Intel Haswell Linux Virtualization: KVM vs. Xen vs. VirtualBox

[garegin]

Xen and KVM are bare metal hypervisors, where virtual box is a hosted hypervisor. Doing your virtualization at application level is always going to be slower than at kernel level unless somethings really wrong.

not only that. they are paravirtualization. the hardware doesn't have to be fully emulated. that's why I can run two guests on KVM or Hyper-V and the processor and the hard drive are not going crazy.

[n3wu53r]

So If I wanted to run a virtualized instance of Windows for windows-only software, what would be the best idea? I'm thinking vmware workstation, but I'm worried since fedora updates their kernels semi-frequently and things might break with the modules and I'll really need this machine to be reliable.
Then again I could just put off updating the kernel for a bit until vmware has an update. Or even use a 3rd party patch.

[Ericg]

It has VT-X but not VT-D. So virtualized performance should still be fine, you just can't directly access PCI devices etc.

Oh? Thanks for the clarification, chris. Better than I thought it was then, thankfully.

[ownagefool]

Xen and KVM are bare metal hypervisors, where virtual box is a hosted hypervisor. Doing your virtualization at application level is always going to be slower than at kernel level unless somethings really wrong.

[chrisb]

Its also important to note that "K-Series" Intel CPU's don't support the virtualization extensions anymore so these tests are being done at a nice big performance penalty regardless of anything done wrong in Michael's software configuration.

It has VT-X but not VT-D. So virtualized performance should still be fine, you just can't directly access PCI devices etc.

[Ericg]

Note to all: remember he doesn't mess with the knobs (heh). Only tests with default settings. So, for these benchmarks, KVM and Xen seemed to be better out of the box than the VB, though I don't know that I've ever seen computational benchmarks where VB beats either of the other two. KVM, for one, could have enabled the host cpu settings for advanced instructions instead of being stuck with sse2 (IIRC), and that should've sped up some of the benchmarks.

Its also important to note that "K-Series" Intel CPU's don't support the virtualization extensions anymore so these tests are being done at a nice big performance penalty regardless of anything done wrong in Michael's software configuration.

[artivision]

I really don't understand those useless benchmarks. If you want to do a real VM benchmark, then test "qemu wine simple3D_game.exe" on ARM-Linux SoCs.

[liam]

Phoronix: Intel Haswell Linux Virtualization: KVM vs. Xen vs. VirtualBox

The latest chapter to our lengthy Intel Haswell on Linux saga is virtualization benchmarks. From Fedora 19 with the very latest software components for Linux virtualization, the performance of KVM, Xen, and VirtualBox were benchmarked from the Intel Core i7 4770K "Haswell" CPU.

http://www.phoronix.com/vr.php?view=18888

Note to all: remember he doesn't mess with the knobs (heh). Only tests with default settings. So, for these benchmarks, KVM and Xen seemed to be better out of the box than the VB, though I don't know that I've ever seen computational benchmarks where VB beats either of the other two. KVM, for one, could have enabled the host cpu settings for advanced instructions instead of being stuck with sse2 (IIRC), and that should've sped up some of the benchmarks.

[birdie]

Michael,

There's something terribly wrong with your VirtualBox installation - it cannot be 2 times slower than bare metal.

Either Intel VT-X or Nested Pages were disabled or you had some other odd problem.

[uid313]

Note that the Intel processors with the K-suffix lacks support for IOMMU (VT-d) which allows PCI passthrough, which is used for letting the guest VM access the graphics card of the host machine.