Based on the stress test, we need Graphene and UDP fastlane. The end.
What I got from the stresstest was that the only bottleneck was the accept-to-mempool limit. Graphene can be done in private so this is fine too (not a protocol change).
Block propagation was a huge issue. TCP is not a good delivery method. UDP fastlane for private networks, block torrent for public. Would love to see that in the next test.
