A method for predicting a residual online time of a peer in a peer-to-peer (P2P) network is provided. The method includes: obtaining M history life cycle sampling data Si of a peer, i=1, . . . , M; determining the number, n, of Gaussian components in a multidimensional Gaussian Mixture Model to be established, where n≧2, and the multidimensional Gaussian Mixture Model denotes a probability distribution of a residual online time of the peer; using Si and n to establish the multidimensional Gaussian Mixture Model; and using the established multidimensional Gaussian Mixture Model to predict the residual online time of the peer. A device for predicting a residual online time of a peer in a P2P network is also provided.