one of the foremost Oracle's RAC experts) will tell you, RAC is not just a bigger box, it's much more complex then that. Performance increase is highly dependent upon the type of the application you plan to run. Realistic testing is the key. Problem is that DML locks are now global and take network communication which makes them much more expensive then the local locks. Also, you cannot track them. The key to the V$LOCK table are UNDO segments which provide the values for ID1 and ID2 columns. You can easily find out who are you waiting for, as long as everybody is using the same undo segments, something that is not the case on a properly set up RAC system. With global locks, you don't know who's holding the lock on the row you need. With all improvements made to RAC/ OPS (cache fusion, dynamic remastering, dynamic locks) global locks are still much more expensive then the local ones and if your application is a high intensity OLTP application in which many users occasionally cluster around just a few blocks, RAC will not provide any benefits to you. It will provide baptism by fire instead. RAC is not a silver bullet and is, in my opinion, the greatest possible tool for DW type applications and warehouses. OLTP applications take careful functional partitioning and very expensive HW, such as EMC disk arrays, Infiniband and BFM (very big machines, in Doom terminology).

-- 
http://www.mladen-gogala.com