GraphSnapShot: Graph Machine Learning Acceleration with Fast Storage and Retrieval

In our recent research, we have developed a framework called GraphSnapShot, which has been proven an useful tool for graph learning acceleration. The core idea of GraphSnapShot is to capture and update the state of local graph structures dynamically, just like taking snapshots of graphs. GraphSnapShot is designed to efficiently capture, store and update the dynamic snapshots of graph data, enabling us to track patterns in the structure of graph networks. This technique is useful for most graph learning tasks that relies on topology analysis or networks are constantly evolving, such as social media analysis, biological networks, or any system where the relationships between entities change over time. The key components of GraphSnapShot is the GraphSDSampler. GraphS-DSampler can efficiently capture, update, retrieve and store graph snapshots of topology while doing computation at the same time, which makes graph learning computation significantly faster. In experiments, GraphSnapShot shows efficiency. It can promote computation speed significantly compared to traditional NeighborhodSamplers implemented in dgl, and it can reduce the GPU & memory usage during training, with little loss of accuracy. Experimental results show that the GraphSnapShot has potential to be a powerful tool for large graph training acceleration.