Re: [Xen-users] HVM on 3.2 kernel hdd speed issue

[Mario <mario@xxxxxxxxxxxxxx>]

On 04/03/2012 06:39 AM, Joseph Glanville wrote:
> On 3 April 2012 06:03, Mario<mario@xxxxxxxxxxxxxx>  wrote:
> > On 04/02/2012 09:55 PM, Joseph Glanville wrote:
> > > On 3 April 2012 05:48, Mario<mario@xxxxxxxxxxxxxx>    wrote:
> > > > On 04/02/2012 09:27 PM, Joseph Glanville wrote:
> > > > > On 3 April 2012 05:17, Mario<mario@xxxxxxxxxxxxxx>      wrote:
> > > > > > On 04/02/2012 08:45 PM, Joseph Glanville wrote:
> > > > > > > On 2 April 2012 22:49, Mario<mario@xxxxxxxxxxxxxx>        wrote:
> > > > > > > > On 04/02/2012 01:55 PM, Heiko Wundram wrote:
> > > > > > > > > Am 02.04.2012 13:33, schrieb Mario:
> > > > > > > > > > On 04/02/2012 12:19 PM, Heiko Wundram wrote:
> > > > > > > > > > > Windows, of course, does not natively support PVonHVM in any way
> > > > > > > > > > > (except
> > > > > > > > > > > when using corresponding drivers to enable that), so if you get
> > > > > > > > > > > excessively slower I/O-speeds on "fully" virtualized Linux DomUs
> > > > > > > > > > > than
> > > > > > > > > > > you do on Windows DomUs which _don't_ have the corresponding
> > > > > > > > > > > PV-drivers
> > > > > > > > > > > installed, something else is amiss here; it'd help if you could
> > > > > > > > > > > describe
> > > > > > > > > > > your setup in a little more detail.
> > > > > > > > > > Why are we still on PVonHVM subject? I do not want that, I want
> > > > > > > > > > regular HVM to work with linux domU's the same way it works with
> > > > > > > > > > windows domU's. I don't have the luxury to install custom drivers
> > > > > > > > > > on
> > > > > > > > > > some domU's, so there is no point in trying to force me to use
> > > > > > > > > > PVonHVM
> > > > > > > > > > because i can't.
> > > > > > > > > > So, anyone else please? :-)
> > > > > > > > >
> > > > > > > > >
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > Read my last paragraph again, please: Linux fully virtualized DomUs
> > > > > > > > > (which use the corresponding NIC and disk emulation as implemented
> > > > > > > > > by
> > > > > > > > > qemu) shouldn't perform any different than a Windows DomU,
> > > > > > > > > I/O-performance wise, as both of them use the same infrastructure in
> > > > > > > > > Dom0 to do I/O (qemu process). You're saying that they are
> > > > > > > > > different,
> > > > > > > > > I/O-wise, so: please be a little more concrete _what_ the problem is
> > > > > > > > > that you're seeing. We don't have crystal balls handy, sorry.
> > > > > > > > Its actualy quite simple, here is an example: Windows hvm domU disk
> > > > > > > > io
> > > > > > > > on
> > > > > > > > my
> > > > > > > > test server is ~60MB/s (sequential read or write). Linux HVM guests
> > > > > > > > (using
> > > > > > > > same config file template) on the same server gives ~10MB/s
> > > > > > > > (sequential
> > > > > > > > read
> > > > > > > > or write), i tried pretty much everything, from tuning scheduler to
> > > > > > > > changing
> > > > > > > > kernel. I am not sure what to do with it, other then roll back to
> > > > > > > > kernel
> > > > > > > > 2.6.34 on my dom0.
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > Hieko and myself have told you what to do to get decent performance.
> > > > > > > Some examples of fully setup PVonHVM guests are available on my file
> > > > > > > mirror.
> > > > > > > http://mirror.orionvm.com.au
> > > > > > >
> > > > > > > Unless you use the PV drivers there isn't really a whole lot more I
> > > > > > > can do for you.
> > > > > > >
> > > > > > > I can't explain why performance would differ under 2.6.34 vs 3.2.
> > > > > > > This makes no sense as qemu-dm runs in userspace. You would have had
> > > > > > > to
> > > > > > > make
> > > > > > > some changes to the toolstack for this performance to differ.
> > > > > > Performance difference between dom0 kernels aside, what I don't
> > > > > > understand
> > > > > > is why windows HVM domU works fine, while linux doesn't?
> > > > > > Isn't HVM supposed to work same for every guest, or does linux actualy
> > > > > > have
> > > > > > something against HVM mode? I simply don't get it.
> > > > >
> > > > >
> > > > > It's definitely not an optimized use case however I have never seen as
> > > > > low performance as you are reporting.
> > > > > There are too many reasons to list as to why performance between
> > > > > Windows HVM and Linux HVM would differ.
> > > > >
> > > > > What Linux guests are you attempting to run?
> > > > >
> > > > > Joseph.
> > > > Running a slackware guest, tried various kernels, did not make any
> > > > difference. Windows in question is 2008 R2.
> > >
> > > Try using a 3.0 guest kernel with XEN_PLATFORM_PCI=y and all of the
> > > Xen device drivers.
> > > This will enable PVHVM and you should be right as rain for performance.
> > So what exactly do I do with linux guests that I don't have kernel sources
> > for? :-)
> Guests older than 2.6.32 could be a problem.
> Everything else should be fine.
> Anything earlier than 2.6.32 you probably want to run in pure PV mode.
I tried 2.6.37 and 3.2.7, same results. Ok it has become clear that 
nobody knows, so lets forget about this one. I will keep xen for pv 
guests only, everything else will go into vmware.
Thanks anyway,
mario


_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxx
http://lists.xen.org/xen-users