NUMA Memory Allocation on System with Excess Installed Memory

Question

While chasing down performance issues on a server application we have stumbled across some strange behavior with the way windows addresses physical memory. Our setup is as follows, dual processor servers with node interleaving disabled and 64GB of RAM. The
server is setup to dual boot Windows Server 2008R2 and Windows 2012. On Windows 2008R2, the 64GB of RAM is recognized but only 32GB of RAM is used by the OS. 
Our software consistes of 2 processes continually running, with each affinity set so each process has a dedicated processor. When we run our two processes on 2008R2, we notice one of the processes runs slower than the other; this slowdown is not noticed on
Windows Server 2012. The Intel memory bandwidth performance counters show that on 2008R2 most of the memory is read and written from socket 0, while 2012 shows equal distribution of memory bandwidth across both sockets. If we remove the unused 32GB of RAM
from the server and run the same test on 2008R2, we also see that the memory bandwidth is equal on both sockets. From this test (and a few others), it appears that when Windows 2008R2 chooses the 32GB of physical memory to address, it is only choosing memory
on the socket 0 memory controller. 
Without removing the 32GB of memory, is there any way to configure Windows to equally address physical memory from both sockets?

Answer 1

Hi Mikekehoe,
I've moved this question to the Windows Server General forum, which I believe is the appropriate place for this question.
I'd suggest consulting the documentation specifying the
Memory Limits for Windows Releases.  If you are using Windows 2008 R2 Standard, the physical memory limit is 32GB.
Regards,
--Daniel