On 8/31/17, 3:13 AM, "Yuanhan Liu" <y...@fridaylinux.org> wrote:
On Wed, Aug 30, 2017 at 07:28:01PM +0000, Darrell Ball wrote: > > [Finn] > > I think we should not further intermix the rxqs distributed to different pmd's, other than initially configured, when setting up hw-offload. If we make a round-robin distribution of the rxqs, a different pmd will most likely receive the hw-offloaded packets - not the same pmd that ran the slow-path originally creating the flow. > > It is usual to optimize caches etc. per pmd and that would not work then. Maybe the further processing of the hw-offloaded packets does not need these optimizations at the moment, however, IMHO I think we would be better off using the first proposal above (use the same rxq as the one creating the flow). > > [Darrell] Several ideas have some validity. > However, this sounds reasonable and simple and we could revisit as needed. > What do you think Yuanhan ? Just want to make sure we are on the same page: do you mean the original solution/workaround I mentioned in the cover letter: record the rxq at recv and pass it down to flow creation? If so, I'm okay with it. [Darrell] This is the relevant part from the cover letter: “One possible solution is to record the rxq and pass it down to the flow creation stage. It would be much better, but it's still far away from being perfect. Because it might have changed the steering rules stealthily, which may break the default RSS setup by OVS-DPDK.” This is a reasonable first cut. However, the flows installed are masked flows but the associated packets would ‘normally’ end up on multiple PMDs due to RSS, right ? But for HWOL, we specify ‘the queue’ to be the one we receive the first packet from. This is what I was getting at b4. So, future workarounds would be ‘auto-splitting flows’ across queues, user specified flow->queue associations etc --yliu _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev