Recently, graph neural networks have achieved impressive results in areas like graph representation learning. However, as the number of layers in graph neural networks continues to increase, the performance of the network does not improve as expected. The primary obstacle to building deep graph neural networks is considered to be the over-smoothing problem. In this paper, we first theoretically analyze the relationship between over-smoothing problem and the algebraic connectivity of graphs, and establish the relationship between algebraic connectivity and edge dropping operations based on the degrees of nodes in the graph. We propose SimDrop, it consists of two parts: a sampler based on algebraic connectivity, and a sampler based on node feature similarity. Our method can perceive both the structural information and node feature information of the graph, aiming to alleviate the over-smoothing problem in deep graph neural networks. Experimental evaluations on various datasets demonstrate that our method achieves better performance in both full-supervised and semi-supervised tasks.