On Mon, Apr 11, 2016 at 2:52 AM, Sugesh Chandran <sugesh.chand...@intel.com> wrote: > Optimizing VxLAN tunneling performance in userspace datapath by offloading > the rx checksum validation on tunnel packets to the NIC when it is supported. > > Signed-off-by: Sugesh Chandran <sugesh.chand...@intel.com>
Thanks for working on this. I think there is a lot of room for OVS DPDK to take advantage of common NIC offloads. I have some high level comments: * I noticed that you've been following up you patches with a second message with performance numbers. I think it would be better to just include this in the patch description since that's what will become the commit message. Even if the numbers might change over time, it's useful to have a reference to what the expected performance benefit was at the time the patch was authored. Related to this, it would also be useful to have more context on the benchmark numbers that you are providing. For example, I'm guessing that the test was run using VXLAN with outer UDP checksums? What NIC was used? * I think we need to have a clearer way to pass around checksum state rather than as an argument, it's just too invasive. A natural place that most OSs use is in metadata associate with the packet itself, which would be struct dp_packet in this case. It's also important to consider how this will be expanded over time. For example, offloading of transmit checksums would likely want to use the same fields. * Please find a way to do this without duplicating the tunnel header extraction code. Having two copies will surely result in bugs over time. Potentially the previous point may help with this. * Please try to generalize your code where possible. In this case it looks like IPv6, Geneve, and GRE (at least the IP checksum) could easily be supported using the same mechanism. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
