[ovs-dev] [ovs-dev, v5] dpif-netdev: Avoid reordering of packets in a batch with same megaflow
Vishal Deep Ajmera
vishal.deep.ajmera at ericsson.com
Fri Jul 27 10:35:28 UTC 2018
> Have you tried solution with direct pushing of all the packets to
> 'flow_map' and per-flow batching only at the end of 'dp_netdev_input__()'?
> Is it really slower?
> If you have some performance data, I'd like to see it, because it looks
> like code should be simpler without this optimization.
I do not have any comparison data. However, for best case where all
packets have a hit in EMC, this will add to the cycles since an extra loop for
batch needs to be run in dp_netdev_input__() after emc processing.
> Also, I'm thinking, maybe it'll be better to keep the receive order
> values somewhere in the packet batch itself modifying the
> 'dp_packet_batch' structure and API. Using this we may restore the
> original order inside the per-flow batches before executing actions and
> completely avoid any modifications in processing related code.
> Maybe, this solution will look more clean and easy to understand.
I could not try the solution suggested by you. However I have addressed your
code review comments in v6 patch. Also, v6 patch is rebased with dfc/smc feature.
More information about the dev