Comprehensive Evaluation of GNN Training Systems: A Data Management Perspective

Many Graph Neural Network (GNN) training systems have emerged recently tosupport efficient GNN training. Since GNNs embody complex data dependenciesbetween training samples, the training of GNNs should address distinctchallenges different from DNN training in data management, such as datapartitioning, batch preparation for mini-batch training, and data transferringbetween CPUs and GPUs. These factors, which take up a large proportion oftraining time, make data management in GNN training more significant. Thispaper reviews GNN training from a data management perspective and provides acomprehensive analysis and evaluation of the representative approaches. Weconduct extensive experiments on various benchmark datasets and show manyinteresting and valuable results. We also provide some practical tips learnedfrom these experiments, which are helpful for designing GNN training systems inthe future.