Geminet: Learning the Duality-based Topology-Agnostic Update Operator for Lightweight...
summary
Geminet learns a lightweight, topology-agnostic update operator that iteratively updates edge-level dual variables (congestion prices) and recovers per-pair split ratios with a softmin recovery, replacing path-level state and deep GNN encoders. The trained model is tiny (0.04%–7% of prior parameter counts), uses <10 GiB memory vs >80 GiB for HARP, consumes ~4% GPU footprint in tests, converges up to 18× faster overall and reaches target policy 3.6× faster on dynamic topologies.
AI-written summary · about
chapters (12)
- 0:00Termite introduction
- 1:45TE algorithm requirements
- 3:56TE approach challenges
- 6:48Topology-aware scalability
- 7:52Topology-agnostic update
- 9:16Edge-level dual transform
- 11:23Path-level overhead
- 13:35Split ratios recovery
- 14:41Geminet workflow
- 15:47Evaluation adaptiveness
- 16:51Conclusion and future
- 17:12Q&A
related talks
16mGemma 4: Compacting Intelligence for the Edge
14mLeapGNN: Accelerating Distributed GNN Training Leveraging Feature-Centric Model Migration
17mTopoOpt: Co-optimizing Network Topology and Parallelization Strategy for Distributed...
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
16mGemel: Model Merging for Memory-Efficient, Real-Time Video Analytics at the Edge
19mOptimizations and abstractions for sparse machine learning
15mIdentifying and Analyzing Pitfalls in GNN Systems
12mTopology Over Predictions: Using Graph Context to Drive Real-Time Lending Decisions