CPU & Memory allocation in Exadata Database

Question

Hi All,
We are using Exadata 1/8th Rack which is having 256 GB memory and 12 cores on each Database server.
Our requirement is to create 15 databases out of which 10 are standalone and 5 are clustered database.
Now we are concerned about how the memory and cpu will be done for all these databases. I know we should not allocate more than 75% of total memory and cpu on the server. But don't know how much sga and pga I should allocate to each db and what is the mathematics here.
Kindly let us know if any one has done same kind of configuration earlier or any best practices to resolve such issues.
Also, please suggest if instance caging will be helpful here, please share detail if anyone has experience instance caging on exadata.
Your assistance is highly appreciated.
Thanks,
Vineet

Answer 1

Hi Vineet,
Some quick math:
With 10 standalone and 5 2-node clustered databases, that would average to 10 instances per database server.  Again going by straight averages, using 75% of total memory and CPU, that would leave you with an average of 19GB of memory and 0.9 CPU cores per instance.  It's a lot of instances but definitely not impossible.
I'd suggest looking at where the databases are sitting now:  assuming they're Oracle databases, how much SGA and PGA do they have allocated now?  What are the advisers telling you about optimal sizes?
With this many instances, instance caging doesn't really make sense:  you'd only have a single CPU core for each instance.  Again here, looking at the existing workload is a good place to start.  How many CPU cores are they using (typical/peak) already?  Would you expect to use more than 75% of CPU capacity during peak usage periods?  If you are, one option to look from the CPU-management side would be the MAX_UTILIZATION_LIMIT, though the main issue is that, as for instance caging, it only refers to a single instance.
And don't ignore I/O capacity.  Even though X4s have a lot of flash, it doesn't hurt to get statistics for read and write volumes on your databases, and see how they compare to Exadata specs.
And lastly, there's no true substitute for real testing.  If you're using 12c, RAT consolidated replay is very nice, but perhaps you can re-purpose application-level testing infrastructure to test on Exadata as well?
There are a lot of good Exadata and general Oracle consolidation presentations from various conferences floating around.  A Google search will show a few, or you may even want to sign up for virtual attendance for one of the major Oracle conferences to get access to the slides and papers.
HTH!
Marc