This archive stores the experimental version that added:
- RED-like early ECN / early-drop AQM
- stricter LSA flood pruning
- dynamic flowlet-gap based on path cost
- dynamic RTO cap based on measured RTT
- stronger receive-buffer preservation of contiguous blocks
This version did not outperform v1 overall.
challenge-incast-12:35.8181 -> 36.7333(better)challenge-topodiscovery-12:41.5880 -> 41.0506(worse)core-ring-80:29.9981 -> 29.3914(worse)
The experiment improved incast queue handling, but the added AQM and tighter control-plane behavior hurt the larger / topology-sensitive datasets enough that it was not kept as the active mainline strategy.