Why does I/O APIC has so much effect to 2D performance?

[gkjvb]

Recently I test all popular virtual machines on Mac and Linux. There are 2 groups.

Hardware:
1. Macbook pro 2014 mid. CPU: i7 4770HQ, GPU: Intel Iris 5200, no independence graphic chip, 16GB memory.
2. My old PC. CPU: AMD Athlon 64 X2 3800+, GPU: AMD HD4850, 2GB memory.

VM Version:
VirtualBox 5.0.10; Parallels Desktop 11.1.1; VMware Fusion 8.1 / workstation 12.1; QEMU 2.4.1

Host OS:
1. OS X 10.11
2. WES7 (Windows Embedded Standard 7 SP1) and Arch Linux

Guest OS:
All are Windows XP SP3

Test Software:
PerformanceTest 8.0.1053

Test note:
I assigned half of total CPU thread to VM, because others (graphic disk) need CPU too.
PerformanceTest identify wrong size of memory in VMware.
Part of PassMark Rating are not accurate, because I didn't test 3D graphic performance, due to not support well.
I didn't test on Windows, because my old PC only have 2GB memory which don't have enough space to run 64 bit Windows.

Here are the results:

OS X 1.png (47.5 KiB) Viewed 3386 times

arch 1.png (47.99 KiB) Viewed 3386 times

The whole tests are here:
https://drive.google.com/open?id=0B2Pn3tNtsOH4dHVJMlB6ZjlSblk
https://drive.google.com/open?id=0B2Pn3tNtsOH4N0g0cC0xczV6d0E

Well I don't know this is a bug or design defect, but 2D graphic performance is importent. (like office or financial chart) It is barely to use on my old PC when enable I/O APIC. Please solve this problem.

[michaln]

The images look interesting, but this is missing crucial information -- VBox.log files (which contain configuration and other information) from the various VirtualBox VMs and test runs.

AMD processors are specifically known to have performance problems with 32-bit Windows XP and APIC. That does not apply to Vista and later guests (I think even later Server 2003 updates). The problem is the TPR aka CR8 register. Intel has a hardware solution, while AMD needs a software solution. Microsoft never updated XP to use that.

You can safely assume that if there's some performance problem affecting a combination of outdated guest OS and minority host CPU, improving that will have quite low priority. A modern Mac host is a different kettle of fish altogether.

[gkjvb]

Thanks michaln replay. I see it is a special combination of OS and CPU which cause the performance fallen. I will upload the VBox.log of WinXP and test to see what happens to server 2003. But I'm still confuse about why there is nearly 2-fold performance difference on Mac.

[michaln]

But I'm still confuse about why there is nearly 2-fold performance difference on Mac.

Me too. That is a very different case and again I'd like to see the VBox.log for that.

Is there any chance you could provide your test VM with some brief instructions on what to run, or provide a detailed recipe explaining how to reproduce your numbers?

[gkjvb]

test_option.png (18.4 KiB) Viewed 3295 times

These are configs and logs.

WinXP enable.7z

(21.36 KiB) Downloaded 5 times

WinXP disable.7z

(23.91 KiB) Downloaded 8 times

I didn't mention someting:
1. On Mac, I put VM working directory into 6GB RAM disk. (Macbook pro has 16GB memory) So I have to re-test again and the test result is the same as before.
2. All platform use the same version of Windows XP SP3 lite which can save disk space (about 1GB after Windows update) and save memory.

About brief instructions: There is no third party applications running on OS X when testing VM. What I can see the running applications are Finder and Terminal. The guset OS is a simplified WinXP and I'm sure no third party applications installed inside VM. Then just run test software, nothing special.