Re: QEMU P2P migration speed

[Andrey Korolyov <andrey@xdel.ru>]

On 02/05/2014 07:15 PM, Paolo Bonzini wrote:
> Il 05/02/2014 11:46, Andrey Korolyov ha scritto:
>> On 02/05/2014 11:27 AM, Paolo Bonzini wrote:
>>> Il 04/02/2014 18:06, Andrey Korolyov ha scritto:
>>>> Migration time is almost independent of VM RSS(varies by ten percent at
>>>> maximum), for situation when VM is active on target host, time is about
>>>> 85 seconds to migrate 8G between hosts, and when it is turned off,
>>>> migration time *increasing* to 120s. For curious ones, frequency
>>>> management is completely inactive on both nodes, neither CStates
>>>> mechanism. Interconnection is relatively fast (20+Gbit/s by IPoIB).
>>>
>>> What version of QEMU?
>>>
>>> Paolo
>>
>> Ancie.. ehm, stable - 1.1.2 from wheezy. Should I try 1.6/1.7?
> 
> Yeah, you can checkout the release notes on wiki.qemu.org to find out
> which versions had good improvements.  You can also try compiling
> straight from git, there are more speedups there.
> 
> Paolo
> 

Took and build 1.6.2 and faced a problem - after a couple of bounce
iterations of migration (1->2->1->2) VM is not able to migrate anymore
back in a probabilistic manner with an error 'internal error unexpected
migration status in setup'. Error may disappear over a time, or may not
disappear at all and it may took a lot of tries in a row to succeed.
There are no obvious hints with default logging level in libvirt/qemu
logs and seemingly libvirt is not a cause because "accumulated error
state" preserves over service restarts. Also every VM is affected, not
ones which are experiencing multiple migration actions. Error happens on
3rd-5th second of the migration procedure, if it may help.

What is more interesting, the original counter-intuitive behavior is not
disappeared but increased relative span: 25 vs 70 seconds for fully
commited 8G VM. I am suspecting some mechanism falling back to the idle
and dropping overall performance therefore but can not image one beyond
standard freq/cstates which are definitely turned off.