Document level relation extraction is a challenging task because different entities in a document are often distributed in different sentences, where the same entity may appear several times in the document. Most importantly, two entities that are far away from each other in the document may also have a latent relationship which makes it difficult to extract the relationship at the document level. However, most of the previous works focus on either only the relationship between local and global semantics or only the representation based on local context. They lack the unity of the relationship between the parts and the whole, which has greatly hampered efficient document-level relation extraction. In this paper, a joint global and local Dual-GCN (JDGCN) model is proposed to solve the above problems. By using two global anchor nodes, the anchor node-based mention graph convolution module and entity graph convolution module are combined. The local and global perception representation of entities is realized through the graph convolution of mention graph, and the relational inference is carried out through entity graph for relation extraction. At last, we have made experiments to validate the model proposed in this paper on the DocRED datasets, and the experiment results show that the proposed method is effective and efficient to document-level relation extraction against the state-of-the-art baseline methods.