Sancus

Graph neural networks (GNNs) have emerged due to their success at modeling graph data. Yet, it is challenging for GNNs efficiently scale large graphs. Thus, distributed come into play. To avoid communication caused by expensive data movement between workers, we propose Sancus, a staleness-aware communication-avoiding decentralized GNN system. By introducing set of novel bounded embedding staleness metrics and adaptively skipping broadcasts, Sancus abstracts processing as sequential matrix multiplication uses historical embeddings via cache. Theoretically, show approximation errors gradients with convergence guarantee. Empirically, evaluate common models different system setups on large-scale benchmark datasets. Compared SOTA works, can up 74% least 1.86X faster throughput average without accuracy loss.