Android Emulator – AMD Processor and Hyper-V Support | Hacker News Reader

Android Emulator – AMD Processor and Hyper-V Support

android-developers.googleblog.com

79 pointsel_duderino30 comments

This is enabled by some new APIs we've been working on in Windows: https://docs.microsoft.com/en-us/virtualization/api/. These allow third-party use of the Windows hypervisor from unprivileged user-mode processes (such as qemu). This is similar in spirit to KVM's ioctl interface and macOS's Hypervisor Framework.

> and macOS's Hypervisor Framework

Tangentially and asking the crowd:

What products on the market use macOS's Hypervisor Framework? Virtualbox apparently doesn't [1], but Docker For Mac does [2]...

Edit: and QEMU added experimental support in 2.12 [3], released in April

[1] https://www.virtualbox.org/ticket/14217

[2] https://docs.docker.com/docker-for-mac/docker-toolbox/

[3] https://wiki.qemu.org/ChangeLog/2.12#Host_support

You can run vagrant boxes and docker-toolbox on top of xhyve, which uses Hypervisor.framework under the hood.

Android Emulator, which is based on a fork of QEMU, has used Hypervisor.framework by default for macOS 10.10+ since version 26.1.0 (last year).

Parallels Desktop Lite is available on the Mac App Store with all of its sandboxing restrictions because it uses Hypervisor.framework: https://itunes.apple.com/us/app/parallels-desktop-lite/id108...

There was Veertu Desktop, but seems that it's dead project now.

It's worth mentioning that the AVD (emulator) always worked on Linux - it uses KVM there and has no problems with AMD chips.

The issue was on Windows (and macOS) where it used Intel HAXM which obviously didn't work with AMD cores.

> The issue was on Windows (and macOS) where it used Intel HAXM which obviously didn't work with AMD cores.

there weren't really that many "official" mac products with amd.

Microsoft's WHPX is closed source: we are stuck with an incomplete instruction emulator, no guest debugging support and other shortcomings. All of this locked to a single host OS. Same thing goes for Apple's HVF. Each vendor seems to care only about their own software and hardware.

I'm afraid that the ecosystem of hypervisor APIs are going the same way as with graphics APIs, e.g. thanks to Direct3D and Metal. I believe we should disregard such APIs in favor of cross-platform open alternatives such as Intel's HAXM: Support for Linux and AMD SVM is planned on my side, as it serves as basis for my PS4 LLE emulator.

I just want to point out that you are free to use your own, more complete or faster instruction emulator with WHP. We exposed an API for the one we use in Hyper-V to make WHP easier to use out of the box, but it has no special advantages or hooks into our hypervisor.

I’m curious what you mean about no guest debugging support. Also curious about the other shortcomings so that we can prioritize appropriately. We have only just shipped our first version.

On API divergence, it’s an interesting point. MacOS, Windows, HAXM, and KVM all have slightly different philosophies about what a hypervisor should do. Maybe over time we can work toward more convergence.

> I just want to point out that you are free to use your own, more complete or faster instruction emulator with WHP.

True. This is an unimportant rant on my side, since it would be nice to have a complete emulator out-of-the-box. But it's true it can be solved in userspace (EDIT: now I wonder how could we inject exceptions back into the guest, should they occur during emulation in userspace).

> I'm curious what you mean about no guest debugging support.

This would allow the host to introduce software/hardware breakpoints and single-stepping through the guest. This is extremely relevant to debug bootloaders/microkernels, or in my case, closed-source kernels that do not support debugging. Some hypervisors, e.g. KVM (and soon HAXM), support this: check "KVM_DEBUG_GUEST" at https://www.kernel.org/doc/Documentation/virtual/kvm/api.txt

> Also curious about the other shortcomings so that we can prioritize appropriately.

The API provides no way to add/override/hook arbitrary MSRs while handling guests's rdmsr/wrmsr instructions, or override CPUID leaves/subleaves. This is required to run certain kernels are designed to run in specific x86 processors. Of course, most hypervisors lack this feature and I understand that WHPX is fairly recent so I'm not complaining about the lack of it. My issue is rather that closed-source implementations are not helpful for researchers since shortcomings like these means the whole library is no longer usable for us. Comparatively adding such features to (say) KVM or HAXM takes just few hours.

Also, don't get me wrong. I feel it's incredibly valuable having finally an API allowing developers benefit from Hyper-V in userland. I'm just concerned about the current trend of hypervisor APIs, as every vendor eventually just ships a closed hypervisor library for its own platform, and developers have to put up with 5 different ways of doing the same thing.

(PS: I've invested quite some time on HAXM, but I'm not affiliated with Intel in any way)

> The API provides no way to add/override/hook arbitrary MSRs while handling guests's rdmsr/wrmsr instructions

Isn’t that what this is for?

https://docs.microsoft.com/en-us/virtualization/api/hypervis...

Or are you looking for a higher-level API or something?

edit: and for CPUID:

https://docs.microsoft.com/en-us/virtualization/api/hypervis...

Yes, that's exactly what I was looking for! I might overlooked it last time I checked around May (was that documented at that time?). Thanks a lot for pointing it out. :-)

Yes, I was frantically adding some of these features at the end of the release, so the docs came a little later :)

Finally, I've been avoiding AMD processors for the better half of the decade for exactly this reason. No reason anymore to continue so, especially in lieu of the recent security issues that have mostly affected Intel CPUs. Great work here Android team!

And they really catched up on Intel performance-wise.

This is great stuff, I am so happy we've gotten this support on Windows when Hyper-V is there. Kudos to everyone involved across borders that made it happen.

FYI for anyone using Visual Studio & Xamarin, we've added support for this in Visual Studio 2017 15.8 Preview 1 or higher with the Mobile development with .NET (Xamarin) workload installed.

Miguel posted details in this blog post: https://blogs.msdn.microsoft.com/visualstudio/2018/05/08/hyp...

Happy coding folks.

(disclaimer - I work at Microsoft)

Still doesn't solve the core problem with the Android emulator - performance. Android needs a "simulator", not an emulator running the entire stack. iOS handles this wonderfully.

Performance with HAXM has been fantastic for a while. So no, it doesn't need a simulator. It doesn't need that at all.

Android emulator has improved massively over the past few years, you might want to go take an actual look at it. Even has quite a few features missing from iOS's simulator, too, like camera scene emulation.

The Android "emulator" is using an x86 build of Android. Running the entire stack will increase memory use but should have no real impact on performance. Given that many people are running the emulator on non-Linux kernels (and this story is explicitly about doing so on Windows), how would a simulator work?

(I work at Google, but not on Android)

Perhaps it is time for an Android Subsystem for Windows that is analogous to the Linux Subsystem for Windows 10.

The Linux subsystem is still slower then a Linux running in Hyper-V because of the Syscall overhead.

Microsoft already wrote most of one (project Astoria) and shipped it in some early W10M builds but scrapped it seemingly shortly before it was done.

In another HN discussion somebody claimed that WSL was developed from the Android emulator. It makes so much sense that I don't want to doubt it ;)

A simulator may have some advantages with respect to I/O and a bit less RAM. CPU-wise it should not make a big difference.

There is also anbox that should be close to what a simulator would be: https://anbox.io/

>Still doesn't solve the core problem with the Android emulator - performance

When was the last time you used the emulator?

I'm not OP, but I last used it a couple of years ago and it was pretty slow (I was using an i5-2500).

Has it improved markedly recently, and if so, how?

It now uses Intel's HAXM on Windows to increase performance tremendously. It also allows for the saving of snapshots to launch the emulator almost instantly.

Android emulator on my Macbook eats 100% i5-7560 CPU, drain battery in a hour and scroll with touchpad does not work

Honestly, I would say that with all of the recent improvements to the emulator over the past few years that Gradle is actually the sluggish, horrible part of the Android development experience. The emulator is actually pretty decent now.

Reply on news.ycombinator.com