Re: [PATCH net-next 4/4] virtio-net: auto-tune mergeable rx buffer size for improved performance

[Rusty Russell <rusty@rustcorp.com.au>]

Eric Dumazet <eric.dumazet@gmail.com> writes:
> On Wed, 2013-11-13 at 15:10 +0800, Jason Wang wrote:
>
>> There's one concern with EWMA. How well does it handle multiple streams
>> each with different packet size? E.g there may be two flows, one with
>> 256 bytes each packet another is 64K.  Looks like it can result we
>> allocate PAGE_SIZE buffer for 256 (which is bad since the
>> payload/truesize is low) bytes or 1500+ for 64K buffer (which is ok
>> since we can do coalescing).
>
> It's hard to predict the future ;)
>
> 256 bytes frames consume 2.5 KB anyway on a traditional NIC.
> If it was a concern, we would have it already.
>
> If you receive a mix of big and small frames, there is no win.

Well, that's not quite true.  The device could optimistically look
through the queue a bit for a small buffer; it does not need to consume
in order.  We'd probably want a feature bit for this.

I look forward to your thoughts on what mixing algorithm of different
sizes to use, of course.

Meanwhile, I suspect your patch works well because of 4k pages.  80%
non-GSO packets won't drop the average len below 4k.  On PPC64 with 64k
pages, that's not true.

I wonder if "last used len" would work about as well; it really might if
we had a smarter device...

Cheers,
Rusty.