historically, virtualization systems (VMWare, QEmu, VirtualPC, etc) on x86
had to use heavy emulation techniques to compensate for the lack of real
hardware virtualization.
Xen uses a different tradeoff, it relies on modifications to the guest OS. it
doesn't need any new processor feature, and doesn't use it if present.
of course, as we all know, that doesn't work with closed source OSs, so the
Xen people began to work on a real virtualization feature. The tradeoff chosen
was to make it depend on the new hardware features so it wouldn't get much worse
overhead than the already working paravirtualization.
In short:
without Pacifica/VT:
- paravirtualized guests (linux, bsd, solaris): very low overhead
- unmodified guests: can't run.
with Pacifica/VT:
- paravirtualized guests (linux, bsd, solaris): very low overhead (same
as
before)
- unmodified guests: run with somewhat bigger overhead; but not so bad as old
software-only solutions.
or, another way to see it:
modifiable OS: can run paravirtualized, very low overhead. doesn't need nor
use any new virtualization feature of the processor.
non-modifiable OS: need Pacifica/VT processors, get higher overhead on disk
and graphics.
there are some plans about using some new CPU features even on
paravirtualized
domains to get lower cross-domain switching times, or easier PCI
administration, or to mix 64bit/PAE/32bit guests.... but all those are in the
future.