DAGWise++: GNN-Based Adaptive Optimization for DAG-BFT Consensus

Nour Diallo, Lei Xu, Yang Lu, Dana R. Alsagheer, Weidong Larry Shi

Abstract: DAG-based Byzantine Fault Tolerant (DAG-BFT) protocols achieve high throughput by decoupling data dissemination from consensus. However, their performance often degrades under dynamic conditions due to statically tuned control parameters. Most DAG-BFT protocols, including Narwhal, Bullshark, and Shoal++, rely on fixed parameters such as batching interval, quorum threshold, and anchor selection, which perform poorly under network asynchrony, validator churn, or adversarial conditions. We present DAGWise++, a GNN-based prediction module for DAG-BFT protocols that replaces these static parameters with lightweight, per-round predictions. DAGWise++ integrates four Graph Neural Network (GNN) heads, each operating on validator-local, per-round certified DAG snapshots, to infer runtime parameters including batching delays, quorum thresholds, anchor sets, and commit scores, all without modifying the underlying consensus logic. Evaluations on geo-distributed clusters with up to 128 validators show up to $2 \times$ lower consensus latency and $1.5 \times$ higher throughput compared to the latest DAG-BFT protocols, including Shoal++, Bullshark, and Mysticeti. These results suggest that GNN-based adaptive prediction can improve the robustness and responsiveness of modern DAG-BFT systems.

External IDs:dblp:conf/brains/DialloXLAS25