Interval board
Busiest link in each interval. Select one to jump there.
FlowState is an SDN controller that classifies every flow, forecasts which links are about to congest, and sends new traffic another way. Below is one simulated run: every train is a real flow on the route its controller chose.
Busiest link in each interval. Select one to jump there.
Milliseconds, Python experiment
Percent; over 100 means over capacity
Five stops, repeated for every new flow
Poll flow statistics and per-port byte counters from every switch.
A Random Forest labels it voice, video, web or bulk, which sets its priority.
A second model predicts each link's utilization for the next interval.
Busy and soon-to-be-busy links get expensive, more so for voice and video.
Dijkstra picks the cheapest path and the controller pushes OpenFlow rules hop by hop.
Generate fresh traffic and route it with both controllers in your browser. Each run uses a new random seed.
Flows from the final interval
Eight strategies, two networks, four scenarios, 10 test seeds. Pick a view and select a strategy to follow it through the charts.
Mean latency (ms) at five load levels, normal traffic
Scores follow the network, scenario and metric selected above
| Strategy | What it does | Role | Mean latency (ms) |
|---|
Best fixed-profile Dijkstra against each adaptive method. Each cell counts test cells where it was significantly better / tied / significantly worse.
11 scenario × load cells, 10 test seeds, paired Wilcoxon tests at p < 0.05.
A class-aware Dijkstra profile, chosen on validation data, performed best. The Q-learning hybrids never significantly beat it on mean or voice/video latency, and the result held on NSFNET, a network it was never tuned on.
Two scikit-learn Random Forests. One labels each flow; the other forecasts next-interval link load from recent history.
375 test flows from 1,500 synthetic flows
| True ↓ / predicted → | voice | video | web | bulk |
|---|
The classes are statistically well separated in this synthetic data, so accuracy is high. Real traffic would be harder.
Random Forest feature importance
The same controller logic on Mininet, Open vSwitch and OpenFlow 1.3. A 60 Mbps bulk transfer saturates the s2–s4 link, then a voice flow starts from h2 to h5.
[static] loaded topology: 8 switches, 12 links, 6 hosts [static] switch connected: dpid=1 … dpid=8 [static] 10.0.0.2 -> 10.0.0.4 routed via switches [2, 4, 7, 6] [static] 10.0.0.2 -> 10.0.0.5 routed via switches [2, 4, 7]
[adaptive] loaded pretrained traffic classifier [adaptive] loaded topology: 8 switches, 12 links, 6 hosts [adaptive] 10.0.0.2:55367 -> 10.0.0.4:5001 class=bulk priority=4 path=[2, 4, 7, 6] [adaptive][live-ML] flow classified as 'bulk' (demand=43.96Mbps, pkt_size=1512B) [adaptive] 10.0.0.2:50331 -> 10.0.0.5:5060 class=voice priority=1 path=[2, 5, 7]
Ubuntu or Debian with root access
sudo apt-get install -y mininet openvswitch-switch iperf pip install -r mininet_sdn/requirements.txt cd mininet_sdn && python3 train_models.py sudo ./run_demo.sh static # baseline sudo ./run_demo.sh adaptive # AI-driven controller
Log excerpts from mininet_sdn/sample_run_logs/. The real
controller assigns initial priority by port; the trained classifier
runs on live flow statistics, and using it to move established flows
is future work.