SIGMA partitioner unifies vertex and edge balancing for distributed GNN training

Distributed Graph Neural Network (GNN) training hinges on how the underlying graph is partitioned across compute resources. Existing partitioners typically specialize in either vertex or edge partitioning and optimize a single objective (e.g., edge cut or vertex cut) with one balance constraint. SIGMA (Streaming Integrated Graph Partitioning with Multi-objective Awareness) breaks this mold by supporting both partitioning strategies within a single streaming framework. It simultaneously accounts for vertex and edge balancing, and its clustering-based preprocessing injects global graph structure to improve partition quality without sacrificing the speed and scalability of streaming. This makes SIGMA adaptable to fundamentally different distributed GNN architectures, such as Dist-GNN (edge-partitioned) and DistDGL (vertex-partitioned).

In experiments on six benchmark graphs spanning diverse domains and scales, SIGMA consistently delivered strong performance. It often surpassed streaming baselines and remained competitive with high-quality in-memory partitioners like METIS, KaHIP, and HEP. The results show that a unified streaming partitioner can effectively navigate trade-offs between partition quality, training efficiency, and memory consumption. This is a significant step forward for model training efficiency in production GNN systems, where scalability and balanced resource utilization are critical.