Le jeudi 16 février 2012 à 12:22 -0500, Neal Cardwell a écrit :

Our team has also run into issues with GRO/LRO "stretch ACKs", and
their negative impact on TCP performance. We have a patch that we've
been working with that deals with the sender-side issues. It turns out
that neither the ABC nor non-ABC byte counting quite fixes the stretch
ACKs issues.

On the receiver side, the approach of changing tcp_grow_window() and
__tcp_grow_window() to adjust things in terms of actual packet size
rather than a fixed 2*MSS sounds great to me as well.

In terms of the original 3.1 vs 3.2 issue report that started this
thread, I didn't see any evidence of LRO or GRO causing stretch ACKs.
(Alexey, would you be able to confim by running "ethtool -k eth0" on
both kernels?) So it seems that probably the original issue is
unrelated to stretch ACKs?

I do believe the "3.1 vs 3.2" issue is a matter of how skb truesize
fixes uncover prior bugs in tcp stack. Some drivers were lying about skb
truesize in hope to avoid these bugs...

For instance, I think we dont open receiver window fast enough in this
case, (and we probably enter the slow __tcp_grow_window() path)

On sender side, there are also issues, because if I have TSO on or off
on my 'netem delay 50ms' machine, performance is completely different.

It might have something to do with BQL. More testings needed.