Matt Sayler (sayler@thewalrus.org) on 5/15/07 wrote:
---------------------------
>I have several OS X desktop boxes with at least 2GB of memory. I've heard that
>things like Parallels and VMWare have problems with machines with that much memory
>(which makes me think that OS X magically turns on PAE at some threshold).

I have a Mac Pro with 4GB. VMWare runs fine on it.

>Any information? (I know OS X has some 64-bit support on G5, but more in the Solaris
>split bit-ian way. I don't think it does on x86)

OS X is a bit different from the operating systems which have been discussed previously.

The most important difference is that it always uses a 4:4 split for 32-bit processes: the kernel is not mapped, it lives in its own address space, so the process and kernel see individual 4GB spaces. Every system call involves a transition from one virtual address space to another (and back). They take the performance hit in order to allow 32-bit processes to map more memory.

Furthermore, while 10.4 (and later this year 10.5) support 64-bit user processes, the kernel itself runs in a 32-bit virtual address space. Only the pieces which need to deal with physical memory (the VM layer and the device driver layer) know about larger addresses.

Apple's goals in kernel design generally seem to be maximum binary compatibility and flexibility over performance. For example, when Apple introduced G5 Macs, the combination of their IOKit device driver layer and an IOMMU allowed them to have old 32-bit drivers automagically support DMA to and from 64-bit physical addresses. This wouldn't have been possible if IOKit didn't have some very powerful abstractions, something which is usually not free from a performance standpoint.

(They did eventually introduce new IOKit APIs for 64-bit physical addresses, since Intel Macs don't have IOMMUs.)