On 13 Jul 2021, at 13:04, Stokes, Ian wrote:

>> On 12 Jul 2021, at 7:51, kumar Amber wrote:
>>
>>> This commits add the mfex optimized options to be
>>> executed as part of scalar DPIF.
>>>
>>> Signed-off-by: kumar Amber <[email protected]>
>>> Acked-by: Flavio Leitner <[email protected]>
>>
>> The patch in general looks good to me, however, the only concern is the
>> performance impact when AVX512 is not enabled compared to the master
>> branch without the MFEX patch.
>>
>> Up until this patch, the performance compared to the master branch is 
>> actually
>> up about .14 ~ .5% for the PVP use case. After applying this patch the
>> performance is down compared to the master branch, -.24 ~ -.6%.
>>
>> So the question to the upstream community is, are we ok with the performance
>> drop for all the users who do not want to enable AVX512 or don’t have AVX512?
>>
>
> Because of the performance drop it causes I'm wondering does it warrant some 
> more investigation and possibly wider base of testing?
>
> IMO we could defer this patch and remove it for the v11 revision of the 
> series. This should help keep it on track for merging.

Guess that would be a good next step. Although you guys mentioned you did not 
see the regression in your setup.

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to