[dpdk-dev] [dpdk-stable] [PATCH 1/2] net/virtio: fix performance regression due to TSO enabling

[Maxime Coquelin]

On 01/30/2017 02:30 PM, Yuanhan Liu wrote:
> On Mon, Jan 16, 2017 at 12:26:41PM +0100, Michal Orsák wrote:
>>>>>>>> For such workload, I don't think it would behaviour worse on ARM.
>>>>>>> No reply yet; I will treat it as no objections, and please shout out if any.
>>>>>>>
>>>>>>> Both applied to dpdk-next-virtio.
>>>>>>>
>>>>>>> 	--yliu
>>>>>> Hello,
>>>>>>
>>>>>>
>>>>>> currently I am running short of time. If you have any test prepared which i
>>>>>> can just ran, please send me a link.
>>>>> No link, but you could try:
>>>>>
>>>>> - a typical PVP test
>>>>>
>>>>> - a txonly test: running txonly fwd mode in guest PMD while running
>>>>>   rxonly in fwd mode.
>>>>>
>>>>> The second is a micro test, thus I saw way bigger boost.
>>>>>
>>>>> When are you available for the testing, btw?
>>>> 25.1.2017+
>>> Okay, I will hold on a while to apply them.
>> Ok, I will send you results when I have them.
>
> Again, no reply yet. I'm appying them to dpdk-next-virtio. I really
> don't think it could perform worse in ARM, as it removes a costly
> cache invalidation operation (which should be more expensive than the
> instruction cycles).
>
> OTOH, again, testing is welcome, if it's later proved to be worse on
> ARM (which I highly doubt), I could revert them.
Or have a per-architecture definition of ASSIGN_UNLESS_EQUAL if it
proved to regress on ARM?

Maxime
>
> 	--yliu
>