[Discuss] choice of hypervisor (was gracefully shutdown guests)

[richard.pieri at gmail.com (Richard Pieri)]

Jerry Feldman wrote:
> Also, WRT: Xen is a different architecture than the traditional
> virtualized system. With the traditional virtualized system, you simply
> load up standard system. With Xen, you need a paravirtualized guest. For

This is very outdated information.

Xen has provided full HVM support, which requires Intel VT-x/AMD-V,
since version 3.

Xen has always provided PV because the basic x86 architecture CAN'T be
virtualized. The choices then are to emulate the x86 CPU and other
devices or paravirtualize guests. VMware took the former approach in
their early products; Xen took the latter.


> performance, I've heard all sorts of things from people at VMWare and
> Ian Pratt.

This is also outdated information.

HVM on x86 will incur a significant performance hit compared to running
on bare metal. It doesn't matter if it's VMware or VirtualBox or Xen or
KVM. HVM is slow on x86 because the guest hardware has to be emulated.
VT-x/AMD-V help by eliminating most of the requisite CPU emulation.
Other devices (network, disk, etc.) still need to be emulated for HVM to
work. As opposed to PV which does require a modified OS but these
modifications in principle permit direct or near-direct access to the
host's devices via pass-through emulation.

I/O MMU virtualization (Intel VT-d/AMD-Vi) and Network virtualization
(Intel VT-c) were developed to address specific aspects of this
particular performance problem when using HVM. That is, in principle the
chipset provides the necessary virtualization to do what the PV
modifications do. In practice, guests require special pass-through
drivers so it's not so different from PV in terms of actual operation.

-- 
Rich P.