Full-batch graph neural network (GNN) training is essential for interdisciplinary applications. Large-scale graph data is usually divided into subgraphs and distributed across multiple compute units to train GNN. Limited by the workload imbalance among computing units and the pressure of data transfer and communication, the full-batch training method is still inefficient. The state-of-the-art load balancing method based on direct graph partitioning assumes that the computational load and communication load are related to the number of vertices and cut-edges of the graph. This assumption is too rough to effectively achieve true load balancing on GPU clusters. We propose ParGNN, an efficient distributed full-batch training system, which employs an adaptive load balance workflow in conjunction with graph repartitioning to alleviate load imbalance and minimize communication traffic. In addition, we present a hierarchical data transmission deduplication approach to reduce CPU-GPU transfer and communication traffic. Based on the over-partitioning result of the proposed workflow, ParGNN implements a pipeline method to effectively hide CPU-GPU data transfer. Experiments have shown that we have successfully implemented the training of the papers100M dataset on 256 GPUs for the first time and verified that ParGNN has the capability to scale to larger clusters.