Re: [Qemu-devel] Re: QEMU-KVM and video performance

[Avi Kivity]

On 04/22/2010 09:04 AM, Gerhard Wiesinger wrote:
> On Wed, 21 Apr 2010, Avi Kivity wrote:
>
> > On 04/21/2010 09:50 PM, Gerhard Wiesinger wrote:
> > > > > I don't think changing VGA window is a problem because there are
> > > > > 500.000-1Mio changes/s possible.
> > > >
> > > > 1MB/s, 500k-1M changes/s.... Coincidence?  Is it taking a page fault
> > > > or trap on every write?
> > >
> > >
> > > To clarify:
> > > Memory Performance writing to segmen A000 is about 1MB/st.
> >
> > That indicates a fault every write (assuming 8-16 bit writes).  If 
> > you're using 256 color vga and not switching banks, this indicates a 
> > bug.
>
> Yes, 256 color VGA and no bank switches involved.
>
> > > Calling INT 10 set/get window function with different windows (e.g. 
> > > toggling between window page 0 and 1) is about 500.000 to 1Mio 
> > > function calls per second.
> >
> > That's suprisingly fast. I'd expect 100-200k/sec.
>
> Sorry, I mixed up the numbers:
> 1.) QEMU-KVM: ~111k
> 2.) QEMU only: 500k-1Mio
>
> > Please run kvm_stat and report output for both tests to confirm.
>
> See below. 2nd column is per second statistic when running the test.

 efer_reload                  0       0
 exits                 18470836  554582
 fpu_reload             2147833    3469
 halt_exits                2083       0
 halt_wakeup               2047       0
 host_state_reload      2148186    3470
 hypercalls                   0       0
 insn_emulation         7688203  554244

This indicates that kvm is emulating instead of direct mapping.  That's 
probably a bug.  If you fix it, performance will increase dramatically.

> > > To get real good VGA performance both parameters should be:
> > > About >50MB/s for writes to segment A000
> > > ~500.000 bank switches per second.
> >
> > First should be doable easily, second is borderline.
> >
> > > I think this is very easy to distingish:
> > > 1.) VGA Segment A000 is legacy and should be handled through QEMU 
> > > and not through KVM (because it is much more faster). Also 16 color 
> > > modes should be fast enough there.
> > > 2.) All other flat PCI memory accesses should be handled through KVM 
> > > (there is a specialized driver loaded for that PCI device in the non 
> > > legacy OS).
> > >
> > > Is that easily possible?
> >
> > No.  Code can run in either qemu or kvm, not both.  You can switch 
> > between them based on access statistics (early versions of qemu-kvm 
> > did that, without the statistics part), but this isn't trivial.
>
> Hmmm. Ok, 2 different opinions about the memory write performance:
> Easily or not possible?

Switching between tcg and kvm is hard, but not needed.  For 256 color 
modes, direct map is possible and should yield good performance.  Bank 
switching can be improved perhaps 3x, but will never be fast.

--
error compiling committee.c: too many arguments to function