Opteron & Exchange Memory Optimization

[Erik Jans]

Hi,

We have a ProLiant DL385 with two Dual core Opterons (4 cpu total).
From reading various articles it has come to my attention that Opteron uses NUMA (non uniform memory architecture), and this raises a question about memory optimizations.
Should Node Interleaving be turned on for a server running Exchange Server?

As for memory configuration, we have populated all of the 8 slots with 512 MB each for a total of 4 GB.


Here's a reference to NUMA: http://www.digit-life.com/articles2/cpu/rmma-numa.html

-

Thanks
/
Erik

[Basil Vizgin · Accepted solution]

Windows 2003 understand NUMA and tries to optimize application behaviour.

http://www.microsoft.com/whdc/system/platform/server/datacenter/numa_isv.mspx

As I know, Exchange itself doesn't use NUMA specifically, but lay on Windows Memory Subsytem.
IMHO, you can use NUMA mode.

[Erik Jans]

Thank you Basil,

That's what I was looking for.

This raises another question though. What kind of performance drop would a Windows 2003 Server standard edition incur, since it does not have NUMA optimizations?

This server is not live yet, and our plans involve using Windows 2003 Server Standard edition.

Best Regards
/Erik

[David Claypool]

While Windows 2000 may not expressly understand NUMA, the Windows Scheduler naturally tries to schedule threads on the same processor time after time to maximize cache hits. This would have the naturally occuring result of optimizing somewhat for NUMA, even though NUMA is not specfically understood. I doubt the difference will be noticeable on a 2-socket system.

[Ruslan]

Hi Etik!

In your situation the only way to adjust Exchange memory utilization is using /3GB switch in boot.ini file

[Erik Jans]

Thanks David and Ruslan,

Ruslan, I might have missed something about the /3GB switch, but I fail to see the relevance in my case concerning NUMA optimizations. Do you care to elaborate?

David, thanks for the thread scheduling lesson. :)

Considering our server is a ProLiant DL385 with 2x Dual Core Opterons. Would this lead to the assumption that it will operate with 2 memory nodes?
Or would a node be created for each actual processor-unit (4)?

Best Regards
/Erik