Re: [9front] [patch] ethervgbe: add rx checksum offloading

[hiro <23hiro@gmail.com>]

this netflix guy mentions some important points:
https://www.youtube.com/watch?v=36qZYL5RlgY

e.g. if they don't enable *all* those optimizations there's barely any
benefit. turn one off and performance degrades miserably.
basically you have to use TCP segmentation offload, TCP large receive
offload, some kind of sendfile() thingy, and other stuff that i don't
understand, all in unison, or you basically get nothing at all out of
it.

for us i think it's best we just skip all this badly designed over-complexity.

On 11/20/22, hiro <23hiro@gmail.com> wrote:
> my hunch is also that checksum offloading is not worth it, since we
> don't have any kind of fastpath in our packet forwarding (we memcpy
> every packet).
>
> also, if we did more computationally intensive work on the packet
> data, like aes encryption, i understand why somebody might want to use
> accelerators for that part. but those checksums are dirt-cheap, on any
> CPU.
>
> On 11/19/22, ori@eigenstate.org <ori@eigenstate.org> wrote:
>> Quoth Arne Meyer <meyer.arne83@netcologne.de>:
>>> Is hw offloading something we want? I ask because I think there is
>>> something wrong with ether8169 offloading and I'm working on a patch to
>>> fix or axe this.
>>>
>>
>> if it makes a real difference in performance and isn't too complex, it
>> may
>> be worth it, but I suspect software is more than fast enough for
>> most of the cards we support.
>>
>>
>