On Fri, 2019-01-25 at 08:58 +0100, Steffen Klassert wrote:

quoted

Finally this will cause GRO/GSO for local UDP packets delivery to non
GSO_SEGMENT sockets. That could be possibly a win or a regression: we
save on netfilter/IP stack traversal, but we add additional work, some
performances figures would probably help.

I did some tests for the local receive path with netperf and iperf, but
in this case the sender that generates the packets is the bottleneck.
So the benchmarks are not that meaningful for the receive path.

I think we can use GSO on the sender if we add some additional code on
the rx side - for testing purpose only - limiting the GRO aggregation
to an (user controlled via sysfs) value.

Beyond that, other options would be using multiple senders threads and
a single rx queue and/or asymmetric CPUs.

Do you have some performance tests for UDP GRO receive?

I have a bunch of ansible(!!!) scripts I can share, if you dare. They
have a lot of hard-coded setting, so I'm not sure how much can be re-
used outside my testbed.

I also hope/wish/think/ I can allocate some time for benchmarking this
on my own in the next week[s], so I'll try to post some results for the
next iteration.

Cheers,

Paolo