VirtualBox performance is much worse when Intel VT-x is enabled

[Michał Gołębiowski-Owczarek]

Binary package hint: virtualbox-ose

I have Dell Latitude E6500 with P8600 CPU. When I run my Windows XP guest on VirtualBox OSE without virtualization VT-x option inactive it runs much faster than in the case I turn it on. It is really visible - with VT-x on I even see how screen flickers during the boot process.

I did Peacekeeper tests in Firefox and I got these results:
1) VT-x off, 3D support off - 1596
2) VT-x off, 3D support on - 1679
3) VT-x on, 3D support off - 900
4) VT-x on, 3D support on - 1314

Google Chrome behaves even more strange - with VT-x off it runs faster than Firefox, while with VT-x on it runs *VERY* slow but looses no frames in graphic tests and it finally gets about 8800 points.

ProblemType: Bug
Architecture: amd64
CheckboxSubmission: b35340b6290029563207fd4b2aa461a1
CheckboxSystem: d00f84de8a555815fa1c4660280da308
Date: Mon Oct 19 04:44:51 2009
DistroRelease: Ubuntu 9.10
InstallationMedia: Ubuntu 9.10 "Karmic Koala" - Alpha amd64 (20091008)
Package: virtualbox-ose-qt 3.0.8-dfsg-1ubuntu1
ProcVersionSignature: Ubuntu 2.6.31-14.48-generic
SourcePackage: virtualbox-ose
Uname: Linux 2.6.31-14-generic x86_64
VirtualBoxOse.DpkgList:
 ii virtualbox-guest-additions 3.0.6-1 guest additions iso image for VirtualBox
 ii virtualbox-ose 3.0.8-dfsg-1ubuntu1 x86 virtualization solution - base binaries
 ii virtualbox-ose-qt 3.0.8-dfsg-1ubuntu1 x86 virtualization solution - Qt based user
 ii virtualbox-ose-source 3.0.8-dfsg-1ubuntu1 x86 virtualization solution - kernel module

[William Davis (hotshotdj)]

I have the same experience with the VirtualBox-3.0 PUEL version from sun's repository. With VT-x/AMD-V enabled, my Windows XP Pro virtual machine is unusable. After I switch VT-x off, it runs well.

[Michał Gołębiowski-Owczarek (mgol)]

I think that about a year ago I had a similar problem on a borrowed notebook. I don't remember if it has Intel or AMD CPU, but it had virtualization support and turning on using it in VirtualBox caused everything to run a few times slower.

So it seems it's a long-time issue... Did sb report it to Sun?

[William Davis (hotshotdj)]

michal: How many virtual CPU's do you have for your VM? Through experimentation, I've just discovered that I can get good performance with VT-x enabled if I set the number of CPU's to 1 on my Core 2 Duo machine (P8700). I was able to set it to 2 CPU's under Jaunty without a problem -- but the performance with it set to 1 CPU in Karmic is excellent, so I don't know if this is a bug or a feature.

[Michał Gołębiowski-Owczarek (mgol)]

@William Davis
1 CPU, IO APIC off. And performance with Intel VT-x is bad, though. It seems that on my Debian etch virtual machine it's the opposite - VT-x increases performace (but I don't have any reliable tests). Windows XP machine works terrible with virtualization on.

[Michał Gołębiowski-Owczarek (mgol)]

A person called Fale here:
http://ubuntuforums.org/showpost.php?p=8167746&postcount=8
claims the problem didn't exist in Jaunty.

[Michał Gołębiowski-Owczarek (mgol)]

Added upstream tracker.

[William Davis (hotshotdj)]

I can confirm that this problem still exists with VirtualBox 3.0.10 (PUEL). Also, my experience is the same as Fale's. I was able to run WinXP in VirtualBox 3.0.8 (PUEL) with VT-x turned on & 2 CPU's without an issue.

[Grizzly (sven-witterstein)]

I am running XP guest on Karmic / 2GB AMD 2x4400+ with Samsung-SSD.
I suspected the filesystem (ext4) but on XFS, nilfs and btrfs it is all the same: XP guest loads for AGES... unusable...
However, for me it is worse when virtualisation is turned off.
When the system is otherwise idle, disk throughput alternates between 3 MB/s and 150 k periodicalle every other second (according to gkrellm). The SSD can deliver 250 MB/s - so I suspect some probs in the way Virtualbox 3.0.10 accesses disks.
(Or the Koala?) No difference between "IDE" and "SATA" mode for the guest emulated controller.
Will leave my phenom-hosted virtual machines on jaunty until this gets clearer.

[tlu (thomas-ludwig-gmx)]

@Michał Gołębiowski: I can confirm that this problem didn't exist in Jaunty where Virtualbox ran very well with SMP (2 cpus) enabled (and, of course, VT-x which is required for SMP). I have an Intel Core2Duo E8400 with 3 GHz.

Now under Karmic the situation is considerably worse: With SMP enabled, VB runs very slowly on my machine. Going back to 1 cpu and leaving VT-x enabled improves the performance a bit (but still not comparable with Jaunty), and disabling VT-x makes it worse again.

So something has changed in Karmic under the hood and not for the better ...

[tlu (thomas-ludwig-gmx)]

I forgot to mention one thing: In the Virtualbox forums there are some posters who wrote that installing the real time kernel (incl. headers) improved the performance of Virtualbox considerably. I tried, and it didn't make a difference for me. But it might be worth a try for others. (Note, however, that the real time kernel is a universe package - personally I don't want to have the kernel of all packages from the universe repos.) Nevertheless the positive experience by the above mentioned users suggests that the problem might be kernel related.

[Michał Gołębiowski-Owczarek (mgol)]

@tlu
rt kernel doesn't do the job for me as well.

[William Davis (hotshotdj)]

Well, I had THOUGHT that running with just one CPU defined in VirtualBox was a fix for me, but apparently, it is not. While it is a huge improvement over defining 2 CPU's, once I try to do anything that needs a little horse-power, everything goes down hill. Today, I was testing a PowerPoint presentation because I'm guest lecturing next week and have no choice but to use the university's computers with MS PowerPoint 2007 -- which I have installed in my XP VM. I use some embedded videos. My virtualized XP became painfully slow and was barely responsive. I couldn't finish testing the presentation until I got to my office later in the morning, completely defeating the purpose of having VirtualBox at all. VERY disappointing.

[Falcone (fausto-falcone)]

I have the same result, poor performance with VT-x on or 2 CPUs. I am running Karmic (amd64) as host, XP (x86) as guest and VB 3.0.10.

[mnoe (matthias-noe)]

I have the same problem of virtualbox being unusably slow since I upgraded to karmic, in jaunty everything was runnig fine. I also noticed that gkrellm lags every 4 seconds for 4 seconds. During these the window doesn't get updated. Other programs run without problems however.

Installing the rt-kernel as suggested in http://forums.virtualbox.org/viewtopic.php?f=7&t=23978 worked for me, so this seems to be a kernel problem. The strange thing is that except virtualbox and gkrellm everything is running fine...

I'm on a MacbookPro5,5 (Core 2 Duo IA64).

[Michał Gołębiowski-Owczarek (mgol)]

@mnoe
For me real time kernel didn't help.

[Romano Giannetti (romano-giannetti)]

Same problem, VBox went really slow after upgrade to Karmic. I do not use any hw virtualization, just standard SW one.

[Grego (grego-mpaja)]

Recompiling Karmic kernel and setting the timer frequency to 250Hz cured this same problem for me, although I'm using the PUEL version of VirtualBox 3.0.12. I'm on amd64 Karmic (Kubuntu 9.10) host, AMD 7750 dual core CPU, WinXP SMP guest.

For kernel compilation http://tinyurl.com/yb7pp5a instructions worked well, but it would be good to have 250Hz kernel as an option upstream, otherwise I need to keep compiling the kernel myself.

[Grego (grego-mpaja)]

Some more notes: I'm using AMD-V and Nested Paging. Especially nested paging made a big positive difference in performance, up to 60-70% faster for compilation when enabled. So the VT-x in the bug description does not apply to me.
Prior to compiling my own kernel I tried linux-rt kernel, but WinXP SMP guest would not even boot with that, died with blue screen every time.

[momcilo (momcilo-majic)]

Hardware
========
Motherboard: GA-P55-UD3P
CPU: I5 750
Memory: 4GB
Software
=======
Host: Ubuntu 9.10 64bit
Guest: Windows XP SP3

Anything but disabling VT-x causes terrible performance of Windows XP guest

Unlike the I7 family I5 does not have certain virtualization features like VT-d.
Is it possible that the missing features are relied upon by VirtualBox?

[William Davis (hotshotdj)]

The kernel issue that caused this problem is fixed in Lucid. I can confirm that VirtualBox (3.2.4) running on Lucid with Win XP and Win 7 guests runs quite nicely now with VT-x enabled with 2 CPU's defined on my Intel Core 2 Duo P8700. Probably safe to mark this bug as such.

[Michał Gołębiowski-Owczarek (mgol)]

It's still not fixed for me with Ubuntu 10.10 and VirtualBox 4.0.4, further information and logs on the upstream bugtracker (which I also reopened).