You probably understand this better than I do, but I have never been
able to
find enough performance difference in PAE vs. default kernels to worry me, at
PAE costs you a few percent on tlb loads. The big hit with > 1GB RAM is
the cost of the remapping of user pages and the TLB flushes it causes,
and on the 'hugemem' > 4GB referencing kernel that grows a lot more as
well as getting a lot of problems with 32bit capable I/O devices and
Intel processors with non IOMMU.
It depends a lot on workload and CPU variant.
I'm writing this in a VM running FC9 under a native
"2.6.22.14-72.fc6PAE" kernel
which is heavily used in native mode using all 4GB memory. I did measure this
against both the non-PAE 32 bit kernel and x86_64 kernel for desktop, gimp, and
kernel builds.
If you are using a VM you've already totally shot your page table
performance to bits so it won't make any difference.
Alan