PAE (Physical Address Extension) and NUMA are two different things... Since your system has 2 GB, no need for PAE
Unfortunately, the S2875 should not support NUMA, there is only one set of memory slots connected to Processor 0.
Whereas the S2885 has memory slots for both processors.
See the Tyan FAQ for the Tyan S2885
' ">http://www.tyan.com/support/html/f_s2885.html (not the board you have)QUOTE
How is memory presented to the processor?
Each processor can have physical memory connected to it. After initialization, the memory map for the system presents all physical memory as a single contiguous extent. This extent is formed in one of two ways. The first way is through the concatenation of all processor memory, also known as NUMA (Non-Uniform Memory Architecture). For example, all of processor 0's memory, followed by all of processor 1's memory and etc. The second way is by the interleaving by page of all processor memory. For example, a page from processor 0's memory, followed by a page from processor 1's memory and etc. The interleaved mechanism (option #2) has better overall uniformity of latency but the concatenation mechanism (option #1) is more usable for an operating system capable of implementing memory affinity management. After initialization, all memory appears equal to the casual programmer. No special addressing mechanisms are necessary to access physical memory attached to a processor other than the processor executing the memory reference instruction.
Is it necessary for the OS to have specific memory management considerations; given the NUMA approach used by AMD Opteron system?
Although not necessary, operating systems that provide careful memory and scheduling affinity improve overall memory latency and bandwidth and reduce traffic on the system's buses, which can have important performance improvement in some applications. Both Windows Extended 64 bit and x86-64 Linux have some NUMA support for memory affinity and thread-biasing to enable optimal application performance considerations. Processes and threads that use local data are advised to allocate that local data space from within the thread or process that uses that data the most, to enable maximum benefit from the OS NUMA support.
If all the memory is associated with one processor, as it is in the S2875, there is little benefit to define processor/memory affinity