Re: [Qemu-devel] Xenoppix (KNOPPIX5.1.1 + Xen3.0.4 + QEMU/KVM + HTTP-FUSE) is released

[Fabrice Bellard <fabrice@bellard.org>]

Kuniyasu Suzaki wrote:
> Dear,
> 
> We released new Xenoppix which is consisted of KNOPPIX5.1.1, Xen3.0.4, QEMU/KVM, 
> and HTTP-FUSE(stackable/network virtual disk). You can compare Xen(3.0.4 on Linux2.6.16) 
> and KVM(Release 12 on Linux2.6.19) on the CD-ROM.
> [...]
> ### Performance 
>  -PI calculation(3 Million-digits) is used to compare. 
>     http://h2np.net/pi/pi_quick_start.tar.gz
>   We confirmed the performance of kvm was very close to native CPU. However the I/O
>   was still slow.
>                | sec   |
>     -----------+-------+----------------------------
>      Native CPU| 14.67 | Core2 Duo (T7200)
>             kvm| 17.90 | IntelVT is effective
>        kvm(off)| 225.1 | "-no-kvm" is used
>     qemu(kqemu)| 24.87 | "-kernel-kqemu" isn't used
>            qemu| 227.1 | "-no-kqemu" is used
>       Xen(DomU)| 14.68 |
>        Xen(HVM)| 15.99 | IntelVT is effective
>     -----------+-------+---------------------------

Hi,

Since your benchmark involve a mostly user task, the performance of 
kqemu must be very close to native CPU time. I suggest you make the 
following tests to improve your benchmarking of qemu/kqemu:

1) Do not use the clock of the virtualized OS to make the measure. QEMU 
may have bugs which make it very inaccurate.

2) For best performances with kqemu, it is better to use Linux 2.4 as 
guest OS (I know this is far from acceptable, but it can help some 
people to get better performance !).

Regards,

Fabrice.