Pasi Kärkkäinen wrote:
> On Mon, Sep 06, 2010 at 12:55:05PM +0300, Pasi Kärkkäinen wrote:
>> On Mon, Sep 06, 2010 at 10:59:26AM +0200, Markus Schuster wrote:
>> > Pasi Kärkkäinen wrote:
>> > > On Sun, Sep 05, 2010 at 01:29:19AM +0200, Markus Schuster wrote:
>> > >> Hi list,
>> > >>
>> > >> I've read about recent efforts to push pv-on-hvm drivers to Linux
>> > >> mainline and I'm curious to know the cause for this. What's the
>> > >> advantage over using pv_ops directly and booting the kernel
>> > >> paravirtualized?
>> > >
>> > > The other point is performance: 32bit PV (paravirtualized) guests
>> > > perform OK, but 64bit PV guests have a performance hit if your
>> > > workload creates a lot of new processes in the guest.
>> > >
>> > > HVM helps there; 64bit Linux guests might be faster as HVM,
>> > > depending on the workload.
>> > Hi Parsi, thanks for your (as usual :)) good answer.
>> > That's the first time I read about a PV performance hit compared to HVM
>> > - maybe you (or someone else) can write a few words about what's
>> > causing that? Could be interesting for other people, maybe?
>> I think there are some XenSummit presentations about it on xen.org
>> website. It has to do with 32bit vs 64bit architecture differences
>> related to memory management.
>> Every time a new process is created by the 64bit PV kernel
>> the guest process pagetables need to be verified/checked by the
>> hypervisor, and this causes a performance hit if you need to create a lof
>> of new processes in the guest.
>> It doesn't affect 'long running' processes in a 64bit PV guest,
>> ie. the performance hit happens only when new processes
>> are created often (kernel compilation, unixbench).
>> For an HVM guest that stuff is handled by the CPU/hardware,
>> so there's no performance hit related to it.
>> HVM guests have some other performance hits though..
>> That's my understanding of it :)
> I think this video interview of Keir Fraser includes that stuff:
>>From around 8 minutes on that video..
Thank's for the link, really interesting stuff. Maybe one should consider
that fact when setting up a domU.
But finally pv-on-hvm makes sense to me :)
Xen-users mailing list