Scaling Parallel Full-graph GNN Training to Thousands of GPUs