I just implemented and tested with software and hardware (this is running
in production for me now).
We now move flowdels first and send a single barrier. This cuts overall
messages sent down by about 20% also allows platforms that do parallel
deletes to take advantage of that ability.
There is also scope for further optimization (for example, insert flows in
Chewing through the backlog here.
I propose to merge the two outstanding PRs as-is, and then proceed with the
learning performance improvements as discussed at NZNOG. It will be better
to address any problems with them after that so that they're not held up
unreasonably and don't drift too far from master.
Hello Switch Vendors,
We have a new directory called src/docs/vendors for vendors so that you can
include a README and a sample faucet.yaml file for your switch so that
users can quickly get started. Please use lower caps for directory names.
If you have questions, please ask.