Re: [PATCH v2 3/4] kvm tools: Use ioeventfd in virtio-net

[Asias He <asias.hejun@gmail.com>]

On 06/01/2011 02:35 PM, Sasha Levin wrote:
> On Wed, 2011-06-01 at 09:41 +0800, Asias He wrote:
>> On 06/01/2011 12:32 AM, Sasha Levin wrote:
>>> On Tue, 2011-05-31 at 10:18 +0300, Pekka Enberg wrote:
>>>> On Tue, May 31, 2011 at 2:53 AM, Asias He <asias.hejun@gmail.com> wrote:
>>>>> TAP based network performance with ioeventfd
>>>>
>>>> Heh, so how did it look _before_ ioeventfd? Did performance improve
>>>> and how much?
>>>
>>> Asias, did you use TCP or UDP values as bandwidth in your previous test?
>>>
>>
>> These commit log have the performance test result before ioeventfd. But
>> the UDP one is missing.
>>
>> commit 739ddbb3b0fe52aa90a84727a6e90da37ce7661b
>> commit 4ed38b41fc034cfb51fec2004f523fe98faa27f6
>>
>>
>>     Netpef test shows this patch changes:
>>
>>     the host to guest bandwidth
>>     from 2866.27 Mbps (cpu 33.96%) to 5548.87 Mbps (cpu 53.87%),
>>
>>     the guest to host bandwitdth
>>     form 1408.86 Mbps (cpu 99.9%) to 1301.29 Mbps (cpu 99.9%).
>>
>>
>> Anyway, I did another test and post the result here:
>>
>> Test shows host -> guest TCP performance drops from 6736.04 to 5562.25.
>> guest -> host TCP performance dumps from 1572.51 to 1731.55.
> 
> That's quite strange. I wasn't expecting any changes with our current
> network code: Our RX thread is blocking on readv() most of the time, so
> it doesn't get affected by IRQ/ioeventfd signals at all, and the TX
> thread should get signaled to wake up just once or twice when the stream
> starts - after which the virtio ring should be full with data.
> 
> I installed netperf and ran the tests (full results are below), guest ->
> host TCP changed from 1862 to 2567 and host -> guest TCP changed from
> 7716 to 8065.

I have no idea why I am seeing this regression in my box. However, your
change should not introduce this regression.


-- 
Best Regards,
Asias He