Machine Learning & Networking
GARRO
Graph Attention Routing with Reinforcement Learning
- Role
- Researcher & sole engineer
- Category
- Machine Learning & Networking
- Links
- GitHub repo
- lower path latency than OSPF
- 55%
- lower path latency than ECMP
- 45%
- lower packet loss than OSPF
- 14.6×
Verified benchmark results
- GARRO35.7 ms
- Hedera36.2 ms
- ECMP (hash)64.6 ms
- OSPF80.5 ms
- ECMP-Detour118.8 ms
- Random123.4 ms
| Algorithm | Latency (ms) | Packet loss (%) |
|---|---|---|
| GARRO | 35.7 | 0.063 |
| Hedera | 36.2 | 0.129 |
| ECMP (hash) | 64.6 | 0.732 |
| OSPF | 80.5 | 0.921 |
| ECMP-Detour | 118.8 | 1.474 |
| Random | 123.4 | 1.652 |
Fat-Tree data-centre network (k=4, 20 switches, 32 links). Evaluation of the 20k-episode checkpoint, from the GARRO repository README.
Overview
Classic routing protocols such as OSPF and ECMP rely on static metrics like hop count, so they struggle with microbursts and uneven traffic. GARRO is an end-to-end deep reinforcement learning routing framework for software-defined networks that replaces those static heuristics with a topology-aware agent.
A Graph Attention Network encodes the live network state (node load, link utilisation and delay), and a Proximal Policy Optimization (PPO) agent learns how to distribute traffic over the K shortest paths while balancing latency, throughput, packet loss and queue variance.
Training happens in a fast digital twin built on an M/M/1/K queueing model with Poisson microburst traffic across NSFNET, GEANT2 and Fat-Tree topologies. The trained policy is then deployed in closed loop on a live Mininet network through a custom OS-Ken OpenFlow 1.3 controller with a real-time web dashboard.
An LLM intent orchestrator translates operator requests such as “prioritise low latency for video calls” into reward weights in real time, without retraining the network.
What I built
- Built a Graph Attention Network and PPO agent that optimises path selection for latency, throughput and packet loss on live Mininet emulation.
- Engineered an OpenFlow 1.3 SDN controller on OS-Ken that applies the agent’s routing decisions to the live network.
- Built an LLM orchestrator that turns natural-language intent into reward weights for the agent.
- Benchmarked against OSPF and ECMP over 500 episodes.