cdall at cs
May 9, 2012, 4:16 AM
Post #24 of 26
On May 9, 2012, at 2:11 AM, Anup Patel <anup [at] brainfault> wrote:
Re: [Android-virt] [ANNOUNCE] Xvisor ARM better than KVM ARM in CPU virtualization
[In reply to]
The intention behind the announcement was to inform people interested in
virtualization about Xvisor. The announcement was an early info about
achievements of Xvisor ARM (for now compared to KVM ARM). Certainly we are
planning to have scientific paper for Xvisor.
consider the members of this list informed
Also, I do agree that KVM ARM can be further optimized but as I mentioned
in my previous replies "KVM ARM will end-up putting more and more stuff
in-kernel". For now you can think of Xvisor ARM = KVM ARM doing everything
in-kernel. Even Xvisor ARM is being optimized so, as time passes Xvisor ARM
is also going to improve further. Its a common wisdom that "No hypervisor
in the world can be better than Native performance". Xvisor ARM is already
very very close to native performance and KVM ARM will come close to native
performance only by increasing its monolithic nature (i.e. doing more
things in-kernel). If monolithic hypervisors are so well performing then
why not to have a monolithic hypervisor made for virtualization purpose
only. The motivation behind writing Xvisor was the same.
Apart from high performance Xvisor has many interesting features such as:
*Ability to work without hardware virtualization support* - Xvisor ARM is
able to boot multiple unmodified Linux guest even on hosts which do not
have virtualization extensions implemented. In contrast, KVM ARM does not
work without virtualization extension. The potential number of host
hardware that Xvisor ARM can support is much more than KVM ARM can support.
Xvisor ARM can in-fact run on old ARMv5 processors too.
*Tree-based configuration* - To create a guest we have to just describe the
guest in form of a device tree (possible even in runtime). In contrast, for
KVM one needs to add the support in QEMU and recompile the binaries.
*Pass-through hardware access* - For hardware not accessed or virtualized
by Xvisor can be used in pass-through mode. Providing the guests a
pass-through accessible device is just matter of adding a tree node and
configure irq routing information in guest tree. Its not just PCI devices,
we can provide any kind of device as pass-through accessible (Note: if
device has in-built DMA then it should have IOMMU or SysMMU otherwise it
would be security breach). We have already tried out Serial Port and NIC as
We can compare KVM advantages with Xvisor as follows:
*Scheduler* - The linux kernel scheduler is very mature and proven OS
scheduler but Hypervisor scheduler can be quite different. Scheduling
processes and scheduling VMs can be very different problems. In case of VMs
we can use info such as: amount of emulated IO done, amount of time spend
in waiting for irq, etc for improving the quality of server consolidation.
*Driver base* - Xvisor has and will have all driver framework APIs similar
to Linux and driver porting will be just one-on-one replacement of APIs in
*User space tools* - For starter Xvisor will use libvirt tools (or similar
open source initiative) for remote management.
*Co-existence host processes* - Xvisor is not an OS. Its made for
virtualization only so no process. Ofcourse, Xvisor has internal threading
framework but most of the time this background threads are sleeping doing
nothing. All the management commands are provided by managment terminal
daemon (which a background thread in Xvisor).
This all sounds fantastic!
But as I said, this list is for the development of KVM/ARM and not a place
for arguing how fantastic Xvisor is as opposed to everything else. Please
keep that in mind.
I am looking forward to your paper.
On Wed, May 9, 2012 at 4:29 AM, Christoffer Dall <cdall [at] cs>wrote:
> Thanks for providing info on your Xvisor project. However, this is a
> mailing list for the development of KVM/ARM and not a scientific forum to
> establish in theory which hypervisor "will always perform better than"
> which other hypervisor.
> If you want to establish that your code base will always perform better
> than all other hosted hypervisors, I strongly encourage you to submit a
> paper about this in a peer reviewed conference. Personally I would find
> that establishing such facts in a scientific way to be extremely
> I understand that you wish to argue Xvisor's superiority in comparison to
> KVM, but I disagree with your conclusions. The code path taken in KVM can
> be optimized to be extremely short and all logic could be placed within the
> KVM module. There are numerous other advantages to using KVM (existing
> driver base, upstream kernel integration, compatibility with existing user
> space tools, co-existence with native host processes, etc.) and with server
> grade hardware I see the reliance on the Linux kernel for scheduling,
> memory management etc. to be a great advantage - and not a drawback. On the
> other hand, when Xvisor matures, feature requests will only increase its
> code size and complexity as well.
> On May 6, 2012, at 7:53 AM, Anup Patel <anup [at] brainfault> wrote:
> 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 [at] linaro>wrote:
>> On 6 May 2012 05:22, Anup Patel <anup [at] brainfault> wrote:
>> > Also can you give example of a code sequence which is faster on model
>> > slower in real world. As far as I know ARM fast models are internally
>> > based models and If a TLM based model is emulating a timer chip of X
>> > 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:
>> "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
>> > 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
>> > 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
>> > to believe Xvisor ARM will perform better than KVM ARM in real-world
>> 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
> Android-virt mailing list
> Android-virt [at] lists