Avi Kivity wrote:
> Aurelien Jarno wrote:
>>   
>>> What is your disk image file format, or are you using a partition?
>>>     
>> I am using a raw image file on an ext3 partition.
>>
>>   
>>> Do the results change (on kvm-14) if you pin the guest to a core with 
>>> 'taskset 1 qemu ...'
>>>     
>> Bingo. It now works even faster than kvm-13!
>>
>> real    0m22.307s
>> user    0m13.935s
>> sys     0m4.720
>>
>>   
> 
> I'm guessing this is due to the glibc aio implementation, which uses 
> threads instead of true aio.  The threads may cause the vcpu to migrate 
> frequently from one code to another.
> 
> There are two possible solutions:
> 
>  - use native aio from http://www.bullopensource.org/posix/.  I think 
> the aio signal patches are not yet in, so this may not work.
>  - teach the scheduler about the cost of migrating vcpus
> The first approach will solve itself eventually, though slowly if the 
> current slow rate of aio merging continues.  We'll have to do the second.
> 

if you prefer the first one, Sébastien will release very soon aio patches for
2.6.20 with an up-to-date libposix-aio.

[advertising] Keep an eye on the bullopensource website. [/advertising]
:-P

Regards,
Laurent
-- 
------------- Laurent.Vivier-6ktuUTfB/bM@public.gmane.org  --------------
       "Any sufficiently advanced technology is
  indistinguishable from magic." - Arthur C. Clarke