Re: Luvalley-2 has been released: running KVM below any operating system

[Avi Kivity <avi@redhat.com>]

Jan Kiszka wrote:
> Avi Kivity wrote:
>   
>> Xiaodong Yi wrote:
>>     
>>> Hi,
>>>
>>> I've tested the guest Linux using UnixBench 5.1.2. The platform is:
>>>   * Intel's Core Due CPU with 2 cores, 2GB RAM
>>>   * CentOS 5.2 as the dom0 Linux, i.e., the host Linux for KVM
>>>   * CentOS 5.2 as the guest Linux, i.e., the Linux running on the
>>> virtual machine provided by Qemu
>>>
>>> The first set of results is for Luvalley, and the second one is for
>>> KVM. As the result, Luvalley's guest Linux is 20% ~ 30% faster than
>>> KVM's guest! It is very surprise to me. I had through Luvalley's guest
>>> should be the same performance as KVM's.
>>>
>>>   
>>>       
>> Yes, it is surprising.
>>
>>     
>>> Double-Precision Whetstone                    12287.7 MWIPS (10.0 s, 2 samples)
>>> Double-Precision Whetstone                     2166.3 MWIPS (10.2 s, 2 samples
>>>       
>> That's by far the biggest difference. Can you confirm it isn't a typo?
>>
>> If not, then it looks like we have a bug in floating point handling. I
>> don't think this benchmark uses sse.
>>
>>     
>
> Even the native Linux numbers are not that high, rather comparable to
> KVM. I suspect Luvalley is fooling the benchmark here...
>   

Most likely timing. Likely Luvally guests time runs to slowly, so they
achieve more loops per unit of guest time.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.