Re: [XenARM] [Android-virt] [ANNOUNCE] Xvisor ARM better than KVM ARM in CPU virtualization

[Anup Patel <anup@xxxxxxxxxxxxxx>]

Hi PMM,

Whether to consider model for measuring performance is one's own opinion. There are number of Tier1 conferences which accept simulation numbers for proving better approaches provided the simulation platform is well accepted by everyone. 

Talking about code sequences both Xvisor ARM and KVM ARM have same set of emulators and drivers. In fact, almost all emulation code has been adopted from QEMU. Many of the crucial drivers are adopted from Linux ARM. Unlike KVM ARM, in Xvisor ARM  there no unnecessary switching between host mode to guest mode and amount of code traversed in handling any fault is also very less hence Xvisor-ARM will have much less code executed compared to KVM ARM.

In Xvisor developement, we have observed that results of any CPU performance test on QEMU or Fast Model naturally scales up on real-hardware. Atleast we have never come across any scenario or test performing better on QEMU or Fast Model compared real-world (this is true for test running on Native Linux or Linux running as guest on Xvisor ARM). 

In our opinion we strongly believe monolithic approaches are always better performing over micro-kernelized approaches (or approaches somewhere in between micro-kernel and monolithic). Hence Xvisor ARM will always perform better than KVM ARM in theory, simulation and real-world.

Best Regards,
Anup Patel

On Sun, May 6, 2012 at 2:21 PM, Peter Maydell <peter.maydell@xxxxxxxxxx> wrote:

On 6 May 2012 05:22, Anup Patel <anup@xxxxxxxxxxxxxx> wrote:
> Also can you give example of a code sequence which is faster on model and
> slower in real world. As far as I know ARM fast models are internally TLM
> based models and If a TLM based model is emulating a timer chip of X clock
> then it is quite precise X clock.

Support for TLM does not require that the underlying model is cycle
accurate (you can have 'loosely timed' behaviour).

You might want to read the Fast Models documentation, which tries
to be clear about what the models do and don't provide. In particular:
 http://infocenter.arm.com/help/topic/com.arm.doc.dui0423l/ch02s01s02.html
"Fast models cannot be used to:
 * model cycle counting
 * model software performance

"

> Ofcourse CPU emulation and computation
> power will be less compared to real world. To see this behaviour try to boot
> linux on Fast model or QEMU and leave it for hours and come back see the
> time elapsed, you will definitely see same amount of time elapsed as real
> world.

Nobody's arguing that the models are faster than hardware!
Let's try a simple example with some numbers representing
relative speeds:

 operation A: h/w: 1    ; model  5
 operation B: h/w  3    ; model  30

Where we're comparing two equivalent code sequences "A A A A" vs "B".
On hardware "B" will be faster. On the model the "A A A A" beats "B".
(both sequences are slower on the model than on the hardware, obviously.)

The point is that some operations will be vastly vastly slower
on the model, and some operations merely moderately slower. Which
of any two code sequences is fastest depends at least as much on
whether it's using operations that are disproportionally worse
on the model. A trivial example of this is VFP -- certainly QEMU
has to do complex software emulation of the floating point ops to
maintain bit-for-bit accuracy, which makes them very slow to the
point where a hand-optimised-integer-assembly codec is likely to
be faster on the model than a Neon/VFP-using codec, even though
of course the Neon codec will be faster on hardware.

[NB: this is itself a big simplification: model performance will
depend on a lot of interacting things and is not purely a
same-every-time slowdown per operation. Some operations effectively
slow down what happens after them, for instance on QEMU if you do
something that makes us flush our cache of translated code. And
if for instance you have a periodic timer then the fact the model
is generally slower means you execute proportionally more insns in
the timer interrupt, so inefficiency or slowness in that code path
has disproportionately more effect on overall speed than it does
on hardware. There are other complications too...]

> The results in the announcemnt are not baseless we have quite amount reasons
> to believe Xvisor ARM will perform better than KVM ARM in real-world too.

I'm not stating a position on whether KVM will be better or worse
than Xvisor. I'm just pointing out that you can't base an argument
on the faulty assumption that performance inside a model can tell
you anything useful about performance on hardware.

-- PMM

_______________________________________________
Xen-arm mailing list
Xen-arm@xxxxxxxxxxxxx
http://lists.xen.org/cgi-bin/mailman/listinfo/xen-arm