On Thu, 2007-02-01 at 16:14 +0100, Petersson, Mats wrote:
> > -----Original Message-----
> > From: xen-users-bounces@xxxxxxxxxxxxxxxxxxx
> > [mailto:xen-users-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of MJang
> > Sent: 01 February 2007 14:52
> > To: xen-users@xxxxxxxxxxxxxxxxxxx
> > Subject: RE: [Xen-users] Config: Paravirtualization and Full
> > Virtualization
> > On Wed, 2007-01-31 at 11:54 +0200, Graeme Gerber wrote:
> > > Where's the difference in the guest installation (or maybe
> > in the host
> > > installation?) when you install a Linux in
> > paravirtualization mode, or
> > > in full virtualization mode (VT or pacifica)?
> > >
> > > g>> Full gives better performance from what I hear. Your hardware
> > > should be suitable and the bios option enabled.
> > > If you know anything about bios pls do let me know as Sony have
> > > disabled this option in there bios.
> > Full gives better performance for the Xen client - but since Para does
> > not require complete hardware emulation (and requires optimized Xen
> > kernels), it results in better performance overall, especially if you
> > have multiple Xen clients.
> Even if you don't touch any IO hardware, I don't think hardware
> virtualization is noticably faster (if faster at all) than
> Para-virtualization (and I have a good reason to NOT say this, but I say
> it, because I believe it's the case, currently at least). Of course if
> the guest is doing absolutely nothing that the hypervisor needs to know
> about, there's very little difference in the two cases, as it's 99.9%
> about the actual speed of the system itself (CPU and memory, as other
> components, such as disk and network, are controlled via the hypervisor
> in one way or another). But assuming we're running something that
> doesn't do disk-access or network-access, but needs a bit of help from
> the hypervisor for other aspects, such as memory management, I would say
> that para-virtual is either going to be faster or same speed for the
> same task.
> One thing that will change this is the ability to use "Nested paging" -
> that will allow the hypervisor to give the VM it's own memory region,
> mapping for example 0..256MB of "guest memory" to a section of "machine
> memory" that is 256MB somewhere in the machine. By this extension to the
> architecture (which is already in the AMD specs), it's possible for the
> guest to run almost autonomously with a very small overhead. A hybrid of
> this technique and para-virtualization is also technically possible,
> where a very thin/small hardware virtualization layer is used in
> conjunction with an otherwise para-virtual OS - that way achieving the
> best of both worlds.
> There are other reasons to use full virtualization today, and one of
> those is the inability to xenify all available operating systems, either
> due to lack of available source code or lack of resources.
I appreciate the clarifications. I get the vague impression that there
are serious debates on the issue w/r/t who or what config saves more
resources. I gave my impressions, which could very well be wrong. But I
bow to your greater expertise.
Xen-users mailing list